From e491157030013ad370cf22164917ddff0e8f8c4f Mon Sep 17 00:00:00 2001
From: Jeongkyu Shin <inureyes@gmail.com>
Date: Sun, 11 Oct 2026 11:06:42 +0900
Subject: [PATCH] fix(ci): store the ROCm verification logs as text instead of
 tarballs

PR #2250 committed docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/guard-logs.tar.gz and rounds/raw-logs.tar.gz without declaring them in scripts/ci/check_binary_assets.py, so the binary assets job has failed on main and on every PR since. Both archives hold small plain-text run logs (212 files), so they are unpacked in place (guard/logs/ and rounds/raw/, the path the report already cites for the raw logs) instead of being declared as a binary exception: text diffs and delta-compresses, and the guard's binary budget stays for assets that need it. The report's guard-log reference now points at guard/logs/.
---
 .../guard/guard-logs.tar.gz                   | Bin 15502 -> 0 bytes
 ...eta-Llama-3.1-8B-Instruct-4bit_r1_base.log |  28 +++++++++
 ...Meta-Llama-3.1-8B-Instruct-4bit_r1_new.log |  25 ++++++++
 ...eta-Llama-3.1-8B-Instruct-4bit_r2_base.log |  26 +++++++++
 ...Meta-Llama-3.1-8B-Instruct-4bit_r2_new.log |  25 ++++++++
 ...eta-Llama-3.1-8B-Instruct-4bit_r3_base.log |  26 +++++++++
 ...Meta-Llama-3.1-8B-Instruct-4bit_r3_new.log |  25 ++++++++
 .../guard/engine_Qwen3-0.6B-4bit_r1_base.log  |  12 ++++
 .../guard/engine_Qwen3-0.6B-4bit_r1_new.log   |  11 ++++
 .../guard/engine_Qwen3-0.6B-4bit_r2_base.log  |  11 ++++
 .../guard/engine_Qwen3-0.6B-4bit_r2_new.log   |  11 ++++
 .../guard/engine_Qwen3-0.6B-4bit_r3_base.log  |  11 ++++
 .../guard/engine_Qwen3-0.6B-4bit_r3_new.log   |  11 ++++
 .../guard/logs/guard/gemma3_probe.log         |   9 +++
 .../guard/logs/guard/paged_pool_past_u32.log  |   8 +++
 ....1-8B-Instruct-4bit_pp512_r1_base-null.log |  11 ++++
 ...ama-3.1-8B-Instruct-4bit_pp512_r1_base.log |  13 +++++
 ...lama-3.1-8B-Instruct-4bit_pp512_r1_new.log |  11 ++++
 ....1-8B-Instruct-4bit_pp512_r2_base-null.log |  11 ++++
 ...ama-3.1-8B-Instruct-4bit_pp512_r2_base.log |  11 ++++
 ...lama-3.1-8B-Instruct-4bit_pp512_r2_new.log |  13 +++++
 ....1-8B-Instruct-4bit_pp512_r3_base-null.log |  13 +++++
 ...ama-3.1-8B-Instruct-4bit_pp512_r3_base.log |  11 ++++
 ...lama-3.1-8B-Instruct-4bit_pp512_r3_new.log |  11 ++++
 ...-Instruct-v0.1-4bit_pp512_r1_base-null.log |  44 ++++++++++++++
 ...-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log |  45 +++++++++++++++
 ...l-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log |  45 +++++++++++++++
 ...-Instruct-v0.1-4bit_pp512_r2_base-null.log |  46 +++++++++++++++
 ...-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log |  44 ++++++++++++++
 ...l-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log |  45 +++++++++++++++
 ...-Instruct-v0.1-4bit_pp512_r3_base-null.log |  45 +++++++++++++++
 ...-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log |  54 ++++++++++++++++++
 ...l-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log |  44 ++++++++++++++
 ...3-Nano-30B-A3B-4bit_pp512_r1_base-null.log |  16 ++++++
 ...tron-3-Nano-30B-A3B-4bit_pp512_r1_base.log |  17 ++++++
 ...otron-3-Nano-30B-A3B-4bit_pp512_r1_new.log |  16 ++++++
 ...3-Nano-30B-A3B-4bit_pp512_r2_base-null.log |  16 ++++++
 ...tron-3-Nano-30B-A3B-4bit_pp512_r2_base.log |  16 ++++++
 ...otron-3-Nano-30B-A3B-4bit_pp512_r2_new.log |  17 ++++++
 ...3-Nano-30B-A3B-4bit_pp512_r3_base-null.log |  17 ++++++
 ...tron-3-Nano-30B-A3B-4bit_pp512_r3_base.log |  16 ++++++
 ...otron-3-Nano-30B-A3B-4bit_pp512_r3_new.log |  16 ++++++
 .../Qwen3-0.6B-4bit_pp2048_r1_base-null.log   |  11 ++++
 .../guard/Qwen3-0.6B-4bit_pp2048_r1_base.log  |  12 ++++
 .../guard/Qwen3-0.6B-4bit_pp2048_r1_new.log   |  12 ++++
 .../Qwen3-0.6B-4bit_pp2048_r2_base-null.log   |   9 +++
 .../guard/Qwen3-0.6B-4bit_pp2048_r2_base.log  |   9 +++
 .../guard/Qwen3-0.6B-4bit_pp2048_r2_new.log   |   8 +++
 .../Qwen3-0.6B-4bit_pp2048_r3_base-null.log   |   8 +++
 .../guard/Qwen3-0.6B-4bit_pp2048_r3_base.log  |   9 +++
 .../guard/Qwen3-0.6B-4bit_pp2048_r3_new.log   |   8 +++
 .../Qwen3-0.6B-4bit_pp512_r1_base-null.log    |   9 +++
 .../guard/Qwen3-0.6B-4bit_pp512_r1_base.log   |  13 +++++
 .../guard/Qwen3-0.6B-4bit_pp512_r1_new.log    |   9 +++
 .../Qwen3-0.6B-4bit_pp512_r2_base-null.log    |   8 +++
 .../guard/Qwen3-0.6B-4bit_pp512_r2_base.log   |   8 +++
 .../guard/Qwen3-0.6B-4bit_pp512_r2_new.log    |   8 +++
 .../Qwen3-0.6B-4bit_pp512_r3_base-null.log    |   8 +++
 .../guard/Qwen3-0.6B-4bit_pp512_r3_base.log   |   8 +++
 .../guard/Qwen3-0.6B-4bit_pp512_r3_new.log    |   8 +++
 ...Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp2048_r1_base.log     |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp2048_r1_new.log      |  23 ++++++++
 ...Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log |  22 +++++++
 .../Qwen3-30B-A3B-4bit_pp2048_r2_base.log     |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp2048_r2_new.log      |  30 ++++++++++
 ...Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp2048_r3_base.log     |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp2048_r3_new.log      |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp512_r1_base-null.log |  16 ++++++
 .../Qwen3-30B-A3B-4bit_pp512_r1_base.log      |  27 +++++++++
 .../guard/Qwen3-30B-A3B-4bit_pp512_r1_new.log |  16 ++++++
 .../Qwen3-30B-A3B-4bit_pp512_r2_base-null.log |  16 ++++++
 .../Qwen3-30B-A3B-4bit_pp512_r2_base.log      |  16 ++++++
 .../guard/Qwen3-30B-A3B-4bit_pp512_r2_new.log |  16 ++++++
 .../Qwen3-30B-A3B-4bit_pp512_r3_base-null.log |  17 ++++++
 .../Qwen3-30B-A3B-4bit_pp512_r3_base.log      |  16 ++++++
 .../guard/Qwen3-30B-A3B-4bit_pp512_r3_new.log |  16 ++++++
 ...gemma-3-4b-it-4bit_pp2048_r1_base-null.log |  11 ++++
 .../gemma-3-4b-it-4bit_pp2048_r1_base.log     |  11 ++++
 .../gemma-3-4b-it-4bit_pp2048_r1_new.log      |  11 ++++
 ...gemma-3-4b-it-4bit_pp2048_r2_base-null.log |  11 ++++
 .../gemma-3-4b-it-4bit_pp2048_r2_base.log     |  11 ++++
 .../gemma-3-4b-it-4bit_pp2048_r2_new.log      |  11 ++++
 ...gemma-3-4b-it-4bit_pp2048_r3_base-null.log |  11 ++++
 .../gemma-3-4b-it-4bit_pp2048_r3_base.log     |  11 ++++
 .../gemma-3-4b-it-4bit_pp2048_r3_new.log      |  11 ++++
 .../gemma-3-4b-it-4bit_pp512_r1_base-null.log |  10 ++++
 .../gemma-3-4b-it-4bit_pp512_r1_base.log      |  11 ++++
 .../guard/gemma-3-4b-it-4bit_pp512_r1_new.log |  10 ++++
 .../gemma-3-4b-it-4bit_pp512_r2_base-null.log |  10 ++++
 .../gemma-3-4b-it-4bit_pp512_r2_base.log      |  10 ++++
 .../guard/gemma-3-4b-it-4bit_pp512_r2_new.log |  10 ++++
 .../gemma-3-4b-it-4bit_pp512_r3_base-null.log |  10 ++++
 .../gemma-3-4b-it-4bit_pp512_r3_base.log      |  10 ++++
 .../guard/gemma-3-4b-it-4bit_pp512_r3_new.log |  10 ++++
 ...pt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log |  20 +++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log    |  30 ++++++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log     |  22 +++++++
 ...pt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log |  22 +++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log    |  21 +++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log     |  31 ++++++++++
 ...pt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log |  31 ++++++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log    |  23 ++++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log     |  21 +++++++
 ...ite-4.0-h-tiny-4bit_pp512_r1_base-null.log |  10 ++++
 .../granite-4.0-h-tiny-4bit_pp512_r1_base.log |  12 ++++
 .../granite-4.0-h-tiny-4bit_pp512_r1_new.log  |  10 ++++
 ...ite-4.0-h-tiny-4bit_pp512_r2_base-null.log |  10 ++++
 .../granite-4.0-h-tiny-4bit_pp512_r2_base.log |  10 ++++
 .../granite-4.0-h-tiny-4bit_pp512_r2_new.log  |  11 ++++
 ...ite-4.0-h-tiny-4bit_pp512_r3_base-null.log |  12 ++++
 .../granite-4.0-h-tiny-4bit_pp512_r3_base.log |  10 ++++
 .../granite-4.0-h-tiny-4bit_pp512_r3_new.log  |  10 ++++
 .../rounds/raw-logs.tar.gz                    | Bin 14119 -> 0 bytes
 ....1-8B-Instruct-4bit_pp512_r1_base-null.log |  22 +++++++
 ...ama-3.1-8B-Instruct-4bit_pp512_r1_base.log |  22 +++++++
 ...lama-3.1-8B-Instruct-4bit_pp512_r1_new.log |  23 ++++++++
 ....1-8B-Instruct-4bit_pp512_r2_base-null.log |  22 +++++++
 ...ama-3.1-8B-Instruct-4bit_pp512_r2_base.log |  22 +++++++
 ...lama-3.1-8B-Instruct-4bit_pp512_r2_new.log |  23 ++++++++
 ....1-8B-Instruct-4bit_pp512_r3_base-null.log |  22 +++++++
 ...ama-3.1-8B-Instruct-4bit_pp512_r3_base.log |  22 +++++++
 ...lama-3.1-8B-Instruct-4bit_pp512_r3_new.log |  23 ++++++++
 ...-Instruct-v0.1-4bit_pp512_r1_base-null.log |  22 +++++++
 ...-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log |  22 +++++++
 ...l-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log |  23 ++++++++
 ...-Instruct-v0.1-4bit_pp512_r2_base-null.log |  22 +++++++
 ...-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log |  22 +++++++
 ...l-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log |  23 ++++++++
 ...-Instruct-v0.1-4bit_pp512_r3_base-null.log |  22 +++++++
 ...-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log |  22 +++++++
 ...l-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log |  23 ++++++++
 ...3-Nano-30B-A3B-4bit_pp512_r1_base-null.log |  27 +++++++++
 ...tron-3-Nano-30B-A3B-4bit_pp512_r1_base.log |  27 +++++++++
 ...otron-3-Nano-30B-A3B-4bit_pp512_r1_new.log |  23 ++++++++
 ...3-Nano-30B-A3B-4bit_pp512_r2_base-null.log |  27 +++++++++
 ...tron-3-Nano-30B-A3B-4bit_pp512_r2_base.log |  27 +++++++++
 ...otron-3-Nano-30B-A3B-4bit_pp512_r2_new.log |  23 ++++++++
 ...3-Nano-30B-A3B-4bit_pp512_r3_base-null.log |  27 +++++++++
 ...tron-3-Nano-30B-A3B-4bit_pp512_r3_base.log |  27 +++++++++
 ...otron-3-Nano-30B-A3B-4bit_pp512_r3_new.log |  23 ++++++++
 .../Qwen3-0.6B-4bit_pp2048_r1_base-null.log   |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp2048_r1_base.log    |  25 ++++++++
 .../raw/Qwen3-0.6B-4bit_pp2048_r1_new.log     |  26 +++++++++
 .../Qwen3-0.6B-4bit_pp2048_r2_base-null.log   |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp2048_r2_base.log    |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp2048_r2_new.log     |  23 ++++++++
 .../Qwen3-0.6B-4bit_pp2048_r3_base-null.log   |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp2048_r3_base.log    |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp2048_r3_new.log     |  23 ++++++++
 .../Qwen3-0.6B-4bit_pp512_r1_base-null.log    |  23 ++++++++
 .../raw/Qwen3-0.6B-4bit_pp512_r1_base.log     |  24 ++++++++
 .../raw/Qwen3-0.6B-4bit_pp512_r1_new.log      |  24 ++++++++
 .../Qwen3-0.6B-4bit_pp512_r2_base-null.log    |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp512_r2_base.log     |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp512_r2_new.log      |  23 ++++++++
 .../Qwen3-0.6B-4bit_pp512_r3_base-null.log    |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp512_r3_base.log     |  22 +++++++
 .../raw/Qwen3-0.6B-4bit_pp512_r3_new.log      |  23 ++++++++
 ...Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp2048_r1_base.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp2048_r1_new.log  |  23 ++++++++
 ...Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp2048_r2_base.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp2048_r2_new.log  |  23 ++++++++
 ...Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp2048_r3_base.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp2048_r3_new.log  |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp512_r1_base-null.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp512_r1_base.log  |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp512_r1_new.log   |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp512_r2_base-null.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp512_r2_base.log  |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp512_r2_new.log   |  23 ++++++++
 .../Qwen3-30B-A3B-4bit_pp512_r3_base-null.log |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp512_r3_base.log  |  22 +++++++
 .../raw/Qwen3-30B-A3B-4bit_pp512_r3_new.log   |  23 ++++++++
 ...gemma-3-4b-it-4bit_pp2048_r1_base-null.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp2048_r1_base.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp2048_r1_new.log  |  24 ++++++++
 ...gemma-3-4b-it-4bit_pp2048_r2_base-null.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp2048_r2_base.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp2048_r2_new.log  |  24 ++++++++
 ...gemma-3-4b-it-4bit_pp2048_r3_base-null.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp2048_r3_base.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp2048_r3_new.log  |  24 ++++++++
 .../gemma-3-4b-it-4bit_pp512_r1_base-null.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp512_r1_base.log  |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp512_r1_new.log   |  24 ++++++++
 .../gemma-3-4b-it-4bit_pp512_r2_base-null.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp512_r2_base.log  |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp512_r2_new.log   |  24 ++++++++
 .../gemma-3-4b-it-4bit_pp512_r3_base-null.log |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp512_r3_base.log  |  23 ++++++++
 .../raw/gemma-3-4b-it-4bit_pp512_r3_new.log   |  24 ++++++++
 ...pt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log |  23 ++++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log    |  23 ++++++++
 .../raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log |  24 ++++++++
 ...pt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log |  23 ++++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log    |  23 ++++++++
 .../raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log |  24 ++++++++
 ...pt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log |  23 ++++++++
 .../gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log    |  23 ++++++++
 .../raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log |  24 ++++++++
 ...ite-4.0-h-tiny-4bit_pp512_r1_base-null.log |  22 +++++++
 .../granite-4.0-h-tiny-4bit_pp512_r1_base.log |  22 +++++++
 .../granite-4.0-h-tiny-4bit_pp512_r1_new.log  |  23 ++++++++
 ...ite-4.0-h-tiny-4bit_pp512_r2_base-null.log |  22 +++++++
 .../granite-4.0-h-tiny-4bit_pp512_r2_base.log |  22 +++++++
 .../granite-4.0-h-tiny-4bit_pp512_r2_new.log  |  23 ++++++++
 ...ite-4.0-h-tiny-4bit_pp512_r3_base-null.log |  22 +++++++
 .../granite-4.0-h-tiny-4bit_pp512_r3_base.log |  22 +++++++
 .../granite-4.0-h-tiny-4bit_pp512_r3_new.log  |  23 ++++++++
 .../rocm-unified-engine-gfx1151-2026-10-08.md |   2 +-
 215 files changed, 4222 insertions(+), 1 deletion(-)
 delete mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/guard-logs.tar.gz
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/gemma3_probe.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/paged_pool_past_u32.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_new.log
 delete mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw-logs.tar.gz
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_new.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base.log
 create mode 100644 docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_new.log

diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/guard-logs.tar.gz b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/guard-logs.tar.gz
deleted file mode 100644
index 6f1c976b366892a654bc298d64b2226c2003d9cb..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 15502
zcma*Oc|4TS`!`OuvX&)V(rQVTP^2=HsBB3RC4>+{MwI0yTiLfF+bAK)QiLdD-`5Zt
zjD5*&7-qZAbC1vG`~5wC{Qh`)y<DBS&bgPl&ULPHuJ`pG@rb<)yZ@M${mf!UuEpnM
z%$YV&(AWGgNLwAgf$J+_K?_uR`G|0x7P%}XaB^RI*f}MWW|(~P_^J7i&wEB3J}Swp
zbf6d3sGm9%Qv4o0Y0h=6!o|2guoLKgwOvLl+Pp&qKJli%PYqJvAGm|%y=A`mz{PnC
z`eRf*I=cEqCDl}X8-@4%C5lQSQzTDexG^A?HZWj-1uuLE)_}Y*+xM;80cADUF*GOt
z-EDME-p<1yWp=RW+{0n0>G|^s43@XeC3`5G?0Wgv_l<%=V9|fsS<Y1s{!aRdrB;yE
zaI>>G8vM+o=S|ctgLZT5=zaW%>xvg3KczVmX5oK}srl`e^28(@E;S;5IkyO37B*FM
zFAWm)Ym={%-<GX0R4ErdxpGbg<x=hYbHo*1ln1|tEg<KNuRkh&kjXPoAr8&{dq!fK
z+?MMX>TsP{e{zK*Kj~o!miz*UY7P7T7+ZqI@NijNDj<P#URC);Nfh__QJNn|Nt|nn
zYsEk!r973Vs;awrb!YPF_VQFU>GDy$GYWryxm-tf82<XaVWx91pSDFK?$|t^8ZnrU
zpfwjf6MW)Cd{|zvT{(6x=Sl9i_?N*u*dJ9DqobC-a3y7DwtaN8oOY+x{MV1at1H!`
zD`Bn+qjT%f)_T?7mh#kycG}8|K3@^A4#C2hq=#+~xVm(gW$Z_AQ&=!!+FonA_}a1{
z+S1R{gEj-t5|_0BH(JHNjG{dD1a8IvLX|iu9tPwY|KsrTd;Eu78667n+;#-(w87NW
z_H9cqukBT=<M&L_YdDCBicPr1Pr-(2y)|n+y95@b{V4Le!=UN4suh^IOhjq0QRJ`A
zF^rYwz@0`%$2Rq7sg^6{vbUyqM(`iwALEoJJCpqX;;}#Q_)%Br@z~ed#nd8oxMR$6
zWMrj0hxfbkh4q9%axTZJiv3vIHl^K<@~LXuv><B;{|l5>ff3((xaB0nL9VvuS<^84
z5vY1}Reg*W<_wRMXCH2FRl6UZ!b^49DbLCD+YR0LOqSg#?5gdrA?Kf26z@8?h|!ED
zLJ)1CeG8`XP$;9s?xd>x(L30;xPPNW%3>7-=PI~;k1)|rquisxFF?C(cWg1xW*1n(
zKJJpv)|vqMd%bZD_X-F&u$CxHwETHC7RcKHrp-wrC_6J=RYh9}+p>oI)}Q<M($3-X
zn-2SrSRNj!OKoT$O5Q4pTEP2OIKuWg=EH)WNl#WqZ2o#jbvDzsMyoG9#rfi+lqyK3
z1U6B-QSn5Dfk!^S;y#nQiAkf&1LZ}d_=wpUxfOr^*!kI=E8ZSHgk|#7=pxbb>*#|O
z^(A?gsLtWBf9pK2XREXb(5Kbkb?phO|J0K`6ud{Sg!VEC)rXv65^4_--e71BS!WXJ
z4#{Iv5a7_R7O3a>Ia~MXea1eyt~I`-_z0UY)r(i3xG`)`Iev@$&3)12iCZY9a(dxg
zup?Ki@uT5vF$2|$E{L+|)AZ*OX3PasOg7W>y+2b+_v)^FDqfvCVN3#Hp+EJw7;aH5
zb^Rv6-&7_r*a@)%OU=;LsyuuOm5SZwR|l>d&Jr*_NWB}F+BRFdZsTZ!@`5M*$D4Ty
zI>VV>5seFTS@!SYIU-VJ#~C=fM}428pfD@Xkrsu6T!`>y&<w3*+N;QUSmUO3jey{Z
z`?kqi44O@6kdCJ7%=>K*#jx$SO+!?TM#xQCRC(z_gf3n)yo{V$Lxd7_M8Fhd!s5!<
zaQ6r=canzfb>TIDg}-6{+*|+qEA`E4So}D+!H!#p=KAwzUB5pK0Yyg-NUAkbo|BLT
zP_l!4JG=w-kr@eK+|F0im*!lJ-F_cC|H9{i4cfw_n$(_{^w^qHdvHO4J8Kaw6w~Xe
z#)<y*tbsu6I@nF2o|pmWN$OJ4py)EK=viAijK6C_1tNPM5I;Z16DWwAnZcIH&M;u6
z1rs3oBvj-qCHS585e)+l-n~U)t%EPq4xZR=n=Ep2|7{_+3+(%tg|&b03l4U!-2bZi
zSl<e`y+TI0(7wV&o8ZgwIQxU7MXZnEJH(UaUKdw@qRG!;x4?D{iQz1Kw0Gn=OObHq
z1(h;`r0!CMv;VsC6vpEjD?Jn=FOQxVesnGBkl-WllIb4`SDJ3STZf8!cLw*4n_qWx
z*MM(4b+7r(YZ|BdkmCWq*S*H4#l$Z5@5@0i^dYx`Zj(n$>RyHCMBQlZCie8Ey6BYx
zuqY-sLNjZ-4P&a{?lv15Ms6EI78u^oG(SE8gynXOgK8Q3yJ>)JUNQ~Htk)}u3$g4!
z{JTSoVd5wz_^0j*7u(<9MT_Lo<zw~ZjL0|7Ld33vaQ}~k<n=PVWM_JPf)kl__WzlB
zn{{OB6ZRr1fx{ZH?!1i50&Hk@ZRBFf1ITB@p<>-bF&Dmotf?C9kG*8<*aYvZ)6zRM
zlsult?gkChOTG&vIe_Jr8mC#c)QtbWY@ri=_hk`-I&9Dmm*Gz=RK`kK2$~OV>X(Q4
zor|50wdTAr*s8#ty=W_B_^X=7K4Q+#KYjx^4Fch#+*{B|P#&$7v5LN)G6*mkitWJZ
zO($?tb>CQ+0T1LLq1Npm^OO9N`?+UJ83JQXW(o|+=4>A3DI8O0_GpIXAtLnrOzCgt
zV`;L?lfA8_6x6$4u%W2W7a+4mk6orNUgzWb=%!r*WxB^suK*@=9`Ip?sj)n<Q{udx
z=_+%V{&l4%2KNlt-D90c9KcHCD0<FT=r!U69kC+k<vL=|JeTWOODL#N*bN*3PA6;Y
z+>QS00ETfE$}zFIt#Xg^C-S~U%pTZ457+vOdi&`#GY$q^ba~c$@p6V*l|v+_Ax89&
zej(rp<M4vsD!u`8pMcX{s2RK|v4ip9d&a!_?1Z0g(8DYhkG4AyK>!T&6>!pDXqG99
zp-#*5tT%)8k*biNx*V<vHbv$_5v==dg)WL;3(f0cc*$yW^hZcq1m|&N`Z=@c!UoZD
z?ymVM`s^D|(I=hvUWNqj)zHT6;b-?fPX`>Cx_c23s&sLi9!2tyq2WyCg4Q%~G5eh#
zGKK0vvsW7j=8U%=e$qK~Pq+U9wUp^N6~-j3hAC#lhGMJ0a~PCGYoRp?V*!uhyGCGA
zNX5RkTtR%Ihvd(c+Ma+Gi_T+tpsAKG#cVb~@;?>XM4%xrjyDVqTr|}wTGhYLNjP!I
zBF=;E7yZCdn2L2?1=%We-v~&u@VfNqJPq=0;{-)j(V2}jY;<Z=#sheDW+N=l7ZWO6
z!W`J6aF%snaBSs`Zkz(WiL8m`?d<wYS~Xgz6raS+bmfUk*Oi8nR~6z(hb%EFIwG+P
z!UYN2W6>mV#{jl`j_(34br9{bjS|g=b)$|3#tC4ey=c3<s@-KPe|7XA?ev58<Kk5E
zz>TwhGFmUe`B7@+6!eYcku7>jit5V6sC~!tg45FtEPeV6Jv|l#=b%%EpJUDfj}ta)
z@Igpe1y_`6Y4ngd)3>GHe%N*8*u%k}O;dQKC^2c(fjz~?tqZ)Y_SQ7h&I_T_6~$Ob
zL$$*+2c`S0dt4mQLTFKuT~+AUc{MvV&T6ix=%6CwoT+hK&%EjmqT>FqN@8;U(@0`&
zhuSHX8OM_7HvQ-SzGR*Kd(rRD+MUB=p}*ZVN?ZHvxm;%^T%^R<Mm9D!&Fb$xZ`@TO
z`$FxX%Z&r}ss_pV-BwQTC;ut^)SvsORG~k3-;Z_hY)XFjxQLF&mhjk=+m>*&>lD(}
z+KY5q6d+x0-N+5Bt7J?uXAc>@&;L4dQ}zDSD}BL&z_3c-Mgti08Eyxk&cQ0$Njnf1
zqkR1QHUF+1)I~dOsytyCiZ^ryeKsxPEw2oxNKEgB3j2jvmz{s>U%Qt2-c+(a6ZfV?
zfmgOWvp#b<8JShF1a2kj$IL585B>B0!OQ+P8Ut~4M^PuI0(@Cb<0i`73MJ|lr3+8h
zyG-1-VEuWaO+40p?U|k;_x<`>6Z@!g8^ju0ipcA}iupn2obAveF)4EIwh0}`mwG)b
z1!*)fk~onyu+Cxx2ShSarA#ps-pq)?*&`8^IT?UP{yQ?oiSOGT>V6I}07cLgqgj2P
z{xF8X$%T^ew)@`XFINB2)XX8K7T{W$UMP9#JGi!Krdj8N^d4(!@3Uu-E}B5p7FMCT
z-R$uX>Idv&UC$x!Q99_hWmflQ`H`~tLGIiQYbv#yNXE^al-`&485CQCn5UJ?Kt(%|
zWSQS`(_KU~kFW}sE`?x#qsG%t@J};|%5w(v!5ukYrkuNNmd}45c<?&W#M!9WCC!!}
zEhH$$f@{34Q~MN!=6B|k6k}<p_81jkN<;H2uU-jqY!N+eAFC&%ym%!{)64C*bx#@(
z;)yla!!(uDZG}`H1Z36Aa;B`lug_dKtj3wLj5J#5jYXu9cUY10OR&}BGlvXo`X7eO
zKA%@`ND1Q(dW1B_6COT3^Cb}JJWUs00_ZKe_yQr~TK(gkgF){2l)T2`8%cSMRzc~H
z9-sO2dq7Y2Trufd4EJSz_iJmXPwBC)gWS*w{K%-wo1J3Xgg8Ww7={PjQq^C?zXIR;
z--G71b8u2!5=lINuhp=L%BZS6RXjsc21cFqR%!ry3@%w<!PH25g>Mgc#5ZHj3wxCV
zatAKgEb!T1IUnfKkeL;rrOL;G3)9qQ)C{%xz=`7O))Icv9R6C78u?g){go68A8P{~
z&&H&yp8StPMm?F#!J(ea&iU`Vb&o-4pzXEjsNe-L{|C17EN#qi^I?{RNvH`;!QZ~O
zgIhyo-=M9Y;yt4jhXv)^NaQH`_st-JfIU^M?)PFF<+Rp(N<=@L(0uB~I+x!Yg*zYp
zOjwkBQ(nLP&e>;F9Br+}{zpMir<NSDAa%D&LHMf8XdRlr)GV1hh{%pajyu+BoPVYD
z!Zbxg78D#X3c^j9E_G<3wLm4$A*~ArRH0O^Fv}yU=`CRXD(cepHI$9#IP~{NJ&ZEi
zpr1<w>-^W#sPY%>SOGzd1KH_W=Ig*sc<v~l?ixzm;xl)b=>x}tgg>J?vj6n>!`4u0
z;W|6wV%gd|;>f$;l;)+_Cm_VDyjaB=A5vqugPwF{*54Ka$LodAuB>9(|Iu}3)(b^~
zUhAH8;oWz8Gn6T+)S{62^X8fFLQCY^da|r6q2YOrZlb)#5t^^xBYu;-{5jd&**fl*
zNaxG2dPCJGLTf-Az%+3zctV+i|INb0*%$fcsFgz3GV`A}|6Kf6dySk~NYq*lF8X`^
z{j$BuKVW(wv#-ztK2M_WE%2(E=~q?2)5GHp484VuatLxyQ&rQw+yu|9owzf2Y<Rdo
zv&O(sg8iI$YV^yf=A106zOM^X+*t%${+Qn9iZ`NqpNoXY+_B=2hzuUSbY>-+C2|In
zk<=yW3k<Q`1|$f#4a~*oP(LQY;QqC;p)T|3O)w!{LdABJf^s(cPR~;(Zk3=Hj;eW*
z{Qg4EpZQI@n4nGtWVV<4va_rf(XCSKxEx&_&Avcee&-}PF&2Hcc%vTuO0<xw`O`F1
zaD8_?s)GPZ7i$uM*Dn}aXnYSW-_m&UuT8%jV5cEZQ=(#*+oTR|1IEf7APwWS-^f;g
z^4`ee_ve!ghS2sEQ7=ZkrC2+~V~vW18ITA7X=v9zV(2mIp)v}yq%vw~fAxj~ID@(v
zGV5zje!dmm`^s&{z4w(ht$VqUq)og>Fn30|iq=zcUw#=%Uf%&tTlwSQjpHH~cc!Ni
z6dlpf*H7~3SOE3Lk4T`;_Z27om>)}n<}90L{LZDt#BB^8X<}2qHy>KZ=bWph>)j%C
zkT7h^&$d({tTM%`Sg&K<G_2SO6U+l{5%F6~*fkf}w3QbW9FMd6Lz)BI8H)9AsxJZG
z;gJl=;(>xf+qE5@pl(0H_ntaQm$Jb1;Mw<?9{72wHMOVxP;g011Eg0D`tC8@BB6To
zf#7U0FsY~mrg7I$jqF<}WP*^azfK~EmZUIexdcAZ!j2!lMHKIN%0R@IFV^loXKH=+
z3YtG;7*b@vqojp|QGYbQYEx|#<%445L}Pmm09#44@1cW!QM!eXqQH$VsHce)$p0l<
zPH4RbcMe$b)!pe!gOP186bxD%Sf7MZ8YnAR-e1`M#|rVRI>FCC1PCneWvRxo+u#1J
z<2clx{Z8ucFK#iG{K2DHLSG()MZX+aL_+opF*Wu(MOSn*7d??UOF1mYiuO1&Ao$?2
z;+CLZM3Np$*{;&oJ^7S#FY%I^-tEk@YSl*%$tgdp>>-K!Pz9}F(`N8-&;{1<w)lrN
zJ+P*j7kyxQ8bTar2lMY|-6U0B12sYIW~Um}p7wo-IyD~{$aLoq30-YbyCK%8rRjk}
zJd-L&U+r;`FQdSRtlL7+$M+6H3nofNC)&hY73O#3)gm=X$AwwNz}2vIDd<Ff<`lL?
z!Oe_AH^j4Fqwcr$DAz5Cbsf^vW=MY;`m%1Bu!AXSCZo(&5NoX|wwTO<_LH+9-<JSP
z9>UNBv^J+T1I?2e@X7n?#7s-Cn<^O<Z!sca$B#jn#X?*|#|Viy$ogTLmpV8FAzPjw
zgILaGlDIHyxA-mV@s5i*aGfl(9!t|8TaWb`^{8-D7Bz;GsaZ9OfOV@_!+KC%SBSZ2
zTGBHMkXt%dPD16oOh8W88afRApya`YOGtQ381<=A1qolTdv9!G?QfKyNK;G6bmi?r
zi<42N(nh&K9{c7ZknQ5!FHH2q17vNT+MMk^qf^1vy+*|gG+TZ*0ShIgo&_2@St?aY
ztx&Ui+IofaC<TJIEMLGjc(7<gGbIetSJ3#@KFUK6GCfUzEzzd(5fs2+v4PqP?O)rF
zsTJ2!Ui6yhPCIw%QkH3rlU%1(-WgcqO59pXyP^?)lX#2N5b>H?c;gO}FbnSRb)CGM
zutsQ}<XTwYL{hD|1PL13GJFANt&*W^M}OJ=_a`fMfWQgG2&W~$r;GfNh~V{HF>vJq
zJbaVVAm$11(3Zhuz%pygNTA;h7j4kIlml{Ddie(I#ti|MQCLT(mcV8!R7wrwwmA|d
zB(}}Zje06_W(ka0S}d1Uo;iq<-_GBH&UTG4gQCAQ*IVAk2qd7`f6v4#uOkvhpp$Ak
zbdp1Jy)++|2KawP>JGYYRvX4Mxpvzu;?h2!N;rhYeXuIQ%u5ZC&#AGORYyhdSxs#&
zW8LafXYbZ12g794$JJqIIfjC7r{;Q|_-zn}0yv+x)085`8OW!{P!+z{2-Qr!PXsp|
z_S_Y*aeDdP#9>cz<V%TwJSh=<h6JN=X6XiK{-oqJ{SsjI4;x0rzpAIAOzD1^X<-ds
z7*LRtMLldP=6P=fHg(r^{|fAvAeLc|I@omWH`frjq8MN$%W{tLI{M`nKxTf}5Y3-#
zet8LdSFt3q3=A)#vfsQ1I^wgS)YarGaJse`+9t5E8-%c3Z$RF|=&fl2YMrLPiu;)2
z8kG$>&p5xJ+=_UajoADyIV^e)TU2D!ljU?j{@1A?!Onm%0M6b=4MBE1B+M+f5iUBa
z(OMr(N+TlS<}FR|-7K`7$pLg|Y-zT!8{&IJX`@u3+b>jaN4zZXIhJ+Jc~eC!M2%eK
zMLa`_xlTkG(gu6r{4!W~pXClX#cMWeEN=gR0=;PCq^#mo^pI9hBYd2Qq9!{K@%8>k
ztbmJ!&%4qehq){I_4u7V_lt#8s(tbz_WlLp#O<nOjJ#LD*w6nPizLKrUZ=Qf*@GC0
z<@ei2UVLg#V!!9WzVpk-CViTMaRNR(2=at_o#g)9MlQGGsbjZINoK)U_0iN`Wcu$T
z(|_7Qjs0AENaWK~PYvca>WA4p_Fbf)WQ7xdf@K`I0&?$U0}rv+rwQy7{0qZ(*1)2S
z?pBf=z?3L1m$DU2;R#y}EHM{(S60StkT`IWAI)<JIV02bbprdiwP`x4?L={3ilHHE
zHW_^6+CtqbC19gY5yS`Mu4OfxB3hod26pNc`3M@u8O$VJ%3OZl50{%9U1+<`rbBT&
z9DaEjC4BWE5fiBxS$~MnPAL7k1pA1D$~o;K#Ln5};U}i`0M>i?;C+~QN(@M^UI6lX
zK0v<oAhP_Aj?0!=#OWFUx~BnL^b>gm(AW&>28$1x8Y8@1N(0!vh7Oi3Glt2QN+wqR
zeKK%$?pA-Y?mqhn-PV&@6jT;D;wF%(qhVv0sgSi^D=0YDD<ethh88X%9?^Xkd`FHV
zQ&=i|Mc)kWSMzfwN91Lfy&Oj=pQD9De+Ur#J@r7!oRt#!vxk||{RO^(gT!{QYb(W|
z(ZON~f$<VYzuW>&Z$E+Wq<_$|$0yL~3UW<=`A_oTPG#9bG#9Qj_+tal<e8|y-ih(W
zfRG9v7;T{BhM>_dh<65tPQE@}V*VbMK{iC1iiOzxBoW_mWg7Dgp6l?73bZTTM{u#K
zJ*|K*{3fCQQ~$ogg~!#MS*Mfz9$hX0CBIS3Funv#76Y9mIL7nDi@64N8lJw3W7`IR
z^<-s7QYYvp^PJ5i)z|oW)<AdVEmGjrm1k*_sM}^N3A~jw`&>w37@GS_E~*#$^bnZt
zU+bEW7MLvs$S$~qT}ioiupFpm^qMV!`R-Sb0ZR(3V$i=3{*jgC-2LN3%ZqiPi#ihZ
zWV}3f48I~*K2TI`D!WtSEz1UL$l}06q(UvcSO%&_7jMzcH75N11}HcG0PcC}HXwIq
zB=5t4hW4MOsGEGLwF39g?4xSp4!$L_vAeS7W^wAKa`M@EAAR^vs@{pWSxcDM=zW-d
zJ|Zt!s=lS@(T4!vY^%BeMbFk?#w1I<t&kJp*;tXMGOf3)Mi1ohaq0%0P5UBkcIr!&
zW|E}PTW+U{9)oX*ck6Mwt-l{~nf{f2d(cVs*R`*ucfF3kHc&V%`={9RK5__umZ<by
zs5DUqNy)l+8~Yzt5Rl`zf=^Tb!=C#elFh$_t|lVcnGjpiJ9AF<US;EmS`F{c{)cxn
zkr5Rq(~+b5kOZPE)kl<E65C4O@79a#w}`T1tec2uQ9VxGfxg0)gL0IMjevL)xS)r@
z_vnD~tIs~l>?Ui+;iS=uw~bTCMmGsPrpo8aGP!igoQ4PgJWjgbHuQl(rB$nT5>ji6
z6)=-vKM4AaLep9ZH<fo=7&LM&nS#_fU!6M1#~K9|6pnnh{a3`gde}J7nms13c2_sf
zWdE;tci2wNJ9For9Sk?q!90cULety)NYe4r5V69OB-_VlTGpfyD~z5)3~;1A_pbSu
z4>;qu2W65|j8DtnQ8=4xX8y(Mu|Y}c-P-=NM~+`#=AG~S{gB1jBPF|$J!e)~H$p(K
z({cBLK0VpczQc&cw3{7r4l*@8PBuTp<*O{&A|?9%u>a)LP49K^fdHk)RYvSa!-FSb
z%d4T};aHIbfb58g!t2=i`;(LFP=J363A@;|daUwF!<5$nqu1}(`91ZMpGl3w7dy3T
zLw_KtBE<)>j<%`<`!mQ;oAJblhaarAx6%Cjy*VT}UqFUojDae~ekYMuXyErvJr>qh
z1tDyLwZ?eoZR=R>p!EyFtjuG2C-2PDE{V}wvY{mk$>yAL`>d%3p_;3!h*8%bJ^rdq
zZy{rlJQKOQQ0y|I+wI-G?iV(<epkN<XK>o|BCoXj=-mN`{U3k5O23@mT6auEOEG<F
z-m~hq6URIjqZLZQr|g}WT`Ao6P~92YX~w`0%b>F(A3(;5NvHr_4m72Fl0jCm{5lZf
zy*hgg0;a3d)EzA2d83`|DR6j`uNt7X73I^E9iW@WDK&oy0qQ(m`E4eOW?&%g<`Ac<
zT3gQ*QHi!5&3lMIT|mSoABna#r2=oMBXxQ2q%x0LC|GM~$-GlI5>k8H^6rJ2e!94A
zMQ>R#xoS=ijm3{mTy8piS7HP8XMX@c=u_)}oRzajfsDlhu)JE6YUatUCJ8tU-_?T8
zU7zH;z#jAjBn9%1#k@jN{u>9l{i2Ku4W5j;7+q`vEm8ta;FHu5l*2nIUiz^$Y8|B*
zUW;Vbjn>qd$VnquCo~TcA&fImV?mrXe;a5~V4Z<%g(5akikv$ikyF04Qw-Wh`CE8%
zDZj{hBZ)N>asLoAP|^+~87@z{zW=yZwuU{+{PR||Zo&VY@>lET8%dS{{xxLG*`3sw
zTL)AjLT~oJo|eVgiW?%&d5pPJ!uF^pSbwn6y|*C6I^}B1AF-~Y7~Z@iDx6)nBPzh6
zC{Y_6Cj}j^cSyv>S#!4GTNJowTy2G7HdGY#qBc}S^aM-}d_z}+YriT^+?AyjFK0r+
zvl~qih#!&4mxKR*rKzn@6I1k~_w^HkhtxTDKnw)n7(7%?bB8=VARmuk12}Q7L6v*5
zV=maf)gdJYV+z*A6~^?;DuC`p8Xf^OB=74S-i6rBs6e^k6cdP%2vh$>m7lPCpq>`+
z<br@fz&~ib)z;=fyv)^m{1G91w)|m;NMz?g*5%oQ1~-iYoQQh-A@q(Qx`+tnvlXfc
z){6FA+Tup(zE$W3eT6{_Fy4C=N3pa)LfO}Ftqjs;+1wV$r57Oz^PicIAoZO|kh=p3
zo*f>C@mEKcg9uS?7bI@zzVl$y?NsJtkzxCo<<Njcvf`&2YT;L?D6w5t|0^JwB~*GS
zmYb;@L{orU#SS#f8SoLTl7T2FqP_>%CCE%CpbXdbXcu0fC;W0O^n{<xxC#!}K)Cq~
zmTmmGlp$s1=CQpA+|ADYQ%y+p^im6Nrm6_DMBrabagSwtBXJblXQVufZIyDM8iX;b
z*vIdiI?5G0rL0jCJ8sq!@Sl`lbE$mx`;h<GxlmHMnGa+T{}&vb1tP0y>itmuZ%`ht
z^;Wm=Cg7u2b}%n;6n6mMCSNqrc`!S3o++(6Zj$qe4ufh?!mVui7cR<QFkyQX*+O6M
z@%sJlj0$g;I<Jc2C~WA;m|}~BKzy(~1&<uHF^}GmTptCJTs!j<o#1;sl|!JEhEY~2
zycE+JmC!U8tj-wG9hH@3<uOH1^ACIllkTGu1|{I<!GMpn)mccHX!)uP#wXFl7q8jA
zGh`Cm_zU`s9}PhXMWF8l88w&(ESvpP9nS4}g5?^klm7nf*!ilwtM17(i7YGGG<2f%
zY3t`wy$r#<ZGCRU+Fvy_u()Yx(Bk2fdS{)e#@_&52saGjsOC}=`U}APYQUp^T_V%}
zbAV90{hw`~0Cwy$X4K4;6kv}JByO=~QDJlQiC=mI7cqSY+Yb&MUI+-YZir5NuYyJn
z&s07!yLm1d?(je(>rO4kt%jHTf})`h^#M4Ux8#(0BC^x7-VxD#yGQiy<I`p$&Q=9q
zZS#r)8=K;@>vI+**j0S|(;}~5BVrmm59{ZwqyFsVZc-%|F2MHGGjLKiY%jnlAN21c
z+5i^JP>cZSDHoL1P>dItIKYK=kg21$jv6P2a+ls3%g$HvEX|EzyLEgihJ8<3#!R48
zFeOft^D?dF$U$p=+x;>7QHs|>*40n3k9<D*T#DDm4+&mz>Wb`Y8?PfL720&PHQChT
zjOGa$brVS~{TqAJ!PjQg$vKoXehLbst=8o|1=#QaQ@v-T$p4~`#_1fEbo7z!Col`d
z=1J`|lTyPUg>rV4_u=$@Ps?2MR_0-DAB)p04q3L{f7(ljwf4_<V@(Epp7)WVZ_gC>
zR_3QizASt!nU$IPw^eHX!!t$p%G~tuJ)aMrG;}>%nT_<yIZ9>)7^dV_x+RMldeypF
zH}+XX$vNbHsF|_3IC^&e1^t@z^hjPV2NBt~>|q@E?vsu%%8O!L3HpX^g8YFA7(W~j
zEUq4VRNX{{el$>_;qy_ABq%_RWcd-TXOITVXLV)-5zPA)M_rDtPXOdqX}T<$Iv_!_
zcvD`OS(IwXTKgMoyWe>UCCsc3;*G{%w5&_>OssKai6l@_t7NZuU|P5g;tT}nCfe3O
zCtk><SKTH+z4l-u4ZO4a+6XWN&}R<Ws@x&?k}%GZ%(Y*p(owhf2kk(LuaHn8+;AMm
zaNa@$`?0T^M}@VHA|GGo@U%)P3<QvXpc+TY5+Dsi*FlkAGWcYB0bIXpF^eiX7B`n5
z(+`E@c5Hy*S=6frYFMr}@M?lMDm+yuuq^78wJmMZa|xklH7xZh96^n1XrU+pcm06c
z$(?-)caIJ-3!Q$8VF(iN&=1g@BJES%e5HQK-}<m$3MU1-ZE9Lo@FcR6*iL&)gMM$5
zv8l12qQDz8J?W`0*Y3BFIFkidFNLfE(fz4;popHPAS>uWrT?dGj}N?i0|$777er^i
zF{6$Yv)1fW7;2b99?CJ!EX0w9a{%6|7jWhDyqye%5-s1*R-JBaYPVAf5JC;G1D-h~
z?4|jKz@1W*+rjaq+lKHms`SmRRwGea&T|UhdpOUXp$kPsECmadfL|pte}VHB2qGEQ
zk*44|1TZ|1X=dWZlhFqwrRHzo2N~tl!*d=ua#6UpvvQVH?gz8$rJp&L-t1^y%$tF{
z-ln8lWJ8;`L`{u!@hjTf7&~_7zp#A~xXAB=CW2y6l3sNWNf3C=9(}(IOVX2}A~oL9
z3R=-a|E)<9<O}sv&h(Dx2z*y!(9OSkwYXGZWDm<In=q@>(NqSecFradkr(#jBP$r&
zcv`nHP$>nIvEV>q2e@}NAe1bZmHOW|&ApYyNwlm>cWs0*C7{?q#tzgb@;ZQQFSU2l
zuGYoFU-_&qJ_fA{3a&vz5NU|@ZtOBtO)hHM^djQL;QMFkNJTypM?GO0_E&C4F@OtN
zcM)yh4V3+o*UmXdBF@XP>HtMH@=oh7r-JQf;+8=`ZINrMg4><xPc&|O@hDn`3=;*^
z!~W$o!c5^n)!@^6zi|GLpBF?~e)vF9>N#-ID+Xa?{A`d637DMt?G^_C!`0oit`y`V
zzCeN|?H-K1t8w&^w*OH4=*aQUm4$r^f;>8%4sx7j`yBRg{^3<-WPKtb_DgrKy_x!J
zSFq5UDNO`!2Ey6xaJLDwL&9RX@YBC^V2z0Pq`JCO#4{skxql(HZRrzB*q~ItOxpIJ
zVKlk1c3@u*oM7E!yI+j;-bt;Hb&K@Vjj&D?J3^9jY|Xq4Fm9)I)7imW?GM3V?6ttg
z|G)w!>Xf3J6zmPy-jh<$uS_X$qqsO7@3V{H=5=+zwklN6B<gQ}qUF>D?2C%#P~wcc
zD5%Zos#wOE?L{26tq8Yzngc9<KI4!BK36Nifl;cfy&r-MM9QusILR1Pbpv25K7j7o
zMNrcH9`O9^h!Et#Id(#_3s>^7!RvL#KZVDrI$?nZMllzzt~;ks?xyf54xrN-PBpRv
z+oyPc^~-cX%m0f7SS+C)0!e+yrQ4>@GODL3fBJ6sK~|T;@9S&66&O)mlbp(Z(3eYr
z(r4^%q9p<?*v$lja+#wMG)v8Z&#7|%yC4)zL3tFV<nUwI)PTa}wVbn5u9aGj`z$8E
zH&jKW{ZkXyDo1)-MaGFGK%K>!GE>Bf*z+D0K(KNXD*6w&kNOX|Unt_~PH~-dT$4`)
zm>=?r#X_nR>LQq!L7&zpiI44``lcyy&%`C<rleMw_Kw0OeW<=Ata=CN%bAAm#Ld5n
zli7rV>1#CCyq%K>m=V;Xf)?f@U`!OhUvPyYAN*OGXjv<aEJ{;5X6MTeMCcA#(h->8
zVUEEK$b^V3I6nTSH@^e=t=j=9l4eowniYU_6)oc&rFFangSHs>-Oo(%Gf%nv6~Cu)
z;vrV>tu<{m$#T`zD+%|{6rX>NYWDLj6LchL8Mx(;wpwAQ!9O77+zBZn#KKsV_U{I_
z$A7%g(leh`a2FCF{09;+awFK(VWcXi50Lq6eBV?zn<#K)BN@pBJA|5YK>(q;ksmdO
z2v<6HhX{{VnBGsGV0++Utuba)?56zXctU&xQ(>=n_?P4AO?|dp&r_!JzJ*><J8777
zHto0c^R_P&$p)=`k48tAGu*rp*1&`g!W!tkc><|+=Uod&2mz<#C;yc&*I%w6b5a_R
ziaGfnKT_|;b0GC@SkHY+cy<8k`Qs>!T#zHsIbdIWQy1Y1a7Q2vg1dRy$OVgo$beAS
z$H;*9X>`WGDM^HKaMz%7V9r4<x3Ec(i@eS?d;dB;^QB?CD)xi?E41=IANrv$$HuS5
z=8u5ZlPkb>aN~zJRetK-DV`os^@J*C&E|5<Kk0y_P|CvR!jsnhfx!jeko|`%D3w78
zt^SrnX__5bO&?!na46Zb3kW{i2^`cDs^PPu8Z<MymrqXyRHWv&ur`d|*AKlS{Npp{
zhLeK7<WzuTl!AZmR6uuney3X(r)j!oG*{Cq`q2Jn<g3~a?_Rxr#1LcBUx+X&SP0I$
zqc=P<)br;LCm<#K*=G8f+9@wZ?v|L^o%Cp?$tM}=GRfjdJ}t(EJkBrFxc&$EwB`t6
zt5~{Lm~DFZ8b&(9P}852{<yUt9$RL$FPSW2W*BeZ-`VaUFi<I=aTciIB0%<U7pm*+
z*GRbv=IFT@R%v58JPB%zA2omv9yJloNSSX07!%lzRnFBD*gBtIgsk2qE*SmT^I%T1
z$RKnVlXn)}VHUawX@m;jx;@|9s({q@*qK<hu0RH%*Qa?}RF6alQxhW-S`a#e2947+
z!TNAae%leO;0(_vRW}`F(|4Mo@^r2PFP%KmvQ6iOG>-pMYT39-R}{A+1cv9;yX=T<
z;{UdH@BA96DYE%KK+HgUci*pU_hUoc|FlMM9$Ui_8*Ivl_7Obj<ct5jLUmYfH}9X?
z-J7>;BLs?PUm1=<N2Y2f;^ue|m_Q^H^6)A*nXD&<LDU8LM-<a~S@P=jWANauDmhs2
zLpsR(C_})S#z`r5`va6Wy9$g7m|g&4NMUOSjI{rB=z*x3y%DU@bG;wHje%PyoHrOH
zjFdG(H6!2!1@B%4gl<Ph%=CXX1SCw9HJYyTb0JpHL<vhGqY*YH&GwM0fKdiMR8^IF
zyTu^x+HU@h-~z^FQ&5zQ^D6X_zj8PK?lYO|@%RsOsOX<FX}8KY4dVrN?y*EH^HB@J
z;~i=`_a4`OgH{q)5(s$^!x<#&_l_Y*cKJj7UFU^%W}y%yw~o{%i3&G;phj^{xxv~^
zdAq_L)~?UXlDL&2F{tqTAd;wak7r-BD74-<U0Hf`s%H8fddhbr?__sIpk4=<Ne1?M
zG2lF|3yQxX2$zQkc=gO-z!`dZd`_to2x%ly#oNSbiSNam2@Y_#kK@&culf(Lijfgo
zYM6hIb<cZogsL>pCdS%HujkW0DFdD0ym33p=@|5(i$JHjR|@(#HBLgQMPDHu7h`i3
z`7qUw6alAdti>+66n?N^yXaYH<tKic{C+Cn7;m9^y|vhVzqj6f`-OET^sk@~{d9a4
z&8N>H-f3^T{s%uN4*fp@0R^REpuBJfG$BC1;$CmI`_2CY1faTM4C5x>EC^0|PI6<Y
zT)eFsP{{v#?39VoH3{bf({-xKhmO=+U1*T3>32}o2w}|N=aoYEq4kW4k(zYThzRZ_
zuEd9thj;9UMO(JP=HhrFQ1JuM66mY!fGn>b>;7`51t2G9B>$H|@8<tVD&xlhmaUfW
z)<IRD=9rTwo<%3=rC75?G-QIDI-!f{I-#2Fq<IBzHDtbX2?+c5KaEccZaJKrc-uqQ
ziTH^%j}5$s-2<xOaRzCydV38ZScjzwRtlyQtX?m(-}0bW3e8}EmW-_hm`q(NDhgHO
zl)?`&44#QaU6Z|a;b@G4$P5o1HeyqZyh36UMrxyo&?Yenu-ZO50v)NBN<<+v&Mt8z
z7{o9X8Mdez9oCE{>?gh2*jrD=;4BE;c-Ky+ue6^ooCMbrBAj;OJ8&5Xb~un5B2PcA
zZFDGZ!{ta^+JSIf&n)TihvX6+1jzQQ4k2vuspGW?a-kA~1D+@QBnGvoR(~j*oh7s?
z<Q^rbGYS2A#BGnspv`?}$X>!OBPG1Qq}9))KICMzKQk>y=5o6uolY<e_%yq$0LzlI
zZvc$>zNB(`;Ks14s{O&c9c^Z*)h|kVlg;EJwU<?dS>-qo*uq!4>T(VNEN~xr1NgMr
zC`B{YaM*2ni7d|Uj!Zv5jtIpdXUZr72k{w~)&U)A*lTU2akyTKFcQ#6{H>N=I>h>*
zNE2v=YI^u1jy|UoMsFmGc$f6%RJjq)dzTzDwi^dAGIO<_mQxw|W>eUBf_NmyI%cD6
z84$jWAhnm8L&38C|I2PXHrYr+;M_VG<^Rt92mFUK5~Mj)Tk2Ff20%)MXBe-89BX6Y
zO`5V0Q|k8nRXX(zLa1;LUU3JlP;=@tw=N^gZ%##w6;IM{3fx+Vj|f_&A|$~J*ag*{
zE?0ykNJx!FR1#BCzQ{1i&wPZb$h*HX)y}8^n;39o12sAaeJ%jJ25-r$Cd^XUe}AiY
z1Ctq5{vJp&*n|q=fn^9*6ZLk|)f@ZP_}k~-q<!i0gC7v8r8z?NM#2#1-=eFbjB-3L
zF77tNkOhCdejMpN6h=@{Gzg)KD1yo<?vc1U6T6DzA^GjaIYIdyZaVCysh+njRod|A
zk)MP}67K%Hx1e1Ke8+%<oC3i2fKmtk#J+l0a0Xb)B7ZZf0D8pb&<fHgVjfTKKyQSA
ztCMN|X5v$%Wf-POpqs`LL9b^4LWjT}`EOA2*HnzGrLZ6dL6`{K>TpDKPz?AmH9mz!
zh?M_TMcS;dTcQuCvOZ2jh1SalqH;+1k!8bR63-Z#YN(1I18OEK5R&pyzm-tVlPu{0
zQnG3r?G8)Y@U2af@b7-h?_<MA%wvN24BiiJR{vQjgoItt({-6Rrg?>VBEzxBmjxv~
zHu2Uy?uN*m`7nsFifwC!)t0mMxPNfm%$hpZ-=3^(NCebH>_;%@2a32!m}i`~*MZD6
z+fD;{{|5jmyPgJ6WqIp3IO)5Mq0;bXfU8~k7A$(-T@$$Ew|y)j@kXBrhEqASZUX(#
zGoOG2!lD#iP&mAWw?&?ba4=5%%|w*rh*H%^2=xzJKViR&oooBJgOqw>;yx2VlQ<a)
zQJSHnL#|Qtr!1e+!wEt>n63f`X5g?oajo0Yfdsa=O4u=}KPT0b<x6D%HIO}_0zO)w
zIfF+FQORMPL2scx>$OF3MNV^AKlHi<Kxmj7u4)F8<giHk$f*;B5b_HJWB^?io=}OX
z#&MR&d=Lc@pGE49L9WpIshR@(W3;cC>V5AmrPv8-81i$gC;6h^ifK==<{)>|98JYP
z_euui?bI+6+nr~rl}`EHEx^Zcf&f?>9;JZ3Apbflva*s2!DGlE;St+D?^c%Q3Ch(s
z@Kbw--1j;13A3h)X@z}tV9*E+Jlpkbge@ZbM%n@qh4?<&kK+25CQkn`P_&k7X@@1~
z-Q^A_Z9!X^;!k0m2vL+6kx8t&6hnl}jz+3T8xPivvm8Q54<G_ro{JyqYu;b5AtJT(
zTANRfRFp1gjs?2zLqP66Qt98@)A1SC5&z$u{8%4#?l5`+RDgiRd}H`ZW(h`rCfJDC
zSf*#V8*vw+k}xZ$z}Ndi-an#tcz*PcKHG&3-Wsq_TnjFBKtv>r9ixY_X{ZjtHNofO
zNEidBX6@z2;O1@YFLAz3s_7ccl>(Vczv)OURd669#EhE+!_coaPhz5?SJM65#vtBr
zv~E*?ONKLOs3k~XEQyRkvgA*ziunIl;~^34gjyJri7|xdzR*~weB+6t9m|2ldp-;C
zeM9C#VSil}^+FKA*?~Jkyr=&5XXh=&*TB9F_{?0OxcXiImNzWOs5@~jii^}O%r1CP
zo&Umo1ZG2ch6lhzR;AJi<bBtYl@|8t6y~()W9uHWp<?9KMHE73q)Cct8~33mmV8x|
zivenB)eDERTrvOkT(*3mYI%FmYr7mTH+wQDo`?@7Ls>Yocow~|7_RmCB(fLpW`K&O
zG(#lZ5>e}l@~fQFA8lj9g5&~~dcM2M{doR0L^DAXiR;2zVVX{X^9udunPX;emBxqZ
zIle87bp=;bLGQOA;7NxL9-sREpaT~=bRez(O1{ya^@aW@o$EC}VzJL8S3g~s(*vQ!
ze%C-E)4^BVR|;9!?Zr>zHp~GuTWktAn*MJIdv2L3U+_hi2ywkGHTv=pH1ZNHXG%aX
zdu$y@K&tY?Ov4$?w6MMm+s6WxZ1LPd6dpy+O5Fg&$9Rv_JD`)0Q1YZ(STtk!0>$~s
zwT$w#ci`xQDwZTv^@o1&cO1-2z|AeU5JR9>UAI&Q3LED+@8Xrj#X#Tzz4gt>0`?LS
zxo3EFziK>nbrmeS0ut8W!Ed2+uxOb{q}z8uZGi|N6@{A5EtR<Qn7B8a8Hzz~h6*Kp
z-0$#aViaT;B_6HE+j96!@+mje9n`*d<%{zYU#$*1`UFxttTey0^0;wN;adk_e;Yb5
ziF*DGv`O{;h2GmOAjlK~%?%Yc*cZ}J<E>*7qY;3z+`bG?F&B5P<o>C-cXGAKodOJW
zz7Oc^S6Vva{4$HSVwsAgjl%G@j?8HC%9QAr`&n-Wa6}AghMb!UJ^SQ^1$I*_aw1ON
z1jDstV)N&<=4&1=I`;o5ql|SPAo47)-X}1Qo)hs|r>u^Qd-<<bJzlFK?^Ml;d*y_8
zbrMbYlWTJwX4h2KQuRowD;LeL3=TCCuN?N`7vnU)f?z3paoszj-etkhx^*NPAm=@?
z8(_IB3{S%<{~HHU(l8{jfSukb_%qr$xCgfUerDjcAc2P6_)sm%Hp?5p)RGxIoo0W>
z{*s68tjzoF^N4+Y3;2rXjJ`ivwCwJC)s(`Ky!xQUcso<q_WbOE59&mGu4}hULtno6
z?SrEh{k2y6R`w|r>pOPzZFe0s1S2OWz?=%GKDG&+Jp+0{dGlY`GS7rwh6qI(C2!8_
z0erO8>Wd}&2JiqMHT8591THz;)vSD_Bc>r3BZ3$5LLGnDIKS{yQ${=eiO(~4L+{|g
z^LIS|@Nsvkn;KrDvn;R(Nv9zX))^uW>Qi&uV65lIqO}jtk5=uOy}AT3okxLuJ=uxW
zVW!y&0aw1fZD!oe#9ZG{Iye4nHO{sMM{rmvdGGh8Sz-pQP^{+P5s<Hx`V;ex5hIO;
zG3#*AA`rMNJ`1tC=+eNc-#(Uq(aER_@7~!1(TjXUv-M@H@?#+e&cJs>EXzKlhItk%
zPxz5RySCB5553Xa*Wst#8yPAa1nftDkK1dx>4p(_BV@YLn~8(hWeUn)t4{9?io#?t
t1T7aKKq(=}-u&SB@Z{|Xe(75uuf-j^8vg&(Ff)cxF<N_GFw8SC{2u}X<V^qo

diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_base.log
new file mode 100644
index 000000000..e1f62dcda
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_base.log
@@ -0,0 +1,28 @@
+2026-10-09T02:44:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:44:37+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:44:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:44:52+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-engine -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label base-r1
+2026-10-09T02:44:52+0900 sample 1: clean gpu=[]
+2026-10-09T02:44:54+0900 sample 2: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:44:55+0900 sample 3: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:44:57+0900 sample 4: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:44:58+0900 sample 5: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:00+0900 sample 6: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:01+0900 sample 7: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:03+0900 sample 8: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:04+0900 sample 9: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:06+0900 sample 10: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:07+0900 sample 11: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:09+0900 sample 12: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:10+0900 sample 13: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:12+0900 sample 14: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:13+0900 sample 15: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:15+0900 sample 16: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:16+0900 sample 17: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:18+0900 sample 18: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:19+0900 sample 19: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:21+0900 sample 20: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:22+0900 sample 21: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:24+0900 sample 22: clean gpu=[1130754:mlxcel-bench-en]
+2026-10-09T02:45:25+0900 monitor: 22 samples
+2026-10-09T02:45:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_new.log
new file mode 100644
index 000000000..9924859b7
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r1_new.log
@@ -0,0 +1,25 @@
+2026-10-09T02:45:30+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:45:30+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:45:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:45:45+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-engine -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label new-r1
+2026-10-09T02:45:45+0900 sample 1: clean gpu=[]
+2026-10-09T02:45:47+0900 sample 2: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:48+0900 sample 3: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:50+0900 sample 4: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:51+0900 sample 5: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:53+0900 sample 6: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:54+0900 sample 7: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:55+0900 sample 8: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:57+0900 sample 9: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:45:58+0900 sample 10: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:00+0900 sample 11: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:01+0900 sample 12: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:03+0900 sample 13: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:04+0900 sample 14: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:06+0900 sample 15: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:07+0900 sample 16: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:09+0900 sample 17: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:10+0900 sample 18: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:12+0900 sample 19: clean gpu=[1161150:mlxcel-bench-en]
+2026-10-09T02:46:13+0900 monitor: 19 samples
+2026-10-09T02:46:13+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_base.log
new file mode 100644
index 000000000..19bd89c8d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_base.log
@@ -0,0 +1,26 @@
+2026-10-09T02:47:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:47:12+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:47:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:47:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-engine -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label base-r2
+2026-10-09T02:47:27+0900 sample 1: clean gpu=[]
+2026-10-09T02:47:28+0900 sample 2: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:30+0900 sample 3: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:31+0900 sample 4: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:33+0900 sample 5: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:34+0900 sample 6: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:36+0900 sample 7: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:37+0900 sample 8: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:39+0900 sample 9: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:40+0900 sample 10: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:42+0900 sample 11: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:43+0900 sample 12: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:45+0900 sample 13: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:46+0900 sample 14: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:48+0900 sample 15: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:49+0900 sample 16: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:51+0900 sample 17: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:52+0900 sample 18: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:54+0900 sample 19: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:55+0900 sample 20: clean gpu=[1217341:mlxcel-bench-en]
+2026-10-09T02:47:56+0900 monitor: 20 samples
+2026-10-09T02:47:56+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_new.log
new file mode 100644
index 000000000..8f463584e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r2_new.log
@@ -0,0 +1,25 @@
+2026-10-09T02:48:01+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:48:01+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:48:16+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:48:16+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-engine -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label new-r2
+2026-10-09T02:48:16+0900 sample 1: clean gpu=[]
+2026-10-09T02:48:18+0900 sample 2: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:19+0900 sample 3: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:21+0900 sample 4: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:22+0900 sample 5: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:24+0900 sample 6: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:25+0900 sample 7: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:27+0900 sample 8: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:28+0900 sample 9: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:30+0900 sample 10: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:31+0900 sample 11: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:33+0900 sample 12: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:34+0900 sample 13: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:36+0900 sample 14: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:37+0900 sample 15: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:39+0900 sample 16: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:40+0900 sample 17: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:42+0900 sample 18: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:43+0900 sample 19: clean gpu=[1245919:mlxcel-bench-en]
+2026-10-09T02:48:44+0900 monitor: 19 samples
+2026-10-09T02:48:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_base.log
new file mode 100644
index 000000000..cd339c9d7
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_base.log
@@ -0,0 +1,26 @@
+2026-10-09T02:49:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:49:43+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:49:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:49:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-engine -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label base-r3
+2026-10-09T02:49:59+0900 sample 1: clean gpu=[]
+2026-10-09T02:50:00+0900 sample 2: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:02+0900 sample 3: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:03+0900 sample 4: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:05+0900 sample 5: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:06+0900 sample 6: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:07+0900 sample 7: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:09+0900 sample 8: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:11+0900 sample 9: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:12+0900 sample 10: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:14+0900 sample 11: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:15+0900 sample 12: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:17+0900 sample 13: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:18+0900 sample 14: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:20+0900 sample 15: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:21+0900 sample 16: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:23+0900 sample 17: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:24+0900 sample 18: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:26+0900 sample 19: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:27+0900 sample 20: clean gpu=[1302244:mlxcel-bench-en]
+2026-10-09T02:50:28+0900 monitor: 20 samples
+2026-10-09T02:50:28+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_new.log
new file mode 100644
index 000000000..6bc98da11
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Meta-Llama-3.1-8B-Instruct-4bit_r3_new.log
@@ -0,0 +1,25 @@
+2026-10-09T02:50:33+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:50:33+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:50:48+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:50:48+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-engine -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label new-r3
+2026-10-09T02:50:48+0900 sample 1: clean gpu=[]
+2026-10-09T02:50:50+0900 sample 2: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:50:51+0900 sample 3: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:50:53+0900 sample 4: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:50:54+0900 sample 5: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:50:56+0900 sample 6: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:50:57+0900 sample 7: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:50:59+0900 sample 8: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:00+0900 sample 9: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:02+0900 sample 10: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:03+0900 sample 11: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:05+0900 sample 12: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:06+0900 sample 13: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:08+0900 sample 14: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:09+0900 sample 15: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:10+0900 sample 16: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:12+0900 sample 17: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:14+0900 sample 18: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:15+0900 sample 19: clean gpu=[1330719:mlxcel-bench-en]
+2026-10-09T02:51:16+0900 monitor: 19 samples
+2026-10-09T02:51:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_base.log
new file mode 100644
index 000000000..9a74ed523
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_base.log
@@ -0,0 +1,12 @@
+2026-10-09T02:43:42+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:43:42+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:43:56+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:43:56+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-engine -m models/mlx/Qwen3-0.6B-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label base-r1
+2026-10-09T02:43:57+0900 sample 1: clean gpu=[]
+2026-10-09T02:43:58+0900 sample 2: clean gpu=[1101389:mlxcel-bench-en]
+2026-10-09T02:44:00+0900 sample 3: clean gpu=[1101389:mlxcel-bench-en]
+2026-10-09T02:44:01+0900 sample 4: clean gpu=[1101389:mlxcel-bench-en]
+2026-10-09T02:44:03+0900 sample 5: clean gpu=[1101389:mlxcel-bench-en]
+2026-10-09T02:44:04+0900 sample 6: clean gpu=[1101389:mlxcel-bench-en]
+2026-10-09T02:44:05+0900 monitor: 6 samples
+2026-10-09T02:44:05+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_new.log
new file mode 100644
index 000000000..62d8bfdae
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r1_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:44:10+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:44:10+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:44:25+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:44:25+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-engine -m models/mlx/Qwen3-0.6B-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label new-r1
+2026-10-09T02:44:25+0900 sample 1: clean gpu=[]
+2026-10-09T02:44:27+0900 sample 2: clean gpu=[1116524:mlxcel-bench-en]
+2026-10-09T02:44:28+0900 sample 3: clean gpu=[1116524:mlxcel-bench-en]
+2026-10-09T02:44:30+0900 sample 4: clean gpu=[1116524:mlxcel-bench-en]
+2026-10-09T02:44:31+0900 sample 5: clean gpu=[1116524:mlxcel-bench-en]
+2026-10-09T02:44:32+0900 monitor: 5 samples
+2026-10-09T02:44:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_base.log
new file mode 100644
index 000000000..480427df9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_base.log
@@ -0,0 +1,11 @@
+2026-10-09T02:46:18+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:46:18+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:46:32+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:46:32+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-engine -m models/mlx/Qwen3-0.6B-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label base-r2
+2026-10-09T02:46:33+0900 sample 1: clean gpu=[]
+2026-10-09T02:46:34+0900 sample 2: clean gpu=[1188757:mlxcel-bench-en]
+2026-10-09T02:46:36+0900 sample 3: clean gpu=[1188757:mlxcel-bench-en]
+2026-10-09T02:46:37+0900 sample 4: clean gpu=[1188757:mlxcel-bench-en]
+2026-10-09T02:46:39+0900 sample 5: clean gpu=[1188757:mlxcel-bench-en]
+2026-10-09T02:46:40+0900 monitor: 5 samples
+2026-10-09T02:46:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_new.log
new file mode 100644
index 000000000..15e3b009c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r2_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:46:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:46:45+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:46:59+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:46:59+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-engine -m models/mlx/Qwen3-0.6B-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label new-r2
+2026-10-09T02:47:00+0900 sample 1: clean gpu=[]
+2026-10-09T02:47:01+0900 sample 2: clean gpu=[1203050:mlxcel-bench-en]
+2026-10-09T02:47:03+0900 sample 3: clean gpu=[1203050:mlxcel-bench-en]
+2026-10-09T02:47:04+0900 sample 4: clean gpu=[1203050:mlxcel-bench-en]
+2026-10-09T02:47:06+0900 sample 5: clean gpu=[1203050:mlxcel-bench-en]
+2026-10-09T02:47:07+0900 monitor: 5 samples
+2026-10-09T02:47:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_base.log
new file mode 100644
index 000000000..76e93a976
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_base.log
@@ -0,0 +1,11 @@
+2026-10-09T02:48:49+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:48:49+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:49:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:49:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-engine -m models/mlx/Qwen3-0.6B-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label base-r3
+2026-10-09T02:49:04+0900 sample 1: clean gpu=[]
+2026-10-09T02:49:06+0900 sample 2: clean gpu=[1273600:mlxcel-bench-en]
+2026-10-09T02:49:07+0900 sample 3: clean gpu=[1273600:mlxcel-bench-en]
+2026-10-09T02:49:09+0900 sample 4: clean gpu=[1273600:mlxcel-bench-en]
+2026-10-09T02:49:10+0900 sample 5: clean gpu=[1273600:mlxcel-bench-en]
+2026-10-09T02:49:11+0900 monitor: 5 samples
+2026-10-09T02:49:11+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_new.log
new file mode 100644
index 000000000..c78b3a5c8
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/engine_Qwen3-0.6B-4bit_r3_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:49:16+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:49:16+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:49:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:49:31+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-engine -m models/mlx/Qwen3-0.6B-4bit --path both --decode-storage dense --prompt-tokens 256,2048 -n 128 --warmup-tokens 20 --csv /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/engine/rounds.csv --label new-r3
+2026-10-09T02:49:31+0900 sample 1: clean gpu=[]
+2026-10-09T02:49:33+0900 sample 2: clean gpu=[1287900:mlxcel-bench-en]
+2026-10-09T02:49:34+0900 sample 3: clean gpu=[1287900:mlxcel-bench-en]
+2026-10-09T02:49:36+0900 sample 4: clean gpu=[1287900:mlxcel-bench-en]
+2026-10-09T02:49:37+0900 sample 5: clean gpu=[1287900:mlxcel-bench-en]
+2026-10-09T02:49:38+0900 monitor: 5 samples
+2026-10-09T02:49:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/gemma3_probe.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/gemma3_probe.log
new file mode 100644
index 000000000..3c3c94426
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/gemma3_probe.log
@@ -0,0 +1,9 @@
+2026-10-09T02:51:39+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:51:39+0900 waiting for 20s of idle GPU and no compiler
+2026-10-09T02:52:08+0900 idle for 20s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:52:08+0900 attempt 1/3: start: bash /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/measure/gemma3_probe.sh
+2026-10-09T02:52:08+0900 sample 1: clean gpu=[]
+2026-10-09T02:52:10+0900 sample 2: clean gpu=[1378646:mlxcel]
+2026-10-09T02:52:11+0900 sample 3: clean gpu=[1378646:mlxcel]
+2026-10-09T02:52:12+0900 monitor: 3 samples
+2026-10-09T02:52:12+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/paged_pool_past_u32.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/paged_pool_past_u32.log
new file mode 100644
index 000000000..37baaff75
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/guard/paged_pool_past_u32.log
@@ -0,0 +1,8 @@
+2026-10-09T02:51:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:51:21+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:51:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:51:36+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/test-fast/deps/mlxcel_core-480914a0f733fc77 paged_pool_past_u32_elements_matches_gather --ignored --test-threads=1 --nocapture
+2026-10-09T02:51:36+0900 sample 1: clean gpu=[]
+2026-10-09T02:51:38+0900 sample 2: clean gpu=[1358129:mlxcel_core-480]
+2026-10-09T02:51:39+0900 monitor: 2 samples
+2026-10-09T02:51:39+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..5f2d4aac8
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log
@@ -0,0 +1,11 @@
+2026-10-09T01:47:04+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:47:04+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:47:18+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:47:18+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:47:19+0900 sample 1: clean gpu=[]
+2026-10-09T01:47:20+0900 sample 2: clean gpu=[3394666:mlxcel-bench-de]
+2026-10-09T01:47:22+0900 sample 3: clean gpu=[3394666:mlxcel-bench-de]
+2026-10-09T01:47:24+0900 sample 4: clean gpu=[3394666:mlxcel-bench-de]
+2026-10-09T01:47:25+0900 sample 5: clean gpu=[3394666:mlxcel-bench-de]
+2026-10-09T01:47:26+0900 monitor: 5 samples
+2026-10-09T01:47:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..97ee9e57d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log
@@ -0,0 +1,13 @@
+2026-10-09T01:46:06+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:46:06+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:46:21+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:46:21+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:46:21+0900 sample 1: clean gpu=[]
+2026-10-09T01:46:23+0900 sample 2: clean gpu=[3363796:mlxcel-bench-de]
+2026-10-09T01:46:24+0900 sample 3: clean gpu=[3363796:mlxcel-bench-de]
+2026-10-09T01:46:26+0900 sample 4: clean gpu=[3363796:mlxcel-bench-de]
+2026-10-09T01:46:27+0900 sample 5: clean gpu=[3363796:mlxcel-bench-de]
+2026-10-09T01:46:29+0900 sample 6: clean gpu=[3363796:mlxcel-bench-de]
+2026-10-09T01:46:30+0900 sample 7: clean gpu=[3363796:mlxcel-bench-de]
+2026-10-09T01:46:31+0900 monitor: 7 samples
+2026-10-09T01:46:31+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..539d1606b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log
@@ -0,0 +1,11 @@
+2026-10-09T01:46:36+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:46:36+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:46:51+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:46:51+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:46:52+0900 sample 1: clean gpu=[]
+2026-10-09T01:46:53+0900 sample 2: clean gpu=[3380296:mlxcel-bench-de]
+2026-10-09T01:46:55+0900 sample 3: clean gpu=[3380296:mlxcel-bench-de]
+2026-10-09T01:46:56+0900 sample 4: clean gpu=[3380296:mlxcel-bench-de]
+2026-10-09T01:46:58+0900 sample 5: clean gpu=[3380296:mlxcel-bench-de]
+2026-10-09T01:46:59+0900 monitor: 5 samples
+2026-10-09T01:46:59+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..f82c9e230
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log
@@ -0,0 +1,11 @@
+2026-10-09T02:06:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:06:37+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:06:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:06:52+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:06:53+0900 sample 1: clean gpu=[]
+2026-10-09T02:06:54+0900 sample 2: clean gpu=[4075655:mlxcel-bench-de]
+2026-10-09T02:06:56+0900 sample 3: clean gpu=[4075655:mlxcel-bench-de]
+2026-10-09T02:06:57+0900 sample 4: clean gpu=[4075655:mlxcel-bench-de]
+2026-10-09T02:06:59+0900 sample 5: clean gpu=[4075655:mlxcel-bench-de]
+2026-10-09T02:07:00+0900 monitor: 5 samples
+2026-10-09T02:07:00+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..67cc4e352
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log
@@ -0,0 +1,11 @@
+2026-10-09T02:07:05+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:07:05+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:07:20+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:07:20+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:07:20+0900 sample 1: clean gpu=[]
+2026-10-09T02:07:21+0900 sample 2: clean gpu=[4089997:mlxcel-bench-de]
+2026-10-09T02:07:23+0900 sample 3: clean gpu=[4089997:mlxcel-bench-de]
+2026-10-09T02:07:25+0900 sample 4: clean gpu=[4089997:mlxcel-bench-de]
+2026-10-09T02:07:26+0900 sample 5: clean gpu=[4089997:mlxcel-bench-de]
+2026-10-09T02:07:27+0900 monitor: 5 samples
+2026-10-09T02:07:27+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..867349bea
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log
@@ -0,0 +1,13 @@
+2026-10-09T02:06:07+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:06:07+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:06:22+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:06:22+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:06:22+0900 sample 1: clean gpu=[]
+2026-10-09T02:06:24+0900 sample 2: clean gpu=[4059413:mlxcel-bench-de]
+2026-10-09T02:06:25+0900 sample 3: clean gpu=[4059413:mlxcel-bench-de]
+2026-10-09T02:06:27+0900 sample 4: clean gpu=[4059413:mlxcel-bench-de]
+2026-10-09T02:06:28+0900 sample 5: clean gpu=[4059413:mlxcel-bench-de]
+2026-10-09T02:06:30+0900 sample 6: clean gpu=[4059413:mlxcel-bench-de]
+2026-10-09T02:06:31+0900 sample 7: clean gpu=[4059413:mlxcel-bench-de]
+2026-10-09T02:06:32+0900 monitor: 7 samples
+2026-10-09T02:06:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..e7d25323d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log
@@ -0,0 +1,13 @@
+2026-10-09T02:26:00+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:26:00+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:26:15+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:26:15+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:26:15+0900 sample 1: clean gpu=[]
+2026-10-09T02:26:17+0900 sample 2: clean gpu=[515159:mlxcel-bench-de]
+2026-10-09T02:26:18+0900 sample 3: clean gpu=[515159:mlxcel-bench-de]
+2026-10-09T02:26:20+0900 sample 4: clean gpu=[515159:mlxcel-bench-de]
+2026-10-09T02:26:21+0900 sample 5: clean gpu=[515159:mlxcel-bench-de]
+2026-10-09T02:26:23+0900 sample 6: clean gpu=[515159:mlxcel-bench-de]
+2026-10-09T02:26:24+0900 sample 7: clean gpu=[515159:mlxcel-bench-de]
+2026-10-09T02:26:25+0900 monitor: 7 samples
+2026-10-09T02:26:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..f4da0a005
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log
@@ -0,0 +1,11 @@
+2026-10-09T02:26:30+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:26:30+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:26:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:26:45+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:26:46+0900 sample 1: clean gpu=[]
+2026-10-09T02:26:47+0900 sample 2: clean gpu=[531413:mlxcel-bench-de]
+2026-10-09T02:26:49+0900 sample 3: clean gpu=[531413:mlxcel-bench-de]
+2026-10-09T02:26:50+0900 sample 4: clean gpu=[531413:mlxcel-bench-de]
+2026-10-09T02:26:52+0900 sample 5: clean gpu=[531413:mlxcel-bench-de]
+2026-10-09T02:26:53+0900 monitor: 5 samples
+2026-10-09T02:26:53+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..33907e75e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:26:58+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:26:58+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:27:12+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:27:12+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:27:13+0900 sample 1: clean gpu=[]
+2026-10-09T02:27:14+0900 sample 2: clean gpu=[545723:mlxcel-bench-de]
+2026-10-09T02:27:16+0900 sample 3: clean gpu=[545723:mlxcel-bench-de]
+2026-10-09T02:27:17+0900 sample 4: clean gpu=[545723:mlxcel-bench-de]
+2026-10-09T02:27:19+0900 sample 5: clean gpu=[545723:mlxcel-bench-de]
+2026-10-09T02:27:20+0900 monitor: 5 samples
+2026-10-09T02:27:20+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..c505ef67b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log
@@ -0,0 +1,44 @@
+2026-10-09T02:02:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:02:31+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:02:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:02:46+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:02:46+0900 sample 1: clean gpu=[]
+2026-10-09T02:02:48+0900 sample 2: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:49+0900 sample 3: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:51+0900 sample 4: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:52+0900 sample 5: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:54+0900 sample 6: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:55+0900 sample 7: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:57+0900 sample 8: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:58+0900 sample 9: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:02:59+0900 sample 10: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:01+0900 sample 11: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:02+0900 sample 12: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:04+0900 sample 13: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:05+0900 sample 14: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:07+0900 sample 15: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:08+0900 sample 16: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:10+0900 sample 17: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:11+0900 sample 18: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:13+0900 sample 19: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:14+0900 sample 20: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:16+0900 sample 21: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:17+0900 sample 22: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:19+0900 sample 23: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:21+0900 sample 24: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:22+0900 sample 25: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:24+0900 sample 26: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:26+0900 sample 27: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:28+0900 sample 28: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:29+0900 sample 29: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:31+0900 sample 30: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:32+0900 sample 31: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:34+0900 sample 32: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:35+0900 sample 33: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:37+0900 sample 34: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:38+0900 sample 35: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:40+0900 sample 36: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:41+0900 sample 37: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:43+0900 sample 38: clean gpu=[3944782:mlxcel-bench-de]
+2026-10-09T02:03:44+0900 monitor: 38 samples
+2026-10-09T02:03:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..23e9423d9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log
@@ -0,0 +1,45 @@
+2026-10-09T01:59:53+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:59:53+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:00:07+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:00:07+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:00:08+0900 sample 1: clean gpu=[]
+2026-10-09T02:00:09+0900 sample 2: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:11+0900 sample 3: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:12+0900 sample 4: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:14+0900 sample 5: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:15+0900 sample 6: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:17+0900 sample 7: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:18+0900 sample 8: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:20+0900 sample 9: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:21+0900 sample 10: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:23+0900 sample 11: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:24+0900 sample 12: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:26+0900 sample 13: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:27+0900 sample 14: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:29+0900 sample 15: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:31+0900 sample 16: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:32+0900 sample 17: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:33+0900 sample 18: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:35+0900 sample 19: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:36+0900 sample 20: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:38+0900 sample 21: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:39+0900 sample 22: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:41+0900 sample 23: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:42+0900 sample 24: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:44+0900 sample 25: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:45+0900 sample 26: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:47+0900 sample 27: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:49+0900 sample 28: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:51+0900 sample 29: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:52+0900 sample 30: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:54+0900 sample 31: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:55+0900 sample 32: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:57+0900 sample 33: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:00:58+0900 sample 34: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:01:00+0900 sample 35: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:01:01+0900 sample 36: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:01:03+0900 sample 37: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:01:04+0900 sample 38: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:01:06+0900 sample 39: clean gpu=[3852117:mlxcel-bench-de]
+2026-10-09T02:01:07+0900 monitor: 39 samples
+2026-10-09T02:01:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..02f99aa8b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log
@@ -0,0 +1,45 @@
+2026-10-09T02:01:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:01:12+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:01:27+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:01:27+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:01:27+0900 sample 1: clean gpu=[]
+2026-10-09T02:01:29+0900 sample 2: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:30+0900 sample 3: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:32+0900 sample 4: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:33+0900 sample 5: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:35+0900 sample 6: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:36+0900 sample 7: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:38+0900 sample 8: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:39+0900 sample 9: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:41+0900 sample 10: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:42+0900 sample 11: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:44+0900 sample 12: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:45+0900 sample 13: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:47+0900 sample 14: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:48+0900 sample 15: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:50+0900 sample 16: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:51+0900 sample 17: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:53+0900 sample 18: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:54+0900 sample 19: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:56+0900 sample 20: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:57+0900 sample 21: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:01:58+0900 sample 22: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:00+0900 sample 23: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:01+0900 sample 24: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:03+0900 sample 25: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:05+0900 sample 26: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:07+0900 sample 27: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:08+0900 sample 28: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:10+0900 sample 29: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:12+0900 sample 30: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:13+0900 sample 31: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:15+0900 sample 32: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:16+0900 sample 33: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:18+0900 sample 34: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:19+0900 sample 35: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:21+0900 sample 36: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:22+0900 sample 37: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:24+0900 sample 38: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:25+0900 sample 39: clean gpu=[3899155:mlxcel-bench-de]
+2026-10-09T02:02:26+0900 monitor: 39 samples
+2026-10-09T02:02:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..474ba7a9f
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log
@@ -0,0 +1,46 @@
+2026-10-09T02:21:04+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:21:04+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:21:19+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:21:19+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:21:19+0900 sample 1: clean gpu=[]
+2026-10-09T02:21:21+0900 sample 2: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:22+0900 sample 3: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:24+0900 sample 4: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:25+0900 sample 5: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:27+0900 sample 6: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:28+0900 sample 7: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:30+0900 sample 8: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:31+0900 sample 9: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:33+0900 sample 10: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:34+0900 sample 11: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:36+0900 sample 12: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:37+0900 sample 13: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:39+0900 sample 14: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:40+0900 sample 15: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:42+0900 sample 16: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:43+0900 sample 17: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:45+0900 sample 18: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:46+0900 sample 19: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:48+0900 sample 20: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:49+0900 sample 21: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:50+0900 sample 22: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:52+0900 sample 23: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:53+0900 sample 24: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:55+0900 sample 25: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:57+0900 sample 26: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:21:58+0900 sample 27: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:00+0900 sample 28: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:02+0900 sample 29: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:04+0900 sample 30: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:05+0900 sample 31: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:07+0900 sample 32: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:08+0900 sample 33: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:10+0900 sample 34: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:11+0900 sample 35: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:13+0900 sample 36: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:14+0900 sample 37: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:16+0900 sample 38: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:17+0900 sample 39: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:19+0900 sample 40: clean gpu=[355365:mlxcel-bench-de]
+2026-10-09T02:22:20+0900 monitor: 40 samples
+2026-10-09T02:22:20+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..f0096413b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log
@@ -0,0 +1,44 @@
+2026-10-09T02:22:25+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:22:25+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:22:39+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:22:39+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:22:40+0900 sample 1: clean gpu=[]
+2026-10-09T02:22:41+0900 sample 2: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:43+0900 sample 3: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:44+0900 sample 4: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:46+0900 sample 5: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:47+0900 sample 6: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:49+0900 sample 7: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:50+0900 sample 8: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:52+0900 sample 9: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:53+0900 sample 10: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:55+0900 sample 11: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:56+0900 sample 12: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:22:58+0900 sample 13: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:00+0900 sample 14: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:01+0900 sample 15: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:02+0900 sample 16: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:04+0900 sample 17: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:05+0900 sample 18: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:07+0900 sample 19: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:08+0900 sample 20: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:10+0900 sample 21: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:11+0900 sample 22: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:13+0900 sample 23: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:14+0900 sample 24: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:16+0900 sample 25: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:18+0900 sample 26: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:20+0900 sample 27: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:22+0900 sample 28: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:23+0900 sample 29: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:25+0900 sample 30: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:26+0900 sample 31: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:28+0900 sample 32: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:29+0900 sample 33: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:31+0900 sample 34: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:32+0900 sample 35: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:34+0900 sample 36: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:36+0900 sample 37: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:37+0900 sample 38: clean gpu=[401749:mlxcel-bench-de]
+2026-10-09T02:23:38+0900 monitor: 38 samples
+2026-10-09T02:23:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..b696cc2bb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log
@@ -0,0 +1,45 @@
+2026-10-09T02:19:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:19:45+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:20:00+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:20:00+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:20:00+0900 sample 1: clean gpu=[]
+2026-10-09T02:20:02+0900 sample 2: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:03+0900 sample 3: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:05+0900 sample 4: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:06+0900 sample 5: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:08+0900 sample 6: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:09+0900 sample 7: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:11+0900 sample 8: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:12+0900 sample 9: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:14+0900 sample 10: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:15+0900 sample 11: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:17+0900 sample 12: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:18+0900 sample 13: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:20+0900 sample 14: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:22+0900 sample 15: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:23+0900 sample 16: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:25+0900 sample 17: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:26+0900 sample 18: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:27+0900 sample 19: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:29+0900 sample 20: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:30+0900 sample 21: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:32+0900 sample 22: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:33+0900 sample 23: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:35+0900 sample 24: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:36+0900 sample 25: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:38+0900 sample 26: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:40+0900 sample 27: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:41+0900 sample 28: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:43+0900 sample 29: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:45+0900 sample 30: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:46+0900 sample 31: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:48+0900 sample 32: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:49+0900 sample 33: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:51+0900 sample 34: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:52+0900 sample 35: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:54+0900 sample 36: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:55+0900 sample 37: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:57+0900 sample 38: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:58+0900 sample 39: clean gpu=[309913:mlxcel-bench-de]
+2026-10-09T02:20:59+0900 monitor: 39 samples
+2026-10-09T02:20:59+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..602d4aab4
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log
@@ -0,0 +1,45 @@
+2026-10-09T02:39:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:39:31+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:39:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:39:45+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:39:46+0900 sample 1: clean gpu=[]
+2026-10-09T02:39:47+0900 sample 2: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:49+0900 sample 3: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:50+0900 sample 4: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:52+0900 sample 5: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:53+0900 sample 6: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:55+0900 sample 7: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:56+0900 sample 8: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:58+0900 sample 9: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:39:59+0900 sample 10: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:01+0900 sample 11: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:02+0900 sample 12: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:04+0900 sample 13: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:05+0900 sample 14: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:07+0900 sample 15: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:08+0900 sample 16: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:10+0900 sample 17: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:11+0900 sample 18: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:12+0900 sample 19: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:14+0900 sample 20: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:15+0900 sample 21: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:17+0900 sample 22: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:18+0900 sample 23: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:20+0900 sample 24: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:21+0900 sample 25: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:23+0900 sample 26: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:25+0900 sample 27: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:26+0900 sample 28: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:28+0900 sample 29: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:30+0900 sample 30: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:31+0900 sample 31: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:33+0900 sample 32: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:34+0900 sample 33: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:36+0900 sample 34: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:37+0900 sample 35: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:39+0900 sample 36: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:41+0900 sample 37: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:42+0900 sample 38: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:44+0900 sample 39: clean gpu=[953720:mlxcel-bench-de]
+2026-10-09T02:40:45+0900 monitor: 39 samples
+2026-10-09T02:40:45+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..5089cd132
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log
@@ -0,0 +1,54 @@
+2026-10-09T02:40:50+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:40:50+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:41:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:41:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:41:05+0900 sample 1: clean gpu=[]
+2026-10-09T02:41:06+0900 sample 2: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:08+0900 sample 3: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:09+0900 sample 4: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:11+0900 sample 5: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:13+0900 sample 6: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:14+0900 sample 7: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:16+0900 sample 8: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:17+0900 sample 9: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:19+0900 sample 10: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:20+0900 sample 11: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:22+0900 sample 12: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:23+0900 sample 13: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:25+0900 sample 14: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:26+0900 sample 15: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:28+0900 sample 16: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:29+0900 sample 17: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:31+0900 sample 18: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:32+0900 sample 19: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:34+0900 sample 20: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:35+0900 sample 21: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:37+0900 sample 22: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:38+0900 sample 23: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:40+0900 sample 24: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:41+0900 sample 25: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:43+0900 sample 26: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:45+0900 sample 27: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:46+0900 sample 28: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:47+0900 sample 29: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:49+0900 sample 30: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:50+0900 sample 31: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:52+0900 sample 32: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:53+0900 sample 33: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:55+0900 sample 34: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:56+0900 sample 35: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:41:58+0900 sample 36: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:00+0900 sample 37: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:01+0900 sample 38: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:03+0900 sample 39: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:04+0900 sample 40: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:06+0900 sample 41: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:08+0900 sample 42: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:09+0900 sample 43: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:11+0900 sample 44: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:12+0900 sample 45: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:14+0900 sample 46: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:15+0900 sample 47: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:17+0900 sample 48: clean gpu=[1000365:mlxcel-bench-de]
+2026-10-09T02:42:18+0900 monitor: 48 samples
+2026-10-09T02:42:18+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..cbe1b6f6d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log
@@ -0,0 +1,44 @@
+2026-10-09T02:42:23+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:42:23+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:42:38+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:42:38+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:42:38+0900 sample 1: clean gpu=[]
+2026-10-09T02:42:40+0900 sample 2: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:41+0900 sample 3: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:43+0900 sample 4: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:44+0900 sample 5: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:46+0900 sample 6: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:47+0900 sample 7: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:49+0900 sample 8: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:50+0900 sample 9: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:52+0900 sample 10: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:53+0900 sample 11: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:55+0900 sample 12: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:56+0900 sample 13: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:58+0900 sample 14: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:42:59+0900 sample 15: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:01+0900 sample 16: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:03+0900 sample 17: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:04+0900 sample 18: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:05+0900 sample 19: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:07+0900 sample 20: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:08+0900 sample 21: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:10+0900 sample 22: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:11+0900 sample 23: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:13+0900 sample 24: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:15+0900 sample 25: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:17+0900 sample 26: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:18+0900 sample 27: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:20+0900 sample 28: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:22+0900 sample 29: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:24+0900 sample 30: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:25+0900 sample 31: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:27+0900 sample 32: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:28+0900 sample 33: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:30+0900 sample 34: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:31+0900 sample 35: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:33+0900 sample 36: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:34+0900 sample 37: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:36+0900 sample 38: clean gpu=[1055761:mlxcel-bench-de]
+2026-10-09T02:43:37+0900 monitor: 38 samples
+2026-10-09T02:43:37+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..98074ebd5
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log
@@ -0,0 +1,16 @@
+2026-10-09T01:56:56+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:56:56+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:57:11+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:57:11+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:57:11+0900 sample 1: clean gpu=[]
+2026-10-09T01:57:13+0900 sample 2: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:15+0900 sample 3: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:16+0900 sample 4: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:18+0900 sample 5: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:19+0900 sample 6: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:21+0900 sample 7: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:22+0900 sample 8: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:24+0900 sample 9: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:25+0900 sample 10: clean gpu=[3752289:mlxcel-bench-de]
+2026-10-09T01:57:26+0900 monitor: 10 samples
+2026-10-09T01:57:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..aadd9d79e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log
@@ -0,0 +1,17 @@
+2026-10-09T01:55:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:55:45+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:56:00+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:56:00+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:56:00+0900 sample 1: clean gpu=[]
+2026-10-09T01:56:02+0900 sample 2: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:03+0900 sample 3: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:05+0900 sample 4: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:06+0900 sample 5: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:08+0900 sample 6: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:09+0900 sample 7: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:11+0900 sample 8: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:12+0900 sample 9: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:14+0900 sample 10: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:15+0900 sample 11: clean gpu=[3713056:mlxcel-bench-de]
+2026-10-09T01:56:16+0900 monitor: 11 samples
+2026-10-09T01:56:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..918416306
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log
@@ -0,0 +1,16 @@
+2026-10-09T01:56:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:56:21+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:56:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:56:36+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:56:36+0900 sample 1: clean gpu=[]
+2026-10-09T01:56:38+0900 sample 2: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:39+0900 sample 3: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:41+0900 sample 4: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:43+0900 sample 5: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:44+0900 sample 6: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:46+0900 sample 7: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:47+0900 sample 8: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:49+0900 sample 9: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:50+0900 sample 10: clean gpu=[3733061:mlxcel-bench-de]
+2026-10-09T01:56:51+0900 monitor: 10 samples
+2026-10-09T01:56:51+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..1833fb3e1
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log
@@ -0,0 +1,16 @@
+2026-10-09T02:16:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:16:12+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:16:27+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:16:27+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:16:27+0900 sample 1: clean gpu=[]
+2026-10-09T02:16:29+0900 sample 2: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:30+0900 sample 3: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:32+0900 sample 4: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:33+0900 sample 5: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:35+0900 sample 6: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:36+0900 sample 7: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:38+0900 sample 8: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:39+0900 sample 9: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:41+0900 sample 10: clean gpu=[193318:mlxcel-bench-de]
+2026-10-09T02:16:42+0900 monitor: 10 samples
+2026-10-09T02:16:42+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..321a1d51a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log
@@ -0,0 +1,16 @@
+2026-10-09T02:16:47+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:16:47+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:17:01+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:17:01+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:17:02+0900 sample 1: clean gpu=[]
+2026-10-09T02:17:03+0900 sample 2: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:05+0900 sample 3: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:06+0900 sample 4: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:08+0900 sample 5: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:09+0900 sample 6: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:11+0900 sample 7: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:12+0900 sample 8: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:14+0900 sample 9: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:15+0900 sample 10: clean gpu=[211756:mlxcel-bench-de]
+2026-10-09T02:17:16+0900 monitor: 10 samples
+2026-10-09T02:17:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..75cf8e9f6
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log
@@ -0,0 +1,17 @@
+2026-10-09T02:15:36+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:15:36+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:15:51+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:15:51+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:15:51+0900 sample 1: clean gpu=[]
+2026-10-09T02:15:53+0900 sample 2: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:15:54+0900 sample 3: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:15:56+0900 sample 4: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:15:57+0900 sample 5: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:15:59+0900 sample 6: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:16:00+0900 sample 7: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:16:02+0900 sample 8: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:16:03+0900 sample 9: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:16:05+0900 sample 10: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:16:06+0900 sample 11: clean gpu=[174046:mlxcel-bench-de]
+2026-10-09T02:16:07+0900 monitor: 11 samples
+2026-10-09T02:16:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..8726437a5
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log
@@ -0,0 +1,17 @@
+2026-10-09T02:35:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:35:21+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:35:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:35:36+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:35:36+0900 sample 1: clean gpu=[]
+2026-10-09T02:35:38+0900 sample 2: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:39+0900 sample 3: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:41+0900 sample 4: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:42+0900 sample 5: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:44+0900 sample 6: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:45+0900 sample 7: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:47+0900 sample 8: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:48+0900 sample 9: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:50+0900 sample 10: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:51+0900 sample 11: clean gpu=[816050:mlxcel-bench-de]
+2026-10-09T02:35:52+0900 monitor: 11 samples
+2026-10-09T02:35:52+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..75c9e792a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log
@@ -0,0 +1,16 @@
+2026-10-09T02:35:57+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:35:57+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:36:12+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:36:12+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:36:12+0900 sample 1: clean gpu=[]
+2026-10-09T02:36:14+0900 sample 2: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:15+0900 sample 3: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:17+0900 sample 4: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:18+0900 sample 5: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:20+0900 sample 6: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:21+0900 sample 7: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:23+0900 sample 8: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:24+0900 sample 9: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:26+0900 sample 10: clean gpu=[835434:mlxcel-bench-de]
+2026-10-09T02:36:27+0900 monitor: 10 samples
+2026-10-09T02:36:27+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..d217ae2e4
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log
@@ -0,0 +1,16 @@
+2026-10-09T02:36:32+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:36:32+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:36:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:36:46+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:36:47+0900 sample 1: clean gpu=[]
+2026-10-09T02:36:48+0900 sample 2: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:50+0900 sample 3: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:51+0900 sample 4: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:53+0900 sample 5: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:54+0900 sample 6: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:56+0900 sample 7: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:57+0900 sample 8: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:36:59+0900 sample 9: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:37:00+0900 sample 10: clean gpu=[853905:mlxcel-bench-de]
+2026-10-09T02:37:01+0900 monitor: 10 samples
+2026-10-09T02:37:01+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base-null.log
new file mode 100644
index 000000000..1741b9b09
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base-null.log
@@ -0,0 +1,11 @@
+2026-10-08T23:28:27+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T00:20:29+0900 caught SIGTERM: stopping command and monitor
+2026-10-09T01:19:10+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:19:10+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:45:57+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:45:57+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:45:57+0900 sample 1: clean gpu=[]
+2026-10-09T01:45:59+0900 sample 2: clean gpu=[3351173:mlxcel-bench-de]
+2026-10-09T01:46:00+0900 sample 3: clean gpu=[3351173:mlxcel-bench-de]
+2026-10-09T01:46:01+0900 monitor: 3 samples
+2026-10-09T01:46:01+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base.log
new file mode 100644
index 000000000..21eedf38a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_base.log
@@ -0,0 +1,12 @@
+2026-10-08T21:33:51+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-08T22:19:00+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 2709s
+2026-10-08T22:19:00+0900 waiting for 60s of idle GPU and no compiler
+2026-10-08T22:19:18+0900 idle streak reset after 12s: gpu=[] compilers=[1173803:cargo]
+2026-10-08T22:23:33+0900 idle streak reset after 1s: gpu=[] compilers=[1334471:cargo]
+2026-10-08T23:05:01+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-08T23:05:01+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-08T23:05:01+0900 sample 1: clean gpu=[]
+2026-10-08T23:05:02+0900 sample 2: clean gpu=[2908527:mlxcel-bench-de]
+2026-10-08T23:05:04+0900 sample 3: clean gpu=[2908527:mlxcel-bench-de]
+2026-10-08T23:05:05+0900 monitor: 3 samples
+2026-10-08T23:05:05+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_new.log
new file mode 100644
index 000000000..2f164e6b9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r1_new.log
@@ -0,0 +1,12 @@
+2026-10-08T23:05:15+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-08T23:13:56+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 521s
+2026-10-08T23:13:56+0900 waiting for 60s of idle GPU and no compiler
+2026-10-08T23:14:31+0900 idle streak reset after 23s: gpu=[3262038:rocminfo] compilers=[3262000:cargo]
+2026-10-08T23:23:02+0900 idle streak reset after 14s: gpu=[] compilers=[3571564:cargo,3571605:rustc,3571606:rustc,3571607:rustc,3571608:rustc,3571609:rustc]
+2026-10-08T23:28:13+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-08T23:28:13+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-08T23:28:13+0900 sample 1: clean gpu=[]
+2026-10-08T23:28:15+0900 sample 2: clean gpu=[3751605:mlxcel-bench-de]
+2026-10-08T23:28:16+0900 sample 3: clean gpu=[3751605:mlxcel-bench-de]
+2026-10-08T23:28:17+0900 monitor: 3 samples
+2026-10-08T23:28:17+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base-null.log
new file mode 100644
index 000000000..6060db29e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base-null.log
@@ -0,0 +1,9 @@
+2026-10-09T02:05:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:05:19+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:05:34+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:05:34+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:05:34+0900 sample 1: clean gpu=[]
+2026-10-09T02:05:36+0900 sample 2: clean gpu=[4034722:mlxcel-bench-de]
+2026-10-09T02:05:37+0900 sample 3: clean gpu=[4034722:mlxcel-bench-de]
+2026-10-09T02:05:38+0900 monitor: 3 samples
+2026-10-09T02:05:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base.log
new file mode 100644
index 000000000..694241de9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_base.log
@@ -0,0 +1,9 @@
+2026-10-09T02:05:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:05:43+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:05:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:05:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:05:58+0900 sample 1: clean gpu=[]
+2026-10-09T02:06:00+0900 sample 2: clean gpu=[4047073:mlxcel-bench-de]
+2026-10-09T02:06:01+0900 sample 3: clean gpu=[4047073:mlxcel-bench-de]
+2026-10-09T02:06:02+0900 monitor: 3 samples
+2026-10-09T02:06:02+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_new.log
new file mode 100644
index 000000000..bd3302986
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r2_new.log
@@ -0,0 +1,8 @@
+2026-10-09T02:04:57+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:04:57+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:05:11+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:05:11+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:05:12+0900 sample 1: clean gpu=[]
+2026-10-09T02:05:13+0900 sample 2: clean gpu=[4023270:mlxcel-bench-de]
+2026-10-09T02:05:14+0900 monitor: 2 samples
+2026-10-09T02:05:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base-null.log
new file mode 100644
index 000000000..5b350e952
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base-null.log
@@ -0,0 +1,8 @@
+2026-10-09T02:24:51+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:24:51+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:25:05+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:25:05+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:25:06+0900 sample 1: clean gpu=[]
+2026-10-09T02:25:07+0900 sample 2: clean gpu=[479954:mlxcel-bench-de]
+2026-10-09T02:25:08+0900 monitor: 2 samples
+2026-10-09T02:25:08+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base.log
new file mode 100644
index 000000000..8a3bbd7a2
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_base.log
@@ -0,0 +1,9 @@
+2026-10-09T02:25:13+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:25:13+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:25:28+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:25:28+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:25:28+0900 sample 1: clean gpu=[]
+2026-10-09T02:25:30+0900 sample 2: clean gpu=[491350:mlxcel-bench-de]
+2026-10-09T02:25:31+0900 sample 3: clean gpu=[491350:mlxcel-bench-de]
+2026-10-09T02:25:32+0900 monitor: 3 samples
+2026-10-09T02:25:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_new.log
new file mode 100644
index 000000000..e15670b14
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp2048_r3_new.log
@@ -0,0 +1,8 @@
+2026-10-09T02:25:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:25:37+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:25:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:25:52+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:25:52+0900 sample 1: clean gpu=[]
+2026-10-09T02:25:54+0900 sample 2: clean gpu=[503716:mlxcel-bench-de]
+2026-10-09T02:25:55+0900 monitor: 2 samples
+2026-10-09T02:25:55+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..dfbcd6f75
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base-null.log
@@ -0,0 +1,9 @@
+2026-10-08T21:22:36+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-08T21:32:09+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 573s
+2026-10-08T21:32:09+0900 waiting for 60s of idle GPU and no compiler
+2026-10-08T21:33:38+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-08T21:33:38+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-08T21:33:39+0900 sample 1: clean gpu=[]
+2026-10-08T21:33:40+0900 sample 2: clean gpu=[3633468:mlxcel-bench-de]
+2026-10-08T21:33:41+0900 monitor: 2 samples
+2026-10-08T21:33:41+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..0f61f6c29
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_base.log
@@ -0,0 +1,13 @@
+2026-10-08T19:17:44+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-08T19:41:27+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 1423s
+2026-10-08T19:41:27+0900 waiting for 60s of idle GPU and no compiler
+2026-10-08T19:42:09+0900 idle streak reset after 28s: gpu=[] compilers=[3313432:cargo]
+2026-10-08T20:39:08+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-08T20:39:08+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-08T20:39:08+0900 sample 1: clean gpu=[]
+2026-10-08T20:39:10+0900 sample 2: clean gpu=[]
+2026-10-08T20:39:11+0900 sample 3: clean gpu=[1515758:mlxcel-bench-de]
+2026-10-08T20:39:13+0900 sample 4: clean gpu=[1515758:mlxcel-bench-de]
+2026-10-08T20:39:14+0900 sample 5: clean gpu=[1515758:mlxcel-bench-de]
+2026-10-08T20:39:15+0900 monitor: 5 samples
+2026-10-08T20:39:15+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..72e8e1dd3
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r1_new.log
@@ -0,0 +1,9 @@
+2026-10-08T20:39:25+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-08T21:20:56+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 2491s
+2026-10-08T21:20:56+0900 waiting for 60s of idle GPU and no compiler
+2026-10-08T21:22:23+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-08T21:22:23+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-08T21:22:23+0900 sample 1: clean gpu=[]
+2026-10-08T21:22:25+0900 sample 2: clean gpu=[3213872:mlxcel-bench-de]
+2026-10-08T21:22:26+0900 monitor: 2 samples
+2026-10-08T21:22:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..d99261dae
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base-null.log
@@ -0,0 +1,8 @@
+2026-10-09T02:04:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:04:12+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:04:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:04:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:04:26+0900 sample 1: clean gpu=[]
+2026-10-09T02:04:28+0900 sample 2: clean gpu=[4000296:mlxcel-bench-de]
+2026-10-09T02:04:29+0900 monitor: 2 samples
+2026-10-09T02:04:29+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..55ff939a4
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_base.log
@@ -0,0 +1,8 @@
+2026-10-09T02:04:34+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:04:34+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:04:49+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:04:49+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:04:49+0900 sample 1: clean gpu=[]
+2026-10-09T02:04:51+0900 sample 2: clean gpu=[4011756:mlxcel-bench-de]
+2026-10-09T02:04:52+0900 monitor: 2 samples
+2026-10-09T02:04:52+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..2dace6b91
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r2_new.log
@@ -0,0 +1,8 @@
+2026-10-09T02:03:49+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:03:49+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:04:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:04:04+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:04:04+0900 sample 1: clean gpu=[]
+2026-10-09T02:04:05+0900 sample 2: clean gpu=[3989149:mlxcel-bench-de]
+2026-10-09T02:04:06+0900 monitor: 2 samples
+2026-10-09T02:04:06+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..b54f6047c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base-null.log
@@ -0,0 +1,8 @@
+2026-10-09T02:23:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:23:43+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:23:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:23:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:23:58+0900 sample 1: clean gpu=[]
+2026-10-09T02:24:00+0900 sample 2: clean gpu=[446099:mlxcel-bench-de]
+2026-10-09T02:24:01+0900 monitor: 2 samples
+2026-10-09T02:24:01+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..87434b7eb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_base.log
@@ -0,0 +1,8 @@
+2026-10-09T02:24:06+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:24:06+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:24:20+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:24:20+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:24:20+0900 sample 1: clean gpu=[]
+2026-10-09T02:24:22+0900 sample 2: clean gpu=[457164:mlxcel-bench-de]
+2026-10-09T02:24:23+0900 monitor: 2 samples
+2026-10-09T02:24:23+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..721abb3c9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-0.6B-4bit_pp512_r3_new.log
@@ -0,0 +1,8 @@
+2026-10-09T02:24:28+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:24:28+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:24:43+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:24:43+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:24:43+0900 sample 1: clean gpu=[]
+2026-10-09T02:24:45+0900 sample 2: clean gpu=[468540:mlxcel-bench-de]
+2026-10-09T02:24:46+0900 monitor: 2 samples
+2026-10-09T02:24:46+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log
new file mode 100644
index 000000000..161f3078a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log
@@ -0,0 +1,23 @@
+2026-10-09T01:53:41+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:53:41+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:53:56+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:53:56+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:53:56+0900 sample 1: clean gpu=[]
+2026-10-09T01:53:58+0900 sample 2: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:53:59+0900 sample 3: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:01+0900 sample 4: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:02+0900 sample 5: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:04+0900 sample 6: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:05+0900 sample 7: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:07+0900 sample 8: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:08+0900 sample 9: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:09+0900 sample 10: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:11+0900 sample 11: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:12+0900 sample 12: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:14+0900 sample 13: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:15+0900 sample 14: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:17+0900 sample 15: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:18+0900 sample 16: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:20+0900 sample 17: clean gpu=[3646745:mlxcel-bench-de]
+2026-10-09T01:54:21+0900 monitor: 17 samples
+2026-10-09T01:54:21+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base.log
new file mode 100644
index 000000000..901482112
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_base.log
@@ -0,0 +1,23 @@
+2026-10-09T01:52:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:52:12+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:52:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:52:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:52:27+0900 sample 1: clean gpu=[]
+2026-10-09T01:52:28+0900 sample 2: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:30+0900 sample 3: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:31+0900 sample 4: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:33+0900 sample 5: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:34+0900 sample 6: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:36+0900 sample 7: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:37+0900 sample 8: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:39+0900 sample 9: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:40+0900 sample 10: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:42+0900 sample 11: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:43+0900 sample 12: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:45+0900 sample 13: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:46+0900 sample 14: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:48+0900 sample 15: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:49+0900 sample 16: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:51+0900 sample 17: clean gpu=[3596927:mlxcel-bench-de]
+2026-10-09T01:52:52+0900 monitor: 17 samples
+2026-10-09T01:52:52+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_new.log
new file mode 100644
index 000000000..c68c48163
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r1_new.log
@@ -0,0 +1,23 @@
+2026-10-09T01:52:57+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:52:57+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:53:11+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:53:11+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:53:12+0900 sample 1: clean gpu=[]
+2026-10-09T01:53:13+0900 sample 2: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:15+0900 sample 3: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:16+0900 sample 4: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:17+0900 sample 5: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:19+0900 sample 6: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:20+0900 sample 7: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:22+0900 sample 8: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:23+0900 sample 9: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:25+0900 sample 10: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:26+0900 sample 11: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:28+0900 sample 12: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:29+0900 sample 13: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:31+0900 sample 14: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:32+0900 sample 15: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:34+0900 sample 16: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:35+0900 sample 17: clean gpu=[3621883:mlxcel-bench-de]
+2026-10-09T01:53:36+0900 monitor: 17 samples
+2026-10-09T01:53:36+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log
new file mode 100644
index 000000000..a80cc0eca
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log
@@ -0,0 +1,22 @@
+2026-10-09T02:12:50+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:12:50+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:13:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:13:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:13:05+0900 sample 1: clean gpu=[]
+2026-10-09T02:13:06+0900 sample 2: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:08+0900 sample 3: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:09+0900 sample 4: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:11+0900 sample 5: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:12+0900 sample 6: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:14+0900 sample 7: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:15+0900 sample 8: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:17+0900 sample 9: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:18+0900 sample 10: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:20+0900 sample 11: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:21+0900 sample 12: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:23+0900 sample 13: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:24+0900 sample 14: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:26+0900 sample 15: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:27+0900 sample 16: clean gpu=[82768:mlxcel-bench-de]
+2026-10-09T02:13:28+0900 monitor: 16 samples
+2026-10-09T02:13:28+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base.log
new file mode 100644
index 000000000..e4455fc97
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_base.log
@@ -0,0 +1,23 @@
+2026-10-09T02:13:33+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:13:33+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:13:48+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:13:48+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:13:48+0900 sample 1: clean gpu=[]
+2026-10-09T02:13:50+0900 sample 2: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:13:51+0900 sample 3: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:13:53+0900 sample 4: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:13:54+0900 sample 5: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:13:56+0900 sample 6: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:13:57+0900 sample 7: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:13:59+0900 sample 8: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:00+0900 sample 9: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:02+0900 sample 10: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:03+0900 sample 11: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:05+0900 sample 12: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:06+0900 sample 13: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:08+0900 sample 14: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:09+0900 sample 15: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:11+0900 sample 16: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:12+0900 sample 17: clean gpu=[107711:mlxcel-bench-de]
+2026-10-09T02:14:13+0900 monitor: 17 samples
+2026-10-09T02:14:13+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_new.log
new file mode 100644
index 000000000..2897db43d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r2_new.log
@@ -0,0 +1,30 @@
+2026-10-09T02:11:54+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:11:54+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:12:09+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:12:09+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:12:09+0900 sample 1: clean gpu=[]
+2026-10-09T02:12:10+0900 sample 2: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:12+0900 sample 3: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:13+0900 sample 4: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:15+0900 sample 5: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:16+0900 sample 6: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:18+0900 sample 7: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:20+0900 sample 8: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:21+0900 sample 9: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:23+0900 sample 10: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:24+0900 sample 11: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:26+0900 sample 12: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:27+0900 sample 13: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:29+0900 sample 14: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:30+0900 sample 15: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:32+0900 sample 16: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:33+0900 sample 17: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:34+0900 sample 18: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:36+0900 sample 19: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:37+0900 sample 20: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:39+0900 sample 21: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:40+0900 sample 22: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:42+0900 sample 23: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:43+0900 sample 24: clean gpu=[50124:mlxcel-bench-de]
+2026-10-09T02:12:44+0900 monitor: 24 samples
+2026-10-09T02:12:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log
new file mode 100644
index 000000000..de93de17b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log
@@ -0,0 +1,23 @@
+2026-10-09T02:31:48+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:31:48+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:32:03+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:32:03+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:32:03+0900 sample 1: clean gpu=[]
+2026-10-09T02:32:05+0900 sample 2: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:06+0900 sample 3: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:08+0900 sample 4: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:09+0900 sample 5: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:11+0900 sample 6: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:12+0900 sample 7: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:14+0900 sample 8: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:15+0900 sample 9: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:17+0900 sample 10: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:18+0900 sample 11: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:19+0900 sample 12: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:21+0900 sample 13: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:22+0900 sample 14: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:24+0900 sample 15: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:25+0900 sample 16: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:27+0900 sample 17: clean gpu=[700676:mlxcel-bench-de]
+2026-10-09T02:32:28+0900 monitor: 17 samples
+2026-10-09T02:32:28+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base.log
new file mode 100644
index 000000000..fa79ced51
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_base.log
@@ -0,0 +1,23 @@
+2026-10-09T02:32:33+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:32:33+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:32:47+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:32:47+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:32:48+0900 sample 1: clean gpu=[]
+2026-10-09T02:32:49+0900 sample 2: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:32:51+0900 sample 3: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:32:52+0900 sample 4: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:32:54+0900 sample 5: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:32:55+0900 sample 6: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:32:57+0900 sample 7: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:32:58+0900 sample 8: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:00+0900 sample 9: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:01+0900 sample 10: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:03+0900 sample 11: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:04+0900 sample 12: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:06+0900 sample 13: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:07+0900 sample 14: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:09+0900 sample 15: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:10+0900 sample 16: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:12+0900 sample 17: clean gpu=[725677:mlxcel-bench-de]
+2026-10-09T02:33:13+0900 monitor: 17 samples
+2026-10-09T02:33:13+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_new.log
new file mode 100644
index 000000000..2913a0b35
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp2048_r3_new.log
@@ -0,0 +1,23 @@
+2026-10-09T02:33:18+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:33:18+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:33:32+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:33:32+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:33:33+0900 sample 1: clean gpu=[]
+2026-10-09T02:33:34+0900 sample 2: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:35+0900 sample 3: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:37+0900 sample 4: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:38+0900 sample 5: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:40+0900 sample 6: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:41+0900 sample 7: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:43+0900 sample 8: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:44+0900 sample 9: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:46+0900 sample 10: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:47+0900 sample 11: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:49+0900 sample 12: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:50+0900 sample 13: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:52+0900 sample 14: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:53+0900 sample 15: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:55+0900 sample 16: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:56+0900 sample 17: clean gpu=[750642:mlxcel-bench-de]
+2026-10-09T02:33:57+0900 monitor: 17 samples
+2026-10-09T02:33:57+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..c1c13baf8
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log
@@ -0,0 +1,16 @@
+2026-10-09T01:51:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:51:37+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:51:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:51:52+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:51:52+0900 sample 1: clean gpu=[]
+2026-10-09T01:51:54+0900 sample 2: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:51:55+0900 sample 3: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:51:57+0900 sample 4: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:51:59+0900 sample 5: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:52:00+0900 sample 6: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:52:01+0900 sample 7: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:52:03+0900 sample 8: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:52:04+0900 sample 9: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:52:06+0900 sample 10: clean gpu=[3577931:mlxcel-bench-de]
+2026-10-09T01:52:07+0900 monitor: 10 samples
+2026-10-09T01:52:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..b393faedb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_base.log
@@ -0,0 +1,27 @@
+2026-10-09T01:50:11+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:50:11+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:50:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:50:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:50:26+0900 sample 1: clean gpu=[]
+2026-10-09T01:50:28+0900 sample 2: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:29+0900 sample 3: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:31+0900 sample 4: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:32+0900 sample 5: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:34+0900 sample 6: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:35+0900 sample 7: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:37+0900 sample 8: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:38+0900 sample 9: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:40+0900 sample 10: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:41+0900 sample 11: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:43+0900 sample 12: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:45+0900 sample 13: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:46+0900 sample 14: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:48+0900 sample 15: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:49+0900 sample 16: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:51+0900 sample 17: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:52+0900 sample 18: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:54+0900 sample 19: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:55+0900 sample 20: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:57+0900 sample 21: clean gpu=[3529109:mlxcel-bench-de]
+2026-10-09T01:50:58+0900 monitor: 21 samples
+2026-10-09T01:50:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..f7d499617
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r1_new.log
@@ -0,0 +1,16 @@
+2026-10-09T01:51:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:51:03+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:51:17+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:51:17+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:51:18+0900 sample 1: clean gpu=[]
+2026-10-09T01:51:19+0900 sample 2: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:21+0900 sample 3: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:22+0900 sample 4: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:24+0900 sample 5: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:25+0900 sample 6: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:27+0900 sample 7: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:28+0900 sample 8: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:30+0900 sample 9: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:31+0900 sample 10: clean gpu=[3558843:mlxcel-bench-de]
+2026-10-09T01:51:32+0900 monitor: 10 samples
+2026-10-09T01:51:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..59dfefe37
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log
@@ -0,0 +1,16 @@
+2026-10-09T02:10:44+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:10:44+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:10:59+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:10:59+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:11:00+0900 sample 1: clean gpu=[]
+2026-10-09T02:11:01+0900 sample 2: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:03+0900 sample 3: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:04+0900 sample 4: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:06+0900 sample 5: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:07+0900 sample 6: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:09+0900 sample 7: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:10+0900 sample 8: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:12+0900 sample 9: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:13+0900 sample 10: clean gpu=[11664:mlxcel-bench-de]
+2026-10-09T02:11:14+0900 monitor: 10 samples
+2026-10-09T02:11:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..895d4c74e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_base.log
@@ -0,0 +1,16 @@
+2026-10-09T02:11:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:11:19+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:11:34+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:11:34+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:11:34+0900 sample 1: clean gpu=[]
+2026-10-09T02:11:36+0900 sample 2: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:37+0900 sample 3: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:39+0900 sample 4: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:40+0900 sample 5: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:42+0900 sample 6: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:43+0900 sample 7: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:45+0900 sample 8: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:46+0900 sample 9: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:48+0900 sample 10: clean gpu=[30912:mlxcel-bench-de]
+2026-10-09T02:11:49+0900 monitor: 10 samples
+2026-10-09T02:11:49+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..f9b69743e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r2_new.log
@@ -0,0 +1,16 @@
+2026-10-09T02:10:10+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:10:10+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:10:25+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:10:25+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:10:25+0900 sample 1: clean gpu=[]
+2026-10-09T02:10:27+0900 sample 2: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:28+0900 sample 3: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:29+0900 sample 4: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:31+0900 sample 5: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:32+0900 sample 6: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:34+0900 sample 7: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:35+0900 sample 8: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:37+0900 sample 9: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:38+0900 sample 10: clean gpu=[4186267:mlxcel-bench-de]
+2026-10-09T02:10:39+0900 monitor: 10 samples
+2026-10-09T02:10:39+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..dacbb1135
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log
@@ -0,0 +1,17 @@
+2026-10-09T02:30:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:30:03+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:30:18+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:30:18+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:30:19+0900 sample 1: clean gpu=[]
+2026-10-09T02:30:20+0900 sample 2: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:21+0900 sample 3: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:23+0900 sample 4: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:24+0900 sample 5: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:26+0900 sample 6: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:27+0900 sample 7: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:29+0900 sample 8: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:30+0900 sample 9: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:32+0900 sample 10: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:33+0900 sample 11: clean gpu=[643275:mlxcel-bench-de]
+2026-10-09T02:30:34+0900 monitor: 11 samples
+2026-10-09T02:30:34+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..21b084363
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_base.log
@@ -0,0 +1,16 @@
+2026-10-09T02:30:39+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:30:39+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:30:54+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:30:54+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:30:54+0900 sample 1: clean gpu=[]
+2026-10-09T02:30:56+0900 sample 2: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:30:57+0900 sample 3: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:30:59+0900 sample 4: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:00+0900 sample 5: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:02+0900 sample 6: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:03+0900 sample 7: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:05+0900 sample 8: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:06+0900 sample 9: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:08+0900 sample 10: clean gpu=[663423:mlxcel-bench-de]
+2026-10-09T02:31:09+0900 monitor: 10 samples
+2026-10-09T02:31:09+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..b51392ba2
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/Qwen3-30B-A3B-4bit_pp512_r3_new.log
@@ -0,0 +1,16 @@
+2026-10-09T02:31:14+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:31:14+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:31:28+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:31:28+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:31:29+0900 sample 1: clean gpu=[]
+2026-10-09T02:31:30+0900 sample 2: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:32+0900 sample 3: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:33+0900 sample 4: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:35+0900 sample 5: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:36+0900 sample 6: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:38+0900 sample 7: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:39+0900 sample 8: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:41+0900 sample 9: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:42+0900 sample 10: clean gpu=[682097:mlxcel-bench-de]
+2026-10-09T02:31:43+0900 monitor: 10 samples
+2026-10-09T02:31:43+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base-null.log
new file mode 100644
index 000000000..e4bde4d97
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base-null.log
@@ -0,0 +1,11 @@
+2026-10-09T01:49:44+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:49:44+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:49:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:49:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:49:59+0900 sample 1: clean gpu=[]
+2026-10-09T01:50:00+0900 sample 2: clean gpu=[3514781:mlxcel-bench-de]
+2026-10-09T01:50:02+0900 sample 3: clean gpu=[3514781:mlxcel-bench-de]
+2026-10-09T01:50:03+0900 sample 4: clean gpu=[3514781:mlxcel-bench-de]
+2026-10-09T01:50:05+0900 sample 5: clean gpu=[3514781:mlxcel-bench-de]
+2026-10-09T01:50:06+0900 monitor: 5 samples
+2026-10-09T01:50:06+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base.log
new file mode 100644
index 000000000..6a91bef7a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_base.log
@@ -0,0 +1,11 @@
+2026-10-09T01:48:50+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:48:50+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:49:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:49:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:49:05+0900 sample 1: clean gpu=[]
+2026-10-09T01:49:06+0900 sample 2: clean gpu=[3486042:mlxcel-bench-de]
+2026-10-09T01:49:08+0900 sample 3: clean gpu=[3486042:mlxcel-bench-de]
+2026-10-09T01:49:09+0900 sample 4: clean gpu=[3486042:mlxcel-bench-de]
+2026-10-09T01:49:11+0900 sample 5: clean gpu=[3486042:mlxcel-bench-de]
+2026-10-09T01:49:12+0900 monitor: 5 samples
+2026-10-09T01:49:12+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_new.log
new file mode 100644
index 000000000..adfb25015
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r1_new.log
@@ -0,0 +1,11 @@
+2026-10-09T01:49:17+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:49:17+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:49:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:49:31+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T01:49:32+0900 sample 1: clean gpu=[]
+2026-10-09T01:49:33+0900 sample 2: clean gpu=[3500474:mlxcel-bench-de]
+2026-10-09T01:49:35+0900 sample 3: clean gpu=[3500474:mlxcel-bench-de]
+2026-10-09T01:49:36+0900 sample 4: clean gpu=[3500474:mlxcel-bench-de]
+2026-10-09T01:49:38+0900 sample 5: clean gpu=[3500474:mlxcel-bench-de]
+2026-10-09T01:49:39+0900 monitor: 5 samples
+2026-10-09T01:49:39+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base-null.log
new file mode 100644
index 000000000..37c2f6162
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base-null.log
@@ -0,0 +1,11 @@
+2026-10-09T02:09:16+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:09:16+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:09:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:09:31+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:09:31+0900 sample 1: clean gpu=[]
+2026-10-09T02:09:33+0900 sample 2: clean gpu=[4158223:mlxcel-bench-de]
+2026-10-09T02:09:34+0900 sample 3: clean gpu=[4158223:mlxcel-bench-de]
+2026-10-09T02:09:36+0900 sample 4: clean gpu=[4158223:mlxcel-bench-de]
+2026-10-09T02:09:37+0900 sample 5: clean gpu=[4158223:mlxcel-bench-de]
+2026-10-09T02:09:38+0900 monitor: 5 samples
+2026-10-09T02:09:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base.log
new file mode 100644
index 000000000..520ff5375
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_base.log
@@ -0,0 +1,11 @@
+2026-10-09T02:09:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:09:43+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:09:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:09:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:09:58+0900 sample 1: clean gpu=[]
+2026-10-09T02:10:00+0900 sample 2: clean gpu=[4172265:mlxcel-bench-de]
+2026-10-09T02:10:01+0900 sample 3: clean gpu=[4172265:mlxcel-bench-de]
+2026-10-09T02:10:02+0900 sample 4: clean gpu=[4172265:mlxcel-bench-de]
+2026-10-09T02:10:04+0900 sample 5: clean gpu=[4172265:mlxcel-bench-de]
+2026-10-09T02:10:05+0900 monitor: 5 samples
+2026-10-09T02:10:05+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_new.log
new file mode 100644
index 000000000..8d0535c4b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r2_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:08:49+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:08:49+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:09:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:09:04+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:09:04+0900 sample 1: clean gpu=[]
+2026-10-09T02:09:05+0900 sample 2: clean gpu=[4144067:mlxcel-bench-de]
+2026-10-09T02:09:07+0900 sample 3: clean gpu=[4144067:mlxcel-bench-de]
+2026-10-09T02:09:08+0900 sample 4: clean gpu=[4144067:mlxcel-bench-de]
+2026-10-09T02:09:10+0900 sample 5: clean gpu=[4144067:mlxcel-bench-de]
+2026-10-09T02:09:11+0900 monitor: 5 samples
+2026-10-09T02:09:11+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base-null.log
new file mode 100644
index 000000000..1aa849f12
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base-null.log
@@ -0,0 +1,11 @@
+2026-10-09T02:28:42+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:28:42+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:28:57+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:28:57+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:28:57+0900 sample 1: clean gpu=[]
+2026-10-09T02:28:59+0900 sample 2: clean gpu=[600016:mlxcel-bench-de]
+2026-10-09T02:29:00+0900 sample 3: clean gpu=[600016:mlxcel-bench-de]
+2026-10-09T02:29:02+0900 sample 4: clean gpu=[600016:mlxcel-bench-de]
+2026-10-09T02:29:03+0900 sample 5: clean gpu=[600016:mlxcel-bench-de]
+2026-10-09T02:29:04+0900 monitor: 5 samples
+2026-10-09T02:29:04+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base.log
new file mode 100644
index 000000000..6a6fbb3a2
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_base.log
@@ -0,0 +1,11 @@
+2026-10-09T02:29:09+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:29:09+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:29:24+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:29:24+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:29:24+0900 sample 1: clean gpu=[]
+2026-10-09T02:29:26+0900 sample 2: clean gpu=[614459:mlxcel-bench-de]
+2026-10-09T02:29:27+0900 sample 3: clean gpu=[614459:mlxcel-bench-de]
+2026-10-09T02:29:29+0900 sample 4: clean gpu=[614459:mlxcel-bench-de]
+2026-10-09T02:29:30+0900 sample 5: clean gpu=[614459:mlxcel-bench-de]
+2026-10-09T02:29:31+0900 monitor: 5 samples
+2026-10-09T02:29:31+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_new.log
new file mode 100644
index 000000000..6e64b9684
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp2048_r3_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:29:36+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:29:36+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:29:51+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:29:51+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+2026-10-09T02:29:51+0900 sample 1: clean gpu=[]
+2026-10-09T02:29:53+0900 sample 2: clean gpu=[628844:mlxcel-bench-de]
+2026-10-09T02:29:54+0900 sample 3: clean gpu=[628844:mlxcel-bench-de]
+2026-10-09T02:29:56+0900 sample 4: clean gpu=[628844:mlxcel-bench-de]
+2026-10-09T02:29:57+0900 sample 5: clean gpu=[628844:mlxcel-bench-de]
+2026-10-09T02:29:58+0900 monitor: 5 samples
+2026-10-09T02:29:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..0c4b5c694
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base-null.log
@@ -0,0 +1,10 @@
+2026-10-09T01:48:24+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:48:24+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:48:39+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:48:39+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:48:39+0900 sample 1: clean gpu=[]
+2026-10-09T01:48:41+0900 sample 2: clean gpu=[3468221:mlxcel-bench-de]
+2026-10-09T01:48:42+0900 sample 3: clean gpu=[3468221:mlxcel-bench-de]
+2026-10-09T01:48:44+0900 sample 4: clean gpu=[3468221:mlxcel-bench-de]
+2026-10-09T01:48:45+0900 monitor: 4 samples
+2026-10-09T01:48:45+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..4c60b7edb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_base.log
@@ -0,0 +1,11 @@
+2026-10-09T01:47:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:47:31+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:47:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:47:46+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:47:46+0900 sample 1: clean gpu=[]
+2026-10-09T01:47:48+0900 sample 2: clean gpu=[3408879:mlxcel-bench-de]
+2026-10-09T01:47:49+0900 sample 3: clean gpu=[3408879:mlxcel-bench-de]
+2026-10-09T01:47:51+0900 sample 4: clean gpu=[3408879:mlxcel-bench-de]
+2026-10-09T01:47:52+0900 sample 5: clean gpu=[3408879:mlxcel-bench-de]
+2026-10-09T01:47:53+0900 monitor: 5 samples
+2026-10-09T01:47:53+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..ee5b08fc8
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r1_new.log
@@ -0,0 +1,10 @@
+2026-10-09T01:47:58+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:47:58+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:48:13+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:48:13+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:48:13+0900 sample 1: clean gpu=[]
+2026-10-09T01:48:15+0900 sample 2: clean gpu=[3437165:mlxcel-bench-de]
+2026-10-09T01:48:16+0900 sample 3: clean gpu=[3437165:mlxcel-bench-de]
+2026-10-09T01:48:18+0900 sample 4: clean gpu=[3437165:mlxcel-bench-de]
+2026-10-09T01:48:19+0900 monitor: 4 samples
+2026-10-09T01:48:19+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..0a064da03
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base-null.log
@@ -0,0 +1,10 @@
+2026-10-09T02:07:58+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:07:58+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:08:12+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:08:12+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:08:13+0900 sample 1: clean gpu=[]
+2026-10-09T02:08:14+0900 sample 2: clean gpu=[4117635:mlxcel-bench-de]
+2026-10-09T02:08:16+0900 sample 3: clean gpu=[4117635:mlxcel-bench-de]
+2026-10-09T02:08:17+0900 sample 4: clean gpu=[4117635:mlxcel-bench-de]
+2026-10-09T02:08:18+0900 monitor: 4 samples
+2026-10-09T02:08:18+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..da0dfe82c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_base.log
@@ -0,0 +1,10 @@
+2026-10-09T02:08:23+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:08:23+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:08:38+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:08:38+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:08:38+0900 sample 1: clean gpu=[]
+2026-10-09T02:08:40+0900 sample 2: clean gpu=[4130867:mlxcel-bench-de]
+2026-10-09T02:08:41+0900 sample 3: clean gpu=[4130867:mlxcel-bench-de]
+2026-10-09T02:08:43+0900 sample 4: clean gpu=[4130867:mlxcel-bench-de]
+2026-10-09T02:08:44+0900 monitor: 4 samples
+2026-10-09T02:08:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..a030fa689
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r2_new.log
@@ -0,0 +1,10 @@
+2026-10-09T02:07:32+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:07:32+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:07:47+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:07:47+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:07:47+0900 sample 1: clean gpu=[]
+2026-10-09T02:07:49+0900 sample 2: clean gpu=[4104326:mlxcel-bench-de]
+2026-10-09T02:07:50+0900 sample 3: clean gpu=[4104326:mlxcel-bench-de]
+2026-10-09T02:07:52+0900 sample 4: clean gpu=[4104326:mlxcel-bench-de]
+2026-10-09T02:07:53+0900 monitor: 4 samples
+2026-10-09T02:07:53+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..8de1c1f88
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base-null.log
@@ -0,0 +1,10 @@
+2026-10-09T02:27:25+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:27:25+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:27:40+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:27:40+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:27:40+0900 sample 1: clean gpu=[]
+2026-10-09T02:27:42+0900 sample 2: clean gpu=[559971:mlxcel-bench-de]
+2026-10-09T02:27:43+0900 sample 3: clean gpu=[559971:mlxcel-bench-de]
+2026-10-09T02:27:45+0900 sample 4: clean gpu=[559971:mlxcel-bench-de]
+2026-10-09T02:27:46+0900 monitor: 4 samples
+2026-10-09T02:27:46+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..f227e46b1
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_base.log
@@ -0,0 +1,10 @@
+2026-10-09T02:27:51+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:27:51+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:28:05+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:28:05+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:28:06+0900 sample 1: clean gpu=[]
+2026-10-09T02:28:07+0900 sample 2: clean gpu=[573242:mlxcel-bench-de]
+2026-10-09T02:28:09+0900 sample 3: clean gpu=[573242:mlxcel-bench-de]
+2026-10-09T02:28:10+0900 sample 4: clean gpu=[573242:mlxcel-bench-de]
+2026-10-09T02:28:11+0900 monitor: 4 samples
+2026-10-09T02:28:11+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..998d1089b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gemma-3-4b-it-4bit_pp512_r3_new.log
@@ -0,0 +1,10 @@
+2026-10-09T02:28:16+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:28:16+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:28:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:28:31+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:28:31+0900 sample 1: clean gpu=[]
+2026-10-09T02:28:33+0900 sample 2: clean gpu=[586534:mlxcel-bench-de]
+2026-10-09T02:28:34+0900 sample 3: clean gpu=[586534:mlxcel-bench-de]
+2026-10-09T02:28:36+0900 sample 4: clean gpu=[586534:mlxcel-bench-de]
+2026-10-09T02:28:37+0900 monitor: 4 samples
+2026-10-09T02:28:37+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log
new file mode 100644
index 000000000..0648dfb72
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log
@@ -0,0 +1,20 @@
+2026-10-09T01:59:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:59:12+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:59:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:59:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:59:27+0900 sample 1: clean gpu=[]
+2026-10-09T01:59:28+0900 sample 2: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:30+0900 sample 3: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:31+0900 sample 4: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:33+0900 sample 5: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:35+0900 sample 6: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:36+0900 sample 7: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:38+0900 sample 8: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:39+0900 sample 9: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:41+0900 sample 10: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:42+0900 sample 11: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:44+0900 sample 12: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:45+0900 sample 13: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:47+0900 sample 14: clean gpu=[3829107:mlxcel-bench-de]
+2026-10-09T01:59:48+0900 monitor: 14 samples
+2026-10-09T01:59:48+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log
new file mode 100644
index 000000000..f8fc978d1
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log
@@ -0,0 +1,30 @@
+2026-10-09T01:57:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:57:31+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:57:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:57:46+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:57:46+0900 sample 1: clean gpu=[]
+2026-10-09T01:57:48+0900 sample 2: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:49+0900 sample 3: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:51+0900 sample 4: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:52+0900 sample 5: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:54+0900 sample 6: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:56+0900 sample 7: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:57+0900 sample 8: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:57:59+0900 sample 9: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:00+0900 sample 10: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:02+0900 sample 11: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:03+0900 sample 12: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:05+0900 sample 13: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:06+0900 sample 14: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:08+0900 sample 15: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:09+0900 sample 16: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:11+0900 sample 17: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:12+0900 sample 18: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:14+0900 sample 19: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:15+0900 sample 20: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:17+0900 sample 21: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:18+0900 sample 22: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:20+0900 sample 23: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:21+0900 sample 24: clean gpu=[3771509:mlxcel-bench-de]
+2026-10-09T01:58:22+0900 monitor: 24 samples
+2026-10-09T01:58:22+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log
new file mode 100644
index 000000000..d2f0b65aa
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log
@@ -0,0 +1,22 @@
+2026-10-09T01:58:27+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:58:27+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:58:42+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:58:42+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:58:43+0900 sample 1: clean gpu=[]
+2026-10-09T01:58:44+0900 sample 2: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:46+0900 sample 3: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:47+0900 sample 4: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:49+0900 sample 5: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:50+0900 sample 6: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:52+0900 sample 7: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:53+0900 sample 8: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:55+0900 sample 9: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:56+0900 sample 10: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:58+0900 sample 11: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:58:59+0900 sample 12: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:59:01+0900 sample 13: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:59:03+0900 sample 14: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:59:04+0900 sample 15: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:59:06+0900 sample 16: clean gpu=[3804148:mlxcel-bench-de]
+2026-10-09T01:59:07+0900 monitor: 16 samples
+2026-10-09T01:59:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log
new file mode 100644
index 000000000..2d5b23539
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log
@@ -0,0 +1,22 @@
+2026-10-09T02:18:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:18:19+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:18:33+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:18:33+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:18:34+0900 sample 1: clean gpu=[]
+2026-10-09T02:18:35+0900 sample 2: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:37+0900 sample 3: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:38+0900 sample 4: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:40+0900 sample 5: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:41+0900 sample 6: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:43+0900 sample 7: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:44+0900 sample 8: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:46+0900 sample 9: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:47+0900 sample 10: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:49+0900 sample 11: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:50+0900 sample 12: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:52+0900 sample 13: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:54+0900 sample 14: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:55+0900 sample 15: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:57+0900 sample 16: clean gpu=[262691:mlxcel-bench-de]
+2026-10-09T02:18:58+0900 monitor: 16 samples
+2026-10-09T02:18:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log
new file mode 100644
index 000000000..f2e7bb2d9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log
@@ -0,0 +1,21 @@
+2026-10-09T02:19:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:19:03+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:19:17+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:19:17+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:19:18+0900 sample 1: clean gpu=[]
+2026-10-09T02:19:19+0900 sample 2: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:21+0900 sample 3: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:22+0900 sample 4: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:24+0900 sample 5: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:25+0900 sample 6: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:27+0900 sample 7: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:28+0900 sample 8: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:30+0900 sample 9: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:31+0900 sample 10: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:33+0900 sample 11: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:34+0900 sample 12: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:36+0900 sample 13: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:37+0900 sample 14: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:39+0900 sample 15: clean gpu=[286819:mlxcel-bench-de]
+2026-10-09T02:19:40+0900 monitor: 15 samples
+2026-10-09T02:19:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log
new file mode 100644
index 000000000..31a731b87
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log
@@ -0,0 +1,31 @@
+2026-10-09T02:17:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:17:21+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:17:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:17:36+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:17:36+0900 sample 1: clean gpu=[]
+2026-10-09T02:17:38+0900 sample 2: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:39+0900 sample 3: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:41+0900 sample 4: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:42+0900 sample 5: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:44+0900 sample 6: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:45+0900 sample 7: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:47+0900 sample 8: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:48+0900 sample 9: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:50+0900 sample 10: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:51+0900 sample 11: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:53+0900 sample 12: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:54+0900 sample 13: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:56+0900 sample 14: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:57+0900 sample 15: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:17:59+0900 sample 16: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:00+0900 sample 17: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:02+0900 sample 18: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:04+0900 sample 19: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:05+0900 sample 20: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:07+0900 sample 21: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:08+0900 sample 22: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:10+0900 sample 23: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:11+0900 sample 24: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:13+0900 sample 25: clean gpu=[230262:mlxcel-bench-de]
+2026-10-09T02:18:14+0900 monitor: 25 samples
+2026-10-09T02:18:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log
new file mode 100644
index 000000000..1270b6386
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log
@@ -0,0 +1,31 @@
+2026-10-09T02:37:06+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:37:06+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:37:21+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:37:21+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:37:21+0900 sample 1: clean gpu=[]
+2026-10-09T02:37:23+0900 sample 2: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:24+0900 sample 3: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:26+0900 sample 4: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:27+0900 sample 5: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:29+0900 sample 6: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:31+0900 sample 7: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:32+0900 sample 8: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:34+0900 sample 9: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:35+0900 sample 10: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:37+0900 sample 11: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:38+0900 sample 12: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:40+0900 sample 13: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:41+0900 sample 14: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:43+0900 sample 15: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:44+0900 sample 16: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:46+0900 sample 17: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:47+0900 sample 18: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:49+0900 sample 19: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:50+0900 sample 20: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:51+0900 sample 21: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:53+0900 sample 22: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:54+0900 sample 23: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:56+0900 sample 24: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:57+0900 sample 25: clean gpu=[872333:mlxcel-bench-de]
+2026-10-09T02:37:58+0900 monitor: 25 samples
+2026-10-09T02:37:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log
new file mode 100644
index 000000000..28d486c30
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log
@@ -0,0 +1,23 @@
+2026-10-09T02:38:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:38:03+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:38:18+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:38:18+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:38:18+0900 sample 1: clean gpu=[]
+2026-10-09T02:38:20+0900 sample 2: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:21+0900 sample 3: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:23+0900 sample 4: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:24+0900 sample 5: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:26+0900 sample 6: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:27+0900 sample 7: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:29+0900 sample 8: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:30+0900 sample 9: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:32+0900 sample 10: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:33+0900 sample 11: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:35+0900 sample 12: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:36+0900 sample 13: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:38+0900 sample 14: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:39+0900 sample 15: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:41+0900 sample 16: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:42+0900 sample 17: clean gpu=[904504:mlxcel-bench-de]
+2026-10-09T02:38:43+0900 monitor: 17 samples
+2026-10-09T02:38:43+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log
new file mode 100644
index 000000000..a7718e62e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log
@@ -0,0 +1,21 @@
+2026-10-09T02:38:48+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:38:48+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:39:03+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:39:03+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:39:03+0900 sample 1: clean gpu=[]
+2026-10-09T02:39:05+0900 sample 2: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:06+0900 sample 3: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:08+0900 sample 4: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:09+0900 sample 5: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:11+0900 sample 6: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:13+0900 sample 7: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:14+0900 sample 8: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:16+0900 sample 9: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:17+0900 sample 10: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:19+0900 sample 11: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:20+0900 sample 12: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:22+0900 sample 13: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:23+0900 sample 14: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:24+0900 sample 15: clean gpu=[930024:mlxcel-bench-de]
+2026-10-09T02:39:25+0900 monitor: 15 samples
+2026-10-09T02:39:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..fe68df73c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log
@@ -0,0 +1,10 @@
+2026-10-09T01:55:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:55:19+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:55:34+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:55:34+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:55:34+0900 sample 1: clean gpu=[]
+2026-10-09T01:55:36+0900 sample 2: clean gpu=[3699816:mlxcel-bench-de]
+2026-10-09T01:55:38+0900 sample 3: clean gpu=[3699816:mlxcel-bench-de]
+2026-10-09T01:55:39+0900 sample 4: clean gpu=[3699816:mlxcel-bench-de]
+2026-10-09T01:55:40+0900 monitor: 4 samples
+2026-10-09T01:55:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..13fb69f38
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_base.log
@@ -0,0 +1,12 @@
+2026-10-09T01:54:26+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:54:26+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:54:40+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:54:40+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:54:41+0900 sample 1: clean gpu=[]
+2026-10-09T01:54:42+0900 sample 2: clean gpu=[3671563:mlxcel-bench-de]
+2026-10-09T01:54:44+0900 sample 3: clean gpu=[3671563:mlxcel-bench-de]
+2026-10-09T01:54:45+0900 sample 4: clean gpu=[3671563:mlxcel-bench-de]
+2026-10-09T01:54:46+0900 sample 5: clean gpu=[3671563:mlxcel-bench-de]
+2026-10-09T01:54:48+0900 sample 6: clean gpu=[3671563:mlxcel-bench-de]
+2026-10-09T01:54:49+0900 monitor: 6 samples
+2026-10-09T01:54:49+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..14f6606ca
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r1_new.log
@@ -0,0 +1,10 @@
+2026-10-09T01:54:54+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T01:54:54+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T01:55:08+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T01:55:08+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T01:55:09+0900 sample 1: clean gpu=[]
+2026-10-09T01:55:10+0900 sample 2: clean gpu=[3686559:mlxcel-bench-de]
+2026-10-09T01:55:12+0900 sample 3: clean gpu=[3686559:mlxcel-bench-de]
+2026-10-09T01:55:13+0900 sample 4: clean gpu=[3686559:mlxcel-bench-de]
+2026-10-09T01:55:14+0900 monitor: 4 samples
+2026-10-09T01:55:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..b256f3565
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log
@@ -0,0 +1,10 @@
+2026-10-09T02:14:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:14:45+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:15:00+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:15:00+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:15:00+0900 sample 1: clean gpu=[]
+2026-10-09T02:15:02+0900 sample 2: clean gpu=[147991:mlxcel-bench-de]
+2026-10-09T02:15:03+0900 sample 3: clean gpu=[147991:mlxcel-bench-de]
+2026-10-09T02:15:05+0900 sample 4: clean gpu=[147991:mlxcel-bench-de]
+2026-10-09T02:15:06+0900 monitor: 4 samples
+2026-10-09T02:15:06+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..bda10692f
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_base.log
@@ -0,0 +1,10 @@
+2026-10-09T02:15:11+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:15:11+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:15:25+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:15:25+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:15:26+0900 sample 1: clean gpu=[]
+2026-10-09T02:15:27+0900 sample 2: clean gpu=[161122:mlxcel-bench-de]
+2026-10-09T02:15:29+0900 sample 3: clean gpu=[161122:mlxcel-bench-de]
+2026-10-09T02:15:30+0900 sample 4: clean gpu=[161122:mlxcel-bench-de]
+2026-10-09T02:15:31+0900 monitor: 4 samples
+2026-10-09T02:15:31+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..4b5b60c21
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r2_new.log
@@ -0,0 +1,11 @@
+2026-10-09T02:14:18+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:14:18+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:14:33+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:14:33+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:14:33+0900 sample 1: clean gpu=[]
+2026-10-09T02:14:35+0900 sample 2: clean gpu=[133603:mlxcel-bench-de]
+2026-10-09T02:14:36+0900 sample 3: clean gpu=[133603:mlxcel-bench-de]
+2026-10-09T02:14:38+0900 sample 4: clean gpu=[133603:mlxcel-bench-de]
+2026-10-09T02:14:39+0900 sample 5: clean gpu=[133603:mlxcel-bench-de]
+2026-10-09T02:14:40+0900 monitor: 5 samples
+2026-10-09T02:14:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..583bb9b7b
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log
@@ -0,0 +1,12 @@
+2026-10-09T02:34:02+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:34:02+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:34:17+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:34:17+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:34:17+0900 sample 1: clean gpu=[]
+2026-10-09T02:34:19+0900 sample 2: clean gpu=[775576:mlxcel-bench-de]
+2026-10-09T02:34:20+0900 sample 3: clean gpu=[775576:mlxcel-bench-de]
+2026-10-09T02:34:22+0900 sample 4: clean gpu=[775576:mlxcel-bench-de]
+2026-10-09T02:34:23+0900 sample 5: clean gpu=[775576:mlxcel-bench-de]
+2026-10-09T02:34:24+0900 sample 6: clean gpu=[775576:mlxcel-bench-de]
+2026-10-09T02:34:25+0900 monitor: 6 samples
+2026-10-09T02:34:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..0785bd5a8
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_base.log
@@ -0,0 +1,10 @@
+2026-10-09T02:34:30+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:34:30+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:34:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:34:45+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:34:45+0900 sample 1: clean gpu=[]
+2026-10-09T02:34:47+0900 sample 2: clean gpu=[790307:mlxcel-bench-de]
+2026-10-09T02:34:48+0900 sample 3: clean gpu=[790307:mlxcel-bench-de]
+2026-10-09T02:34:50+0900 sample 4: clean gpu=[790307:mlxcel-bench-de]
+2026-10-09T02:34:51+0900 monitor: 4 samples
+2026-10-09T02:34:51+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..17728d15a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/guard/logs/rounds/guard/granite-4.0-h-tiny-4bit_pp512_r3_new.log
@@ -0,0 +1,10 @@
+2026-10-09T02:34:56+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+2026-10-09T02:34:56+0900 waiting for 10s of idle GPU and no compiler
+2026-10-09T02:35:10+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+2026-10-09T02:35:10+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+2026-10-09T02:35:11+0900 sample 1: clean gpu=[]
+2026-10-09T02:35:12+0900 sample 2: clean gpu=[803170:mlxcel-bench-de]
+2026-10-09T02:35:14+0900 sample 3: clean gpu=[803170:mlxcel-bench-de]
+2026-10-09T02:35:15+0900 sample 4: clean gpu=[803170:mlxcel-bench-de]
+2026-10-09T02:35:16+0900 monitor: 4 samples
+2026-10-09T02:35:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw-logs.tar.gz b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw-logs.tar.gz
deleted file mode 100644
index fc22a43bb7d386ca7d8f61dd269055eaed4568a7..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 14119
zcmYj%by$;c)IT-~0s;b3qf3wyL24=zl1hporGRuPDSJ>vx?5sWinM^zOhQJZv=S4L
z28n@eW83{3e&6?Zy}PdcyXQXVKIc<=VA0f6|2`)+a;%fH6h42!W*g+yWlcMc5+vV>
zS)}%F^cT7(4fJo!<ox_#9OXMQ)B5ov`d6VWlB_++#!hV`-~G;FS<op)AuZw8vd_~&
z^HbJ?NLq2N$P`NuEr|XK$p*Z<eqoI#(NOg!8`FziD%CyKk?Oc)n^Rws`?j%mm!QMu
z&qJDGm!AQQzDKf&l+UyUzBhpDYmc+ao$MiMyYPp)$?<)H=-&%bh?sUs&(HO<X5p=E
zx2im5AZ^jUKk%>jxr8&kc2B<LRkTb)BLHhM9=*OuSlCy*|Bw*E!^auJ`J2Ra0eJ!W
zZ7U(1Awxn4`9)8m=_IRln_bSU%L&(xy`qj%SqpuV7$ShX8e7p8AY+|7<XD&Nx%A|G
z-0-i$_iDj8?{hvO&c2Im{WWHHN-C;t{+zC)BQI{11YYtq8Y{LvmbTV+oGT&akIq<L
zZE61-rHXs{>}*<z1od*?Zx?)t4f~62pL)VMDI8vs;S6?Be4|24i!s<dWJr4VMXqxs
z{h$X6cKf+(-XL_>3@MR@mC_zW3B<KBn)WRsGi>H{7yP*tPX_(Alxv|&#+oS5gevr(
zKH!THk{`5JFt>dDB{Q^v;K#CVFl8#|nf8lC;lh{BQ*l_q`vVg?!sgG<72My~<a%SJ
z8EvxiD*8StPeW^;|Iku}K9`)AQYe3LqW*r_STTu6P`R7)M*r4ZLSXQ-XleAXAANQ3
z5bMxB&kM&`&+c*HoGyxcyi3Mi&nR19y^Z-jwtGs@#`hA6VFR*KAuYHmSeLaTsFH2G
znO^>Qs%p}q!L#&Bcz-oZ-4lQF)0=_kAv}znom9=iSf%@~+urXnXb<&OFj}tuP&6uk
z&Y+vLx-6x%>{@Dju)MUerLCDR@7z67y3C4{$Y0+qeomE@g}rs2W-=}TD_<42-!H)#
zbdBB}uFrChaaKqYVdV4clAvMz(9jpLn2xdKZJCx!feu#T_v+6jt?~t0ajv-DEw}*<
z#9J?6f1RqJU6}7f_Yh7UmycG!<F?XKtpGQ*4u}tf_x9qmR(Hp;(orpUjM3>k;WMP;
zL!Z+Te?p@#i;tRAru{)D418$3OR!cIUPFzDcv=72opg7g_y24eW;XJJ_IWzjk3AyZ
zeL6B^<Ifi(b1wg9X(@||!i*)k9NdJane5V{I;#&p={Gs$_any+BFFY3duk1~%xL7i
zd2w<v6Ey<lZRU^zC;B>-?H7zK8&o(a?5VMdZdF5T6%{xaaj!qtW+=6Sd#-`~SbH9G
zV^09lGKZu_=oiT@t7y-p$<!L^qAn3NQMA}sYWQ*OY{9w{v+4kAE<2Zs>Dt0fj&p#n
zA+mSik#}pvyB@p66Kn3Uk6s5|O^#*stq}GgeQWTBe`&7A%IZf<`J;pu+}16<=@V2{
zv0##mEc^zXY)=EaT76<_s+vo9$%&u*z!oO0Mocd)S%#}9CYx8a(wgwU8sa+tYS7{J
z^;0}guc`T8xgD}1(EPMGgE~A+#IrdekZwu&?cAyKlo70dzSX(pEPKxYt1n$eLiLMJ
z?eIHU#O};R^v7NP7q>^N(rChp7c^xU5yaC>6Z=m05zH@drMu$p-ROLS|6ar`Ag(eM
zn%i$s;dp!a#f8ZD8Akr<!I%vB&oQ|3Ke9mf6P2vZGcOLek#8+`kOV<K36Pnvbn&(d
z(r$rh_U$JOe2A8De+|x+h=bS=P)7Q;i(EuKASbh_vD*Lb;2R#>hN-{25miTA&R50J
zZ*mtk#C}`vg^CHBpNW2ESHZ*5Gg!Vju9sJ`>reofdmU93q!dgu+y!xtZ~?C}9KDVm
z3FS~2_Sh2H@ULH&0k3lmkTKER1LNguJGZ7J*Lk$xK}B0*wdpp9pihrg@0S;74KY0P
zSCjiS0X>JGLAIL|YRv0mboAWyWpcvTDm2!tx(Mt;S6)W0jSJ#x{O#4{+P*%$X2QgK
z_KL$CRz^Z*h(kicE-FzV{w(XomSS!e&D_zqsfy9q-EWW1g<Vzs`_TVh@WVhqMYqHI
ztHxi$E0n@c|5krRy*JHFb=@@jtJ`n&IDzn#cki2=?YUOmXJ<Z)zn2j$C=C4kRITUM
zLU>rxCeiQhNF7f(QP}$s_Q)6uOP$_Do^=wGevAOtH(lb1%YU+ug<_GE8H^bsJL0{l
z5x;fsBRDU@quRS4t&GJqA7`+X9>}>0pFI)v@R|TTkTK~zgADK6w*y2ih2FoH8xsQc
z{&vmCd&sMNlM(2xNXU@qJ}=<AfI5-gddg`Le#MWd8Rq{oLeYVQ(kQgo(1D%B64PBa
z6G5Rt6)(BZ#GG71^a}?kxhv~_?807ODPM&DHn|&lW8xFBjuiwjx(NF@s|wtTx8^6?
z{<-;P4)!Sc7jiEkpow?YmA{`O{xmQo2OU&Sz?%`P%|@ZAWh{wmIes^5Jw*Z+n+q%d
zz)noF1D)+S5D~{DoWXC4A9_`*>ojL%mkYehRd$ivnNpolLkT|tQYu^9nYWBA5#cH_
z*_`1!2V>G}V3<?lk6ZuJ5jI&K65e!uqCdkgzgXK*i;KNckthFW=F_10yQWvuX-G4}
zdO=BQb=TL-{B@1WRV{M*d-AL%%ZxTrn_fS@yl!M-5d&4K0qu%zlxe5)JX5!W6cAsA
zYmeZpqr@q*o^}cyd}RP@UGd3`+|kJf>`OFULC_>I<JtN${f&ac+(%Ii74Ie(+Le$9
zB)49Vw6WiV7MNq(gl7?^e68qGzvN3<p5a4Ir$7ZG1v#X|Pzd>H-X<jXruEgxdX~Ux
zGv^zlDUIh$7W_X(^)*_Tzptsi8EDaqY#lpOXHDC=-XjpRP`dB>_mehyba;n)HDetA
z=Al-P!pi;GqDPhn=rNV36C)puo7de|Cl?d8H5AjRSCRY{ib4abOL5fV&kYBIMz~&J
zyIOd*ZxZX?EBV97Mkgv3pi|XE$J-C?#BlPG<&Qwf>yt0-j&()v(fTs|g4^M@j_trG
zD+*gLe!8z_^N36E?k|@0;%{D8Ey$0W>B$SK!uI5jhj>YkX%6tP^7y{ot?-+_rxn(M
z4?iKBhcaskkk<>qOvCq5s_y7X4jfqgg}-3&LC_*RCT7R_-CD*DYQHqFv)?z}d6X(P
z(V2TyU-4Yau!f=)mp|RixhXBJ!2X}Z3)dQ*<DIKjeR5704f98ZFL0#$M6?WWD-MNK
zCw1cKW~JtDc_d?o_IWB@>(^I^1C66(OXkJLh?#~n>5!4d#1^TD3CzfztWO2ETLJx@
zO`v}JCkfIhOzkYj6NgSwW(_8r4&qf_5g~dcOmp?(Ct`7U#@ZTl>Wg$auX^FWc;|<S
zN<<Nt>Uz!>urQh(p>SC<Es-oEvUJq5v~&r$%<yCGb?V49G!zhO<EYg?^+0PwC<njA
z&HH7OsOe39<V2EhcO}F5ha2C1yD$&;E^pa|rFG%x(+{59Co5ei7mOnm9+21F;mkif
z9PyP+P>NXLR9Z~Im65?AhZFg)XdS6U*j*4$Pk+(QRsGII1#@`Q+IAa@k+g1)NUJ*(
zW%N0`1Ewh#c>wqdv2YquSCTycaCPp=mE@c|<i1aTyP>D7#>xs{u_SH<lvk_;<z*qb
zvuT?WM|$$}(M9bHUqekIY`U3Nq}w>seyCQ)JX;g9T-h|Lb&$^pdeyd7i!S%6T;mIA
zQDM9CFzE0`c4z<C*w1S>>29SjxP>U>S0LTI7fK!qTGa<pS52JSUb)h|6~Y-5@JPpB
zMe8z{fs=fcKu-=})ropWc8rk#mlD&)-mzu}{CBuc$9Ef)RlhI=#NiJR%I1hNy^alg
z{Nt|P_-2PQd4Fy^{DQ6`MoeG5KJ*%<l5_H7By22HLzrfIm;YH-OvnuH1z2vxbq`nG
zkLHRy7}79&>nrr~{SHiPk^rM%8}D4fDAb`g3h^(kp%T*w9%&$M5E`aTI=m0U)WQOt
zZpNMmAAJw=N?lI8Ckp3Hun0^=f(V1&%#0?i_3Y%Xb<Oi+5Lil{5+s{#;No>M2~as{
zD5f@v%>n%jrAvY9aKE8fZFqV!7`)&B<PMLk$jpC>)ui{(;)p28;Lj^g%Std8V+d7)
zs2h2^*F%igoLOFvWcf$jin;b%NI8OCr+9Vq?_GCd=gS*jo@Zy{G7UWSo_!k3+}g&K
zl;*V7?pV|R!O-o%t}eq%4#1sf%SSw%CKvvEEpr)Vd31p1ga7vTC(uuZB`%_Ob_<Hn
zlI}VNj5T+s!43@K!wnfR_sPW-R>R%#qa!Lmb_-l-!3teucheaGhgb;#-^43KKZAoU
zD<WFpJ5kt;%qlyqEwe&>NP%O@yk4^hN1UK^hukrac!D3=wJs-p)AQkbPejk8?Xro2
z0Hwpc4MCM^!u}AQetfq4XX1qVFi|ouP(;C4Wg*AO`xVamru`sv`w5xf@&Q?vcc3rQ
zze$_rG`{lfCpg)Nb--Kg`~Y;y1l<oT&Byp>Q3qQ#7e562^0;iv?BbwX(6`M(wk~`6
zA!tmQYV8?o#b`p9bcDxfPV`N0CV^kx7f5Eum)MbGFBru6mc7}b<AnW^S=XA?_nz+@
zUOTta3L#&*>~oKu$Xn`uUOs!*x(!Lu4Y_LD1;=$&9!icnt(j1dB7dQxF}^`}_SHY-
z_N|I#7qd>A-4`-$(-*ovfC&1{ZT_7uY0<Ig^LFAJiRZfE^J%ZM-Kf5eCiuM=cD8m_
zu-Gq}f9Kffyq(Y{a^Hf*h;RQXDVas!B6Z+ZHABlE4U^xCa-Nlt?pe>i>sG3E&fDC%
z|0xyFb^`6%KJfCxAczr6?LU$MIr-hqpzbY&a`;~pQKwA!KLc|FryRU74TeU?pR9hv
zEW^p@$4p$AIq&%4GQR7n+`|!~$ko&CxC^{F6;%N`0-_mL)eG*_$O;U5Yp<*DB+_LL
zH!^|qoFFJ1WK@9C;Qv(UhEK4AD)9MFa4>3~AB5cHG*gfEA`OTGu5Ku`-k}c)+J=o`
zeO%y$3AJIJceQ-Tk6N~2VnMYBzu;F3)ef_Wj#6Z^9koHoiR>6z4kjpzL96-_P<Z7s
zxMxJRt|Mf*?U6ABe=tE$aT*4vF@Jvu9@Nj3t5#)>Q^n6;7@7=${UWF<9_Z*9XxXoQ
z8>r}eagCpG{^sX_TWTWX?tAYZ2vq#Kv=VtW9cP=1DKprl$9uD^S3Fxg=N8|h^D!3a
zc7k)6%MkDL1@LwBsDqis-gU_X0{`*N<6^SM&D2kn9zLI5g7?MG;1fp_*YZ>IMVwxz
zkIZY*V^+^zKaVIhJTqQMEY3}BY`5Zv%XrlJd<&g)sU|*qef#+n>4jTU$9O-|fb(2O
zJ*56lw&nn~S`Y4sm%e|J24s#T%ibj!iewQxYF=C$wEMVSQyi#hlS|KX-NQ684k$O1
z7)JUL8Hz3(c!Qkw+g(ug$Cx{sWFwE`ULs)8?aUIY&X!TZ!{G&D`sPM`W%k6q?6UFX
z-0ZfXQSkNt8VSnaqdfEWgQd5VVIrBoDSKZ@3_m{xB#IUDL7lxxJnNEnS!m0>YW+9&
zP8!7s1nB;-%+O~XUXooTls{zmpJl~v$ZkKJT>ix!d0y~*rz)pWtQ5P0vcZ#D*BdZx
z6|=4A5l5pB2vyD%#Uma)<mPIGUGWHVomdS$GBP(DJ%dQ{Nm%3{(3L9qOb)+55v);^
zQZvW7@dmj56d-RVXLpEMm_zRfg?=VxT-3b$ou%^ro34)dvaWO8F1{C{^r|!Sa2Ep1
z*<J-|cxkb_IiBMYOsS{iaS=^?_V)H1*85|^U`j^ar<6&NAS~wC;IH5{U{@H&luvI?
zmgOm5r{)R+-x!GD14rsao}0|!A@igUx;y${YBFLod<IWX?+j;d<lgB*MC^~Z2oWDO
zu5nJy!Q8vx(_ghje6vD)aTA}&9(m<oQ4T%pP;t>-Nj9-4eg$)7AMzoq>7pTO2Q0`;
zObhGud!a4#pF{IPX6=Hdk~k~iS68B$-CT6sRhVhI$~r$5ZH{`xs@qsIIhG2_?|8k;
z>+6qy&ic6YX%Ex0hxubKyS~kw`0;n}X}dz-y#1vK^DXuTuKM;yS3k}QEQ6l}xoOk*
zQ0-&CPvPaEMK_6|{mTQy3C%r^GvtK;;+b|`ERrV;#AvOzcPy==F4q3-g*xw&>RujE
zCAT=7u~Pybmj?4HN2jlN4p97FMUJk5QV$ROTz(Dzdt!G4jY?%A>|vfP*$wwSthfQ^
zVWw4Bcrvn9Nnm)l(w!#KZFmqmh$YWNgM`s#ThWAAAjc=+62&D5XgK%)jeI%?gTEq5
z(q{mDhYpG`RKvxO?pVv!MJQsYzCsdU>?i%5v>WkWQEw+);ZGlSS8s?JDn)1EE7dlw
zNx#L@jix%ZBI2y{-JTi_K`KHJV)zAM=)uQFje=-55<UO0w%u>)9aIBleB$vj?DHtg
zu@gHWcO0SOD~9`g@cAQ}&SP!prx)wR;hW!z-l0i66O!d0HO8M&uaEMDZjLimT#I&Q
zPLAR2W^ao<Tfu+T&@);0aDo0AvHFTGy5#y@MOiM{#1o6sI{aXNf$X*n7oyy;Ar|KN
zud4(#&0>(~vY84ZJ`ka@HwC}ErhtAl%)sIB?IOAbpEx0qrl#oPkqV1kL<i%(Gg%-6
zbfFI{h}%;y8<W)d-kgWHcoS1XgAYP(+l2<xr<jtPP=|h!WTxZrJMhJSK7Tl0Nhj@p
zMOeDAr{_OwrF4_5l(|X4iNDLJH(E4bp;Rpz7x;8MP)Ol0t4QH)S?DWYfaj|{mnRDE
z`E7=RyA5}{xQXGfB~x~82o3<57CqCpelHJu-kuJ}{o>}1(8t9@h&C@96jy*34FkYV
z<3SqBb!`&8%xx0=ywMN?Mf+*=LlOpIXF5W)a{GP1v^2abIjB6$@K2y_Srmu)=a(Cd
z&4gTGPCXfNdyPQ<I=t0rC&m!T%+&iv^{mJ`q4T$=afFJQOn2Qy{|54Ly?xmTofh%4
zZ{Mq_u~&Ke$v=R!M|NM|c*xPj!L}4D!G=)#yM#Kn3#5;@tv(J2zLM>usBZgHl(^%6
zoMu(L-apHMS;fL&9%Pi}5X9Eo_=JJg&J|xtPv&Q~%iKXu-M@XZR9;~s?ANvH1^dqH
z$EfanGL2v_kE7CZ{g}Y=BO1Ie0UCb$7>3lV<$yO~5zaK+98y(R`q(xBoO=-T8(ym;
zrt(kwnuE>n+f)7vH@Q_%ybmAcr#%x>B3zaFmAGx8cY4|5C>%*nxhWunz_QOYMqw<=
z>I~iS=S`|nPh)6)P&_<Z;Gp-jO6BFBi?I?vXVm_M7l(XtV@=EL0f&e2TbWpzraFqN
zcO#+$T0#8wRS@la*5U}Z1YY-Vfy+-e<H4x<$EzI_l|!NQF_nv5G_k0%P?1b>YU{*^
z)l2;5SQ6gmQyia4$LQR7lNZ)ths62(oLM4!q^t_-)$58qPq|CXFMMcVSa(rwq82K#
zNdGl7pL)^c6zj+jW4%Iw`i!ulzS@#p&w`>cw3*w<g??0aCPz!7t_1UZl#@Oe^Du7G
zZ_Z>!^kDEQpWI)KXDpRVmG8{;zuxp8OTIth^PTv?lPWM)aCvk_*tq=t-cz$J8P}g-
z8NoZgeQG~`Ml`YHT)xUppG3(K4o|(<_zG!l0ku&AG@#X&;_H{IoBUsdV2Nb!1%`L+
zDKZiBki4)EtNM}Xi1)A|7btQJyax53)35Eqyhm;gS&)4Wf5H7&8`81Jm0pVY*w{L@
zjUgpdecpw=QzCcVUIxjpe&DS6hh>V24EmR$D^LC<6k$gR1=n}rf_Gy;B<C*H-P@VU
zD^|M}wx*wiS(#bAy(YHyLEl7YO;kkv1R?7}lKCYYb;m&1qncy!(&qy#LKe<6X8Gkf
z*NIhmM|e3-S<U(@yH2QI1ucgY>o6<S)7cNl^h~3>^9GxE9w;!kcv}sd9mF4VUx2Fz
zDUsET(U6S=7)Yqq4Vsh1?n5*bSBnzY2{}eC6%b%Uk}5llY?~GEn&ah2IMsRoz_<?w
z`(u0TxBQn!wnhp>hbYJNDdNMGurVt0uMJ%M-;7|@5108e3-20mdhb2WK>YvB2*{`w
zpY_=3u3djs?UsSZ34H$kX2%YtRHEBXXFT>}iOxB7I>583bR^ezDzb0(&enTtL$OG~
zNR|&`QAw7{-D(?G#uA@AL@1_%@9#&^Ewi|7+|v~T;=tR_+Wjusyb%yW@rz#QYvJir
zQ>3}=okcR)CEZ-qT(E|W!&RY#OieV5hu)`_Q|d35Afz)0kF!-T-HdrK7^<Yg<~bU$
zW^iGmE#hpd`;Bvvk1swJD7j&$&@Cb@2NO;8s=t)3%>20QVEWKMq=M(#tIxIBuAHZC
z56z<zSAThu9~o*VgE3$r3yP#9BadZ<0_VyLc!LeC5(0c(3~RlF!sFvB4CjQfn9D@X
z@ouOy&*urQ5>}9%0hpQJIW^D2Z#=y}qQTs&TcNZI<I447?S|~ojcRcn_#t;?dXSSy
z_xCSgtRiYY{s7YPlCc4U*JdBf80{kW*%N`ktg*j0xu(LQAGEq5)*0jaU)9`dKV~*>
z*r97p@=ts<qRN)eILG-)ZBO>1bf7C?QG+u`p*?Kf$uq~6d!f3jZ9CMS6~V)a@I3fR
z(0gpVTvJWnTGar8e=S~b{~a-fd!V|$2UiUALG8OphT0AAC4g+fh)tsHzY82&G$?7Q
zvEnf>R$)pJ=Zh8F4jK1O{AJJF(Y}l?&NOgf%^kP%FAq!C2(GW$W=xDIfiCpn_MWN#
z)P7CJwxqSMFKBKbBAl;jvcjeG*G9`O!=0M@l4wH7uX`6YzIpH{n{L&CMU>LnS7Ez(
zGuN=T8K<zAQxabaaFuVsO@$s}db=_!bCx=Y*Lw@zZ=L_J2lOrN$re%|vLDI>GGHHX
zaI*vY#&m3B2WEC2oPVboil_H0UC`Ku#hG*Zonq;PSU*z*;pyusxsD!@-hU7XNlQk)
zc=O*i#utr1w%nxjt0a_14(3T5raTQpCHki@hZ}WP*4a0bI1cC23|9+8oRAvXD&e;5
z`hdsW3@JF5==>$iuvo!t_j<zSeQPVp`V$lGSwF7iULuzYwfOats2#-Oh&4!17z@0*
zV#zcKWgxp#nU%{;OGpryQ|2fBPlk3UG(Hx!0Fj#7)mNVU5stRiY-jC=b9fj?>qF>M
z?Ub^ley+DoU0spSkurC=guTtZJm-z=PN*1lkmlDJmFi%Tag?}*)Z>87*cVz(DdwYI
ze<~TdBLoP4CszKZ7970J5{QyAKKZkpDP^vVdWLnoKT4lw9@Y07sksA7wkOg@mw|N3
zp%{O7W#@vjiDIG|{YCvRQ2k|=RJ_MM;+g1GR>}q7JpzcDCBvx3JEk`DVD15WXdAOF
z5O;%Ipyb`kSTm!5#Z(30x4iMoA!O@P(Eg<#VlrI9Su<iSwW=1xRsJh?!!t1gl$Bn>
zqCDPU(ngvR@td58f-9GnWIP%i5F9?nBfA_+%stLm*+o5d>N<NoY{Z0J5bMDT&8HNo
zH@oBZUuA^q5I4FmZSqZ)5?S2{9$Mgx8!kRc21G~pQ^urO;2N4JpJ7Ml*mzQcLq1Dg
zfON#U1a73ij4!x6S;5UM%@{X)lk_3KGFsRrFZGoUM5p+s;oIP&BvOO7Q6ImNt)hdN
zw|Mdwwcio3t-7hgEQNyHKCQQCH|~3VR?eMKSg?%<+s1^gV(t+&Pf8QtOJJ?fVJ-FR
z1E#clpmCSm<PN1=F|y-B%5(|<+f=+H7Q>E{Y*4QLGJTcAo?jsJR$iff;H1z6%Be4U
z;VR|dyU~J{f$MHYnemid3_3x8!WQ8P_GH#&yvLn>9T_P#5(NQLLDDL)jVEK+v8d>I
z`Rd*mESZ=%9|_Bnq3W*2XS*B{)94gsgc-HNyr(@;-zFXbYvbe>;y~~(J0e2A(QU2Z
z^HiF=G@726uAw&e_3U-_W$l6vl+{4xi<UPaBpq)pRzvn-<Y-;;{P<|~5He`)T7Z*^
zl<$G9$kK~wK=2E@I6+U|Tw1+})NEQ1_%xE~;ztkri^_vIYb8l_+*VYi`Wg7D;1ujy
zxB)>DnpB1w17@H5KgZ2rm!i!Pq-fI8j_u*xpY4E0+*^d6K<FR<svHt;!Ar{5Aw3Y|
zLtZru1kS_Vd|0TpJgRkCSs9AH291Zf=pP&q9ue*}e}}u1Jsa!Lm7_||#OL)vKQU|X
zSIkBQAM<t_QU_SZ&ZfrR=KZ9q`>DS3ONQF**TsL5YR=s-k_ez?W?~XomI&{R^G(y(
zuX-k)t^0shRR1(t+geuEp*ij9s8;{tU9Fl~AISzgM8!tzj4$Kb`FT|2A!I?+)VdBr
z7s>GgOM;iQ;9#@U#h6?WxrEB8?>*WIZnA~$Yb7k^(=O8Gw`8d!x#&0CS+TS<N`uO$
zf)0z6idB{aJAL>eW%|aJ{yd^z^`<g0{Vg!N`(qo%A_Oi+cH=PuOV<yLjr6mU{_S77
zHdyQ&%O(tEVoduz%DRc4RS_fo3vmJaiXp5$9yyO}JD-UQ1QRQYg@-Q`G4q9DW6q}a
z-?gA@%bT}DI!<Q_xSqZri@l+g5&U4Ll$c%x%(N^(xaegpDgiGU8E_}2KLLEQ;@?F&
zTNMf|!^a@UCISvqwjXL19^ZEd#C;fUa`*3y7$Ltjy6xoGCi?Oz4L2JhbUcwk^SXd~
zq2JoK5*NpOJFQzqXXVp=Q(N-pop#I<5z6YewxzS>*)`j$hbMjCv_EUoc>Ra(EoIs{
z)Q6)^hW8OA=Nf=~WWeCfAI8;`My<<<fn#^NpiiYoFzzsi`ntmJK+bvzJU@DpeQI%c
zN6aYn`3g^t0w;NtC;Mrm)=?3~c=}Qi;+dKLEoyS;s(M*4F+CrcVbunpe4>gLLncVd
zNbeN+UxNOVp?d_dQ$4@|q0{5>TzWAm6Vr~5RZK#Su82w^F^!tl2{<mbjmbXRD0aC{
zTST%J=Yn5vOvH6bEJuW|H@fM>+RN{Q?n-`_Yvxd}zeDm<sA){bl;F$qz&a9nZkxF0
z%^`Ye0|_+5!6g#w@tDMZNVGpFe|$#M319h_vbxmzT~MJ?$@AbH_|QY!7O5Y$5_h9T
zKXH^lInG=qFi|Mfk?^}b%q&jmM4Z2YUl_VKGtL~x@OO>tg3s+^4FztuMwhVy8(%-(
zR{v^6;4T7C)mofkQDmw1r(wg-o;<#$TlL_I*pNhd&D7xe`!sdmiGCU~{+&eYdx$uO
z(?<<iR%BEX8^QjQ5}k4Hd%**<Okny2;=^JB@dgTfCY&Fw`X#U!E%J#6L$5BMvz>SD
zz-j6Z9z;M<>TU3tAvAnC_}CC`x%M8>&*9&1^NL*Cfd>cJlXYyZlG<vvVYx4XmLi!y
zl$b66g5-#%WoERB*sPkKQFSGPUlkp~r^jNwWy@qoug~L@V1x-%$o0z%dxbd`#I73|
ziXHr@kK7xl-szwCu77o3<^0c{yR|h_4z<xXzT|L3YT(30DDpdr;|tWh3K%Tk;5`T<
z|LSsHs9AwyMfZ>aiX=xe%N|mRf>Y7xy$~<yvly>iu!-yNLQ#!U`#YAGtmjuGsF=dF
zTX%W8CzS*ydQR?#{|vu+{F3J7aw+}9Vmn043E2UL6W_W8WsTsyYs$A0IB5xEL4P3Z
zRub=dVYTcY&oVrIYxoZ)k%;caqmE;bE}DYSq2zxE{X~p(Bn2^y_;^QhAF+;;+@aMV
zG(X0_#JGzTsTmRk(n2a7Py?gZ*vzNilM0f{@L4l$x+S=3x#8|qtt~m4QlSi(7zbV=
z3lQh>Kk_tkMLTwb2bK5m^o9fASFr6+!xCt{gD9YYL!SrR#fp0k8vYWLpkqhLb-XMJ
zBR!w_w2Q{F&Xs{7g4IzdIP^PRqoKb5=cY}dXmjaRd7#b=k%dCQeOaI;g}puiDIY)r
z6e%>8SXN6k3ySTA$}Wk?TnFZ6*U5EX0^3ZS`5pbseyVV`!96RQ0)tz-Zm%O)6b;eG
z<Q_^$Jbmol=zKFTq1eGet1C5g`WjW7?p>zM%We108Eq-B;Y$BbfnJOr@$?Aqdfk-0
zBJEX4@GzImg@#hHh?2tB$?2WV=`8(k|4B?ZU($~CCZ&PL@dgfL3rTQ&X~CJi-qT{k
ze?JK?nE|LSUi>b~SEwms4lm*`x9=Ia(D@5)m^}m)9yaFSCBL1<OX?FQhM>tcAT6#-
zj9<iiIQ9EDeK%hM!=g)|`Y9Ir!AR+z34*_>_XmYQOWGLPBb8JdSV&F3DOI6byl+gm
zdDz|hl$L+xt5No0V{j&~^7G&F4z<NYdPLLk-zbhK>~}gpOrh-0ueF9?@lW0I`>exq
zf?jUKZDJ+D#C?ZHw+NuC%0WT<3W{M+1_(+I?Q8ln5C6A&-El=xl-6F30|9@|dAt^u
zw$@eZeo~Ql``j}<<vXt~$bF<S$R?_%IO%SdDJMoGkLtj#J}>amt*;1kg(XWZND|Oj
z?Ts1E2~yjc*kyPMgfA~M+R=q^$`x3M$prhmk~zuuZ`8Ral-zwz8T*xrpBg(DV*-?N
zji{*$B#(8eFQJ&18bWN+h?2UAAWVZ1;Y|*0B@k%|kXeVv`oOZ43d;62CjlZDp%qbd
zD0O41>?BMV=SU|}gQhd;4QABF9Cr8g-E|am#K9ZQ;~g&tV7-TFx{Y2pAG}K9T1}sP
zT&>ihfd)bupi9afM^7zBMB7&Y9Xqx3OB?>*X~lHAAPI`_rI*hiPR9PzXPGcrxIR1Z
z5pRwJ`=$uJ9k`jUPZ|z+(&6t?cYH^n)Q!l@OMLM}ktnt~_!IXc#~XZmTOF4eryOuw
zj=_%*%AB81qA=ubI7FuZOuF2puo`5iN?w@$3)#3Xp`3*J2@pj`+4U??9{dqkNgTdP
zf=QTNUV=|hu=v}-M0!^Eauc|W>|4C&-b(zt!}@tjxkXtyLsl0w=I0MU<U?eutLhe#
zdr$kh<KQVZ#ME+=WKgiF@`}%#P3ywJmWTVkZSW*9kTCEj%6%3ygIQ%-#V?Ua<f$ue
z)yi)E>h)*uE~l^SC$|@k>R#CUrV9*jZ`_DesOscQaYkQ$s3#ZVkBUeff1^CoFMi?N
z2eAh|ave1{@$A3Y>Jy8vlyaZfqkEP&s`RJDr>26;Q^hmw>ekFXy@$mt>hV~SLxn&o
z%zktaFy^iwL4As;o{;Q+_QGJ3vN#?RX0l6=EK0A<!}7ON+xmPR_MT${Q9}vukWPH7
zTHH$JDW7tF7s+bZubWbRQo%BVYnv)tssl@)lCL;lMAiw|DPz?ho#gKK^^i?rV*y#F
zU{QdKJ|#~=B^`TA!T;P5t9M)ANdN^rWbOgYK9e|Lk8s5n^p8YMe2w44A-z-nI#IQ?
z1bn?AX0oEbyI{8fsjEPponS9j=~aB$-RWXw?j0|GqNNSdKrVU8t<u|*Xt0lNmCU_x
z9*}St3|K#arpd*u>-g0jTb(-itg^DZD58$UAM6)EUlk7oY|n$$1D6qSSq2nI4~a_b
zL96(oz}Q~MX$h*eEYASP|D0R@+J>44-}ivK(Z0>}(S>`!XGt*(&ZKo@FD6a-^91Bn
z@o@V27;QaX<-0*imgJU4>r}+<7SSfR4dbMM_Xd4_9N5bE*MKvZ+<-d7k6>-F_aq}!
zi&X#YcZH+ht4$m<r8gaVhwcsylO|yOlJQEPuQyN6{-F<#Cv*`%^<}4&8A>ms(cvVy
zkMC}={j?9Mu&fz|2D#ImOrw}eY0a@bUB((}RbMl%%(*YVN%VG|%@x!5ppX&R_n_qs
zz1lkw0lqP<i#>WiScb>rL4Sgk_t>qe!f(BcnZn7u0fSm+f)1gxlt1hLLcf#&|39V2
z0BVku-sE@V9r(I)7jbK9sg7>$n~BC)MXu+Ui~hq>Bk}Jfl01K&Suo7c4D%FVj%`JD
zu`cp+y0cK;(P^IG<9ri#{uPYmPVd+eKjW`*;j{ge`3@?PN<oN>)+WIju40c&p67G4
ze;~QWs$dt^)q|__r5G=oMf7?UD6j6oudOOjAR!WDjxQ|iDTe7S4@z-e?&#E1e8;Cj
zG}w5{3PL?kzMkAcA(8HT?i+QqbDQtW5wnQcf<wzJd9tu8afp`ujq;WP@L<O#u@&Zm
z@4Npe6fy0GDyJ`+_CL?L?)i3HOWo{2q~<-Z2N9Ys(kbKZIs!qALDhx}7FL&XTPw6m
zYws#HDkvDX7~L_IT@ae*7gR_}eJv|V_0hKHhDO$}JTV4Z?bgx5O$d+XQao4jrDkdg
zXOnSi2NNdo#fp4_;|SW}kGe@H3b)*pCRzVa)uuf-dcQC#hZPXKPJD-4r|@&HhfQ_T
zm7nJHsE3}nM|Lf~KPeUo!)<rdb!GhMNW$!e59OR>jq&}KvnYLhlP7*R!lP{mMML!s
z2Kdu}2pgGS=BTwvY?*nWBk+is1LU)_!F?ci6%AO>@lR9DzxT&4JXu%peH1{f|AvE6
zjWii9YF@%dZA<g@*Umkny1Pq#?8dk*<RKz`|J_&jEK%uhT40hque69bDSJ_q{9_0j
zUV<WjzXnzRY)C`yO5DIVVB4bc8~*d^MT5bkYyw{inblL1`bm$aH#BXeKdRMSQ8|5`
zWDmQ=cI9NN5blKeM5LTde}!{({vTl{v*Zt#Sm4DP6@JDQ4XAg;G1uct*xEuCKgpjy
zF6bFNJg)BD9#KtugXjCu`h$qfQd9Mu8m$WIf*Zr@Oo?-BTHZ*lg#RBTWP!2eVR^P!
z%>TkDORdLf$8XSQ%8*&hkXYBgLwAxu@CflWD^pV+(U%St%bT;m&EYs$uy#gJcv-&v
zobfh4>O$w~nlA)0AMGbDeC(bo=j?ILSH880nhGa2pIBS)7R=yTm_e358SO-*XOkri
zQT^v2mNJ0GJIZ%>0P>*~9@;^v@9>LjND5&e#Q}+U*iEW-bOS1?Am5B_G7{_%M&7`7
zY#0@o86(MCu3EJmj#`|b&G3>x_b>?rxDq%|<AgJ?_{Y8pHkytR|Kbq(H$m;ciDyVt
zLt|J>k~W~W+G3b34Ts0PD%dlCN><kMyB}x^%nc($5o})R<I%TeT9nVRX!KO^!J{VM
zDo35xF+7N;w(}K#seI0>--*VQTEOF}{*>|30*ZPSL&;%m@$nfrB!kQZMTSWB*eIR>
zV@C_x{g`5(vc9eTixTcrp4sma(bzTC%f2pydf4SrQXCnh6%#ertJ42i!s9)DlX#)|
z#gq5e8$mZ)Y-(#b9BZQ;cy~}e>%f&5f2RN(XB^1}dUm-kbn>aLqDQdX4JkaN1Q<k*
zC3Js0MMtV41PXaR<->SKnVspd1yVh8Z(vR_IM}S$SWX*<c)Y9pU;>+65xLI_!~EG8
zvry+ZB)bQ+r}rQ!-T|VuZa^+b?0hXLY`Sqjkc48Y+`$<@q6US)Z{GPUaRPO@te(8(
zA#XrWQ2xm!?36t?T_1$20_RB!Z-ZA+``gD?(2L|dl0V2**PjnV_>Lv~$N+Y!W1(X!
zSKhjRcinnJKAwJx-Y<NJ1M5F-&~?}}bQl!uusQ%WG?1QSYY=o<Vc(V~(^*Nc4jb|N
zi%WBOi&}imX=9&8)C6oKemIuwg$*BRQ9%^V-dkAim>D6P!g)~nSLUJ?X*)U09{Jg;
z!L`b_3?;7I?ElWzp{Jwa*kgVw;X|78N$0a-KNn*r?+s2C&(_9;C5uL1l9uj$J*J>+
zQ_N;u`ynAMjMuGY*-GT=%xhL|x?2BU1gj(vTmcu2ACMo7b^@(GK<|-N(u=))NciY-
z^+rL^yt6)Ox!=QOsqU+>hy#LVS^h@nfO1RglKQU2@_mtKTr>)C_TC7l$=xh=ca64u
zS9--=4%NEM>$IEw<*k=Z*`j=frv_O|X`_JNG?<D4FTO*G#t}z$$0tgRbUf+5pfH?i
z08;XzAblomc~0{U-Yb}hb0$`%pSBW!`_4lZiKu>obS5l+o5RPFI2H{({3;LJF!Jr3
z=0#qz<NT$WYV?H!Y9mvX)?VoBA*450w*(&ykOTbae-^%*7H_>S#|J(vxw&|gjYzCb
zrOPPubx=VEy6v_-bW>-Zb-=Fmbj`jTQKa>&{>6N=@f7`S_!F+PUMWuu3#2ENlI%_r
zVkI4n9SD6N(>_Xt6{ywdGP4R9S|lN5unwoYD`&m7l0$Z6g=ceIm2CaLp&?NMBnQ7w
z=BLDOC;urMpS#!oXQq3h8_5)VvbOn~qNR#{%YV$+ZRz)zBQ80+zKaW)IF2t55{Q#3
zJk%0elt1_GMZ>(j_yw4&#&@Dd%qzPdk5lc-bh30BN-GWs&9_|rxp);-aAF=xo&Y(W
zAmfz`pfLx6M~qETQYt$Xa{C=8xCT`I>#pM*+T@yowG!N$ksXY9)`%lEx&WtrLTw`(
z(=NXe!LT?m^BB$S8{OsKn$EYX<ltQqpoi=IvIAoCGZPvhCgB)Y;0--;_oM#2HuLu|
zyaPMZ=T`W*3L>BRNYvB=lC}y*RFfO(|C(5Nls591fXZi|?8w;Zi{mt*T&Bm*x;!p2
zOgt%KbEo3Y`PaNX7BrQd>JFScR__wgQ4O(_yu$Y=uei(O*ExGz?$2XB-}#V%$@-|A
zs?*JFCD*m1k0+M&r1hHSe;U%?j|<w$|2<vtBiU`4u1aE>Vp1}DYyi>P9ZzpG0Ns%8
zI|}?K5}&-2cy|;M@qtJ#gE8Rpa3BM;ups#TQyYf+Jt~)V<u>>1tk&o1zVS2e<CV{&
zZ14FE96!DBA-EtVI?zVS#KCc>FJ~P68+9{xM$V;f;E7*DY#hir7V>BwO1aYk3DIWY
zC0_(cUQ?z4bE3HT!UecYH$(v}5kyKRY@#o|{f>n=_R&d1O6M;B`oQVcPEd%~;sQpw
zIBPb6ag!V81zyupss6a8Hvjeq9!4+>TxNtCY7gZ@Kt06*-laJC%mwI<D^BsH-I>*W
zYyD;lueL7yOAjANgJ*xvU1ot+k0UL0eKr*G;?}c4-izWVJy9a)UM5Y1F_ctfjc541
z&dI3JD|cZikC0FAbaFS=RNGefQk(J#uNa*eoXuc5D?vf*H$RWEw&;HDA%@u0ui24F
z@3_R_(m`Y`IGpH%X!{<3OsSXW!P&SPvJPX={u@J}zCwrP+E$G;`DMm9_oL7Z0sqET
zxqj6h3T}|~>)n=w&px{T;zf|!#L%i&gimBO!EoF(MO7J|X^|VCtPJyuy&GINX1y&5
zy}!+%J2v=yai#-}MzTwWFTm}Zo5qITp^?h&1IRGirc<U(ZcTHH^){(lTm1&NzzgKw
zMb9AOomxavInsFjeLkV5sV1(%<w8zlQ-uSvk;w8PQ&MOk_d~!{x%gZCx&sTv<}&<K
z_YGyocy!OmZd%%Q`p=9!Z|k;)-IbSDzs_CUH0AIXQMwcltjNcIEkeZ$sFQ4>cx$Qw
zBC`Fp`?0Em+7FMjLqKHYp51d%z!Vuh7i8XduZ%ThSI0s^{R*S7KI54LV_Wv)#7-)B
z5x2a%WD2`VSk=Til2Ay~{spE)LS!i?%$jdCS(i{MNaBPc6ZS-<`O9>M(^L4CBI9cs
zL!+oPnXg2}%li#h<USS|M7m~qsZM>LCc^F!NgNO<{QL6~9OijU`aUdS9iB8P1zZSE
zpV_I;LnOpa-}-t5vgGP^OXZ?43prg<v$j-3>a28~rM;b%<*AQWsYO)Sm0RO3X#1E#
zwKYagLv?h(_?-!o`2BZu>L-K0UnHHdvuh;SMqVf_oDfX>B`Rq9lIOt-xy1J7blQCH
zPncN?q4l(+1Ri5|x5M#!0SRC}1;XJgR{=WP)@~j}>rK*R<Y;8rj$=`hc}NOh(&5Y_
zhgP*4Q4M=D(^K_Ly0TL#d+lpPaipE}d{t?9MfQYck(_9D^9YAgtR9@mPW31Cgg3)#
z3Kcp!&^K<r<OVJMiR@dr`!wRBcFWF{;U9AvIDCupZvFW4c&cvDmv<=^=vf0Dxy>=+
zd)tJxebkYK;HbZ^8))$-LoHI44y<lD5&~1wcVSX8oK&t#QJv6|j+^n$*m0Z1iq4|9
z22UB2G9`IAg(GD@TP`Xr%uAV6?~A26=oaahBwx`jx@6R`5W{@)RPJzhO8Z;gV@%WC
rt7aL!7t+$ZF7;UmJvG`r_P}(;O*Y!0FHPbT2`-%Lf17IM1l9io!voy2

diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..7bc5c487e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:47:04+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:47:04+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:47:18+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:47:18+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.171 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.81 GB, active+cache 5.56 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.42 GB, active+cache 5.18 GB, phase peak 6.17 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          510.48 ms (1002.98 tok/s)
+  Decode:           3399.29 ms (37.65 tok/s)
+  MLX peak memory:  6.17 GB
+rocm_gpu_guard: 2026-10-09T01:47:26+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T01:47:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..701febea2
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:46:06+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:46:06+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:46:21+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:46:21+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.207 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.68 GB, active+cache 5.43 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.38 GB, active+cache 5.13 GB, phase peak 6.15 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          506.72 ms (1010.42 tok/s)
+  Decode:           3397.42 ms (37.68 tok/s)
+  MLX peak memory:  6.15 GB
+rocm_gpu_guard: 2026-10-09T01:46:31+0900 monitor: 7 samples
+rocm_gpu_guard: 2026-10-09T01:46:31+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..8d333433a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r1_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:46:36+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:46:36+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:46:51+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:46:51+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.200 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.38 GB, active+cache 5.13 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.38 GB, active+cache 5.13 GB, phase peak 6.15 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          506.60 ms (1010.65 tok/s)
+  Decode:           3356.95 ms (38.13 tok/s)
+  Decode path:      engine
+  MLX peak memory:  6.15 GB
+rocm_gpu_guard: 2026-10-09T01:46:59+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T01:46:59+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..9fc0d9f1a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:06:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:06:37+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:06:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:06:52+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.186 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.69 GB, active+cache 5.44 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.43 GB, active+cache 5.18 GB, phase peak 6.15 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          513.71 ms (996.67 tok/s)
+  Decode:           3382.28 ms (37.84 tok/s)
+  MLX peak memory:  6.15 GB
+rocm_gpu_guard: 2026-10-09T02:07:00+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:07:00+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..11e0ad7bf
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:07:05+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:07:05+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:07:20+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:07:20+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.189 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.81 GB, active+cache 5.56 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.43 GB, active+cache 5.18 GB, phase peak 6.15 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          507.96 ms (1007.95 tok/s)
+  Decode:           3388.98 ms (37.77 tok/s)
+  MLX peak memory:  6.15 GB
+rocm_gpu_guard: 2026-10-09T02:07:27+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:07:27+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..cb54b69d5
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:06:07+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:06:07+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:06:22+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:06:22+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.190 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.42 GB, active+cache 5.18 GB, phase peak 5.11 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.38 GB, active+cache 5.13 GB, phase peak 6.15 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          932.40 ms (549.12 tok/s)
+  Decode:           3416.33 ms (37.47 tok/s)
+  Decode path:      engine
+  MLX peak memory:  6.15 GB
+rocm_gpu_guard: 2026-10-09T02:06:32+0900 monitor: 7 samples
+rocm_gpu_guard: 2026-10-09T02:06:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..c0c263329
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:26:00+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:26:00+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:26:15+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:26:15+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.193 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.68 GB, active+cache 5.43 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.43 GB, active+cache 5.18 GB, phase peak 6.15 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          516.54 ms (991.20 tok/s)
+  Decode:           3403.14 ms (37.61 tok/s)
+  MLX peak memory:  6.15 GB
+rocm_gpu_guard: 2026-10-09T02:26:25+0900 monitor: 7 samples
+rocm_gpu_guard: 2026-10-09T02:26:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..aff3ad82c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:26:30+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:26:30+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:26:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:26:45+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.190 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.81 GB, active+cache 5.56 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.43 GB, active+cache 5.18 GB, phase peak 6.24 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          509.34 ms (1005.22 tok/s)
+  Decode:           3397.71 ms (37.67 tok/s)
+  MLX peak memory:  6.24 GB
+rocm_gpu_guard: 2026-10-09T02:26:53+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:26:53+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..16455f3bc
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Meta-Llama-3.1-8B-Instruct-4bit_pp512_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:26:58+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:26:58+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:27:12+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:27:12+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Meta-Llama-3.1-8B-Instruct-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.191 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 4.75 GB, cache 0.38 GB, active+cache 5.13 GB, phase peak 5.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.75 GB, cache 0.38 GB, active+cache 5.13 GB, phase peak 6.25 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          513.78 ms (996.53 tok/s)
+  Decode:           3398.23 ms (37.67 tok/s)
+  Decode path:      engine
+  MLX peak memory:  6.25 GB
+rocm_gpu_guard: 2026-10-09T02:27:20+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:27:20+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..1efd8e5a7
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:02:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:02:31+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:02:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:02:46+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.055 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 2.26 GB, active+cache 28.66 GB, phase peak 27.52 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.37 GB, active+cache 26.76 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19862.45 ms (25.78 tok/s)
+  Decode:           12226.42 ms (10.47 tok/s)
+  MLX peak memory:  27.52 GB
+rocm_gpu_guard: 2026-10-09T02:03:44+0900 monitor: 38 samples
+rocm_gpu_guard: 2026-10-09T02:03:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..53e1d3596
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:59:53+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:59:53+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:00:07+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:00:07+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.057 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 2.20 GB, active+cache 28.60 GB, phase peak 27.52 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.36 GB, active+cache 26.76 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19714.36 ms (25.97 tok/s)
+  Decode:           13213.77 ms (9.69 tok/s)
+  MLX peak memory:  27.52 GB
+rocm_gpu_guard: 2026-10-09T02:01:07+0900 monitor: 39 samples
+rocm_gpu_guard: 2026-10-09T02:01:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..d6ee35222
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r1_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:01:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:01:12+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:01:27+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:01:27+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.057 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 0.22 GB, active+cache 26.61 GB, phase peak 27.51 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.22 GB, active+cache 26.61 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19839.49 ms (25.81 tok/s)
+  Decode:           13411.61 ms (9.54 tok/s)
+  Decode path:      engine
+  MLX peak memory:  27.51 GB
+rocm_gpu_guard: 2026-10-09T02:02:26+0900 monitor: 39 samples
+rocm_gpu_guard: 2026-10-09T02:02:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..7c8d955a1
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:21:04+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:21:04+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:21:19+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:21:19+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.053 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 1.72 GB, active+cache 28.11 GB, phase peak 27.52 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.37 GB, active+cache 26.76 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19823.17 ms (25.83 tok/s)
+  Decode:           12615.91 ms (10.15 tok/s)
+  MLX peak memory:  27.52 GB
+rocm_gpu_guard: 2026-10-09T02:22:20+0900 monitor: 40 samples
+rocm_gpu_guard: 2026-10-09T02:22:20+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..1458cbfe7
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:22:25+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:22:25+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:22:39+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:22:39+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.070 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 2.26 GB, active+cache 28.66 GB, phase peak 27.52 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.37 GB, active+cache 26.76 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19874.59 ms (25.76 tok/s)
+  Decode:           12839.71 ms (9.97 tok/s)
+  MLX peak memory:  27.52 GB
+rocm_gpu_guard: 2026-10-09T02:23:38+0900 monitor: 38 samples
+rocm_gpu_guard: 2026-10-09T02:23:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..d8c70518a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:19:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:19:45+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:20:00+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:20:00+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.054 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 0.34 GB, active+cache 26.73 GB, phase peak 27.51 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.33 GB, active+cache 26.73 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19734.06 ms (25.94 tok/s)
+  Decode:           12681.93 ms (10.09 tok/s)
+  Decode path:      engine
+  MLX peak memory:  27.51 GB
+rocm_gpu_guard: 2026-10-09T02:20:59+0900 monitor: 39 samples
+rocm_gpu_guard: 2026-10-09T02:20:59+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..113b99b6d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:39:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:39:31+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:39:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:39:45+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.055 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 2.26 GB, active+cache 28.66 GB, phase peak 27.52 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.37 GB, active+cache 26.76 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19711.80 ms (25.97 tok/s)
+  Decode:           13062.31 ms (9.80 tok/s)
+  MLX peak memory:  27.52 GB
+rocm_gpu_guard: 2026-10-09T02:40:45+0900 monitor: 39 samples
+rocm_gpu_guard: 2026-10-09T02:40:45+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..0ddc91c4e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:40:50+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:40:50+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:41:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:41:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.054 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 2.26 GB, active+cache 28.66 GB, phase peak 27.52 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.37 GB, active+cache 26.76 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19663.82 ms (26.04 tok/s)
+  Decode:           12839.24 ms (9.97 tok/s)
+  MLX peak memory:  27.52 GB
+rocm_gpu_guard: 2026-10-09T02:42:18+0900 monitor: 48 samples
+rocm_gpu_guard: 2026-10-09T02:42:18+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..9ae178084
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Mixtral-8x7B-Instruct-v0.1-4bit_pp512_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:42:23+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:42:23+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:42:38+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:42:38+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Mixtral-8x7B-Instruct-v0.1-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.065 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 26.40 GB, cache 0.37 GB, active+cache 26.77 GB, phase peak 27.51 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(32768), reserved 128 for generation)
+[Memory] after measured pass: active 26.40 GB, cache 0.34 GB, active+cache 26.74 GB, phase peak 27.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          19878.37 ms (25.76 tok/s)
+  Decode:           12548.46 ms (10.20 tok/s)
+  Decode path:      engine
+  MLX peak memory:  27.51 GB
+rocm_gpu_guard: 2026-10-09T02:43:37+0900 monitor: 38 samples
+rocm_gpu_guard: 2026-10-09T02:43:37+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..1b37a199e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base-null.log
@@ -0,0 +1,27 @@
+rocm_gpu_guard: 2026-10-09T01:56:56+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:56:56+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:57:11+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:57:11+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[NemotronH] Loading config...
+[NemotronH] Config loaded: 52 layers (23 mamba, 6 attention, 0 mlp, 23 moe)
+[NemotronH] Loading weights from safetensors...
+[NemotronH] Building model...
+[NemotronH] Model loaded successfully
+[Load] wall: 0.321 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.30 GB, active+cache 18.32 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.07 GB, cache 0.21 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1851.40 ms (276.55 tok/s)
+  Decode:           1697.24 ms (75.42 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T01:57:26+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T01:57:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..19ea57f93
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_base.log
@@ -0,0 +1,27 @@
+rocm_gpu_guard: 2026-10-09T01:55:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:55:45+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:56:00+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:56:00+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[NemotronH] Loading config...
+[NemotronH] Config loaded: 52 layers (23 mamba, 6 attention, 0 mlp, 23 moe)
+[NemotronH] Loading weights from safetensors...
+[NemotronH] Building model...
+[NemotronH] Model loaded successfully
+[Load] wall: 0.327 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.30 GB, active+cache 18.32 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.07 GB, cache 0.21 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1827.53 ms (280.16 tok/s)
+  Decode:           1698.51 ms (75.36 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T01:56:16+0900 monitor: 11 samples
+rocm_gpu_guard: 2026-10-09T01:56:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..66f2305b7
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r1_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:56:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:56:21+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:56:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:56:36+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.330 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.17 GB, active+cache 18.18 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.01 GB, cache 0.26 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1719.00 ms (297.85 tok/s)
+  Decode:           1695.18 ms (75.51 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T01:56:51+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T01:56:51+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..faad783e9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base-null.log
@@ -0,0 +1,27 @@
+rocm_gpu_guard: 2026-10-09T02:16:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:16:12+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:16:27+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:16:27+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[NemotronH] Loading config...
+[NemotronH] Config loaded: 52 layers (23 mamba, 6 attention, 0 mlp, 23 moe)
+[NemotronH] Loading weights from safetensors...
+[NemotronH] Building model...
+[NemotronH] Model loaded successfully
+[Load] wall: 0.330 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.30 GB, active+cache 18.32 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.07 GB, cache 0.21 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1819.72 ms (281.36 tok/s)
+  Decode:           1698.86 ms (75.34 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:16:42+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:16:42+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..04b8096e8
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_base.log
@@ -0,0 +1,27 @@
+rocm_gpu_guard: 2026-10-09T02:16:47+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:16:47+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:17:01+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:17:01+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[NemotronH] Loading config...
+[NemotronH] Config loaded: 52 layers (23 mamba, 6 attention, 0 mlp, 23 moe)
+[NemotronH] Loading weights from safetensors...
+[NemotronH] Building model...
+[NemotronH] Model loaded successfully
+[Load] wall: 0.329 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.30 GB, active+cache 18.31 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.07 GB, cache 0.14 GB, active+cache 18.21 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1767.78 ms (289.63 tok/s)
+  Decode:           1703.63 ms (75.13 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:17:16+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:17:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..2d77f0b75
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:15:36+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:15:36+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:15:51+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:15:51+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.337 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.18 GB, active+cache 18.19 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.01 GB, cache 0.26 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          2000.43 ms (255.95 tok/s)
+  Decode:           1700.67 ms (75.26 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:16:07+0900 monitor: 11 samples
+rocm_gpu_guard: 2026-10-09T02:16:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..301464bad
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base-null.log
@@ -0,0 +1,27 @@
+rocm_gpu_guard: 2026-10-09T02:35:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:35:21+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:35:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:35:36+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[NemotronH] Loading config...
+[NemotronH] Config loaded: 52 layers (23 mamba, 6 attention, 0 mlp, 23 moe)
+[NemotronH] Loading weights from safetensors...
+[NemotronH] Building model...
+[NemotronH] Model loaded successfully
+[Load] wall: 0.332 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.30 GB, active+cache 18.31 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.07 GB, cache 0.21 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1878.15 ms (272.61 tok/s)
+  Decode:           1696.57 ms (75.45 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:35:52+0900 monitor: 11 samples
+rocm_gpu_guard: 2026-10-09T02:35:52+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..89793f6f2
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_base.log
@@ -0,0 +1,27 @@
+rocm_gpu_guard: 2026-10-09T02:35:57+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:35:57+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:36:12+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:36:12+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[NemotronH] Loading config...
+[NemotronH] Config loaded: 52 layers (23 mamba, 6 attention, 0 mlp, 23 moe)
+[NemotronH] Loading weights from safetensors...
+[NemotronH] Building model...
+[NemotronH] Model loaded successfully
+[Load] wall: 0.333 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.30 GB, active+cache 18.32 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.07 GB, cache 0.21 GB, active+cache 18.27 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1794.91 ms (285.25 tok/s)
+  Decode:           1700.18 ms (75.29 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:36:27+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:36:27+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..38c5bd715
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit_pp512_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:36:32+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:36:32+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:36:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:36:46+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.336 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 18.01 GB, cache 0.17 GB, active+cache 18.18 GB, phase peak 18.17 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 18.01 GB, cache 0.23 GB, active+cache 18.24 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1856.01 ms (275.86 tok/s)
+  Decode:           1694.83 ms (75.52 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:37:01+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:37:01+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base-null.log
new file mode 100644
index 000000000..04c95815c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:19:10+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:19:10+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:45:57+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:45:57+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.273 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 1.80 GB, active+cache 2.18 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          448.08 ms (4570.58 tok/s)
+  Decode:           590.32 ms (216.83 tok/s)
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-09T01:46:01+0900 monitor: 3 samples
+rocm_gpu_guard: 2026-10-09T01:46:01+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base.log
new file mode 100644
index 000000000..31b35ff19
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_base.log
@@ -0,0 +1,25 @@
+rocm_gpu_guard: 2026-10-08T21:33:51+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-08T22:19:00+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 2709s
+rocm_gpu_guard: 2026-10-08T22:19:00+0900 waiting for 60s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-08T22:19:18+0900 idle streak reset after 12s: gpu=[] compilers=[1173803:cargo]
+rocm_gpu_guard: 2026-10-08T22:23:33+0900 idle streak reset after 1s: gpu=[] compilers=[1334471:cargo]
+rocm_gpu_guard: 2026-10-08T23:05:01+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-08T23:05:01+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.252 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 1.80 GB, active+cache 2.18 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          432.66 ms (4733.48 tok/s)
+  Decode:           594.48 ms (215.31 tok/s)
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-08T23:05:05+0900 monitor: 3 samples
+rocm_gpu_guard: 2026-10-08T23:05:05+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_new.log
new file mode 100644
index 000000000..f449f5bf9
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r1_new.log
@@ -0,0 +1,26 @@
+rocm_gpu_guard: 2026-10-08T23:05:15+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-08T23:13:56+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 521s
+rocm_gpu_guard: 2026-10-08T23:13:56+0900 waiting for 60s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-08T23:14:31+0900 idle streak reset after 23s: gpu=[3262038:rocminfo] compilers=[3262000:cargo]
+rocm_gpu_guard: 2026-10-08T23:23:02+0900 idle streak reset after 14s: gpu=[] compilers=[3571564:cargo,3571605:rustc,3571606:rustc,3571607:rustc,3571608:rustc,3571609:rustc]
+rocm_gpu_guard: 2026-10-08T23:28:13+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-08T23:28:13+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.250 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          474.47 ms (4316.35 tok/s)
+  Decode:           580.45 ms (220.52 tok/s)
+  Decode path:      engine
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-08T23:28:17+0900 monitor: 3 samples
+rocm_gpu_guard: 2026-10-08T23:28:17+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base-null.log
new file mode 100644
index 000000000..f3165af6c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:05:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:05:19+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:05:34+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:05:34+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.195 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 1.81 GB, active+cache 2.19 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          460.43 ms (4447.97 tok/s)
+  Decode:           586.82 ms (218.13 tok/s)
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-09T02:05:38+0900 monitor: 3 samples
+rocm_gpu_guard: 2026-10-09T02:05:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base.log
new file mode 100644
index 000000000..34c9ffe01
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:05:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:05:43+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:05:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:05:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.199 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 1.81 GB, active+cache 2.19 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 1.00 GB, active+cache 1.39 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          453.61 ms (4514.86 tok/s)
+  Decode:           574.66 ms (222.74 tok/s)
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-09T02:06:02+0900 monitor: 3 samples
+rocm_gpu_guard: 2026-10-09T02:06:02+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_new.log
new file mode 100644
index 000000000..c9068650d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:04:57+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:04:57+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:05:11+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:05:11+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.193 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.99 GB, active+cache 1.37 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          450.84 ms (4542.64 tok/s)
+  Decode:           590.98 ms (216.59 tok/s)
+  Decode path:      engine
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-09T02:05:14+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:05:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base-null.log
new file mode 100644
index 000000000..aff771ccf
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:24:51+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:24:51+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:25:05+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:25:05+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.200 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 1.82 GB, active+cache 2.20 GB, phase peak 1.86 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 1.00 GB, active+cache 1.39 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          452.00 ms (4530.98 tok/s)
+  Decode:           588.40 ms (217.54 tok/s)
+  MLX peak memory:  1.86 GB
+rocm_gpu_guard: 2026-10-09T02:25:08+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:25:08+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base.log
new file mode 100644
index 000000000..ba360683d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:25:13+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:25:13+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:25:28+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:25:28+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.205 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 1.81 GB, active+cache 2.19 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 1.00 GB, active+cache 1.39 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          462.41 ms (4428.99 tok/s)
+  Decode:           586.97 ms (218.07 tok/s)
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-09T02:25:32+0900 monitor: 3 samples
+rocm_gpu_guard: 2026-10-09T02:25:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_new.log
new file mode 100644
index 000000000..4e6ad39a4
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp2048_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:25:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:25:37+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:25:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:25:52+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.198 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.88 GB, active+cache 1.27 GB, phase peak 1.84 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.96 GB, active+cache 1.35 GB, phase peak 1.85 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          454.49 ms (4506.18 tok/s)
+  Decode:           591.13 ms (216.53 tok/s)
+  Decode path:      engine
+  MLX peak memory:  1.85 GB
+rocm_gpu_guard: 2026-10-09T02:25:55+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:25:55+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..d4f5a5e65
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-08T21:22:36+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-08T21:32:09+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 573s
+rocm_gpu_guard: 2026-10-08T21:32:09+0900 waiting for 60s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-08T21:33:38+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-08T21:33:38+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.192 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.77 GB, active+cache 1.15 GB, phase peak 1.04 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.31 GB, active+cache 0.69 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          70.41 ms (7271.58 tok/s)
+  Decode:           462.36 ms (276.84 tok/s)
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-08T21:33:41+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-08T21:33:41+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..35457a283
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_base.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-08T19:17:44+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-08T19:41:27+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 1423s
+rocm_gpu_guard: 2026-10-08T19:41:27+0900 waiting for 60s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-08T19:42:09+0900 idle streak reset after 28s: gpu=[] compilers=[3313432:cargo]
+rocm_gpu_guard: 2026-10-08T20:39:08+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-08T20:39:08+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.741 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.36 GB, active+cache 0.74 GB, phase peak 0.63 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.26 GB, active+cache 0.65 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          73.12 ms (7002.12 tok/s)
+  Decode:           465.33 ms (275.07 tok/s)
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-08T20:39:15+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-08T20:39:15+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..7112e48e7
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r1_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-08T20:39:25+0900 waiting for guard lock /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-08T21:20:56+0900 acquired guard lock /tmp/mlxcel-rocm-gpu-guard.lock after 2491s
+rocm_gpu_guard: 2026-10-08T21:20:56+0900 waiting for 60s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-08T21:22:23+0900 idle for 60s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-08T21:22:23+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.256 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.26 GB, active+cache 0.64 GB, phase peak 0.60 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.26 GB, active+cache 0.65 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          59.61 ms (8589.40 tok/s)
+  Decode:           468.02 ms (273.49 tok/s)
+  Decode path:      engine
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-08T21:22:26+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-08T21:22:26+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..a8b18a976
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:04:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:04:12+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:04:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:04:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.197 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.67 GB, active+cache 1.06 GB, phase peak 0.95 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.31 GB, active+cache 0.69 GB, phase peak 1.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          61.93 ms (8267.07 tok/s)
+  Decode:           464.92 ms (275.31 tok/s)
+  MLX peak memory:  1.47 GB
+rocm_gpu_guard: 2026-10-09T02:04:29+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:04:29+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..85f974b44
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:04:34+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:04:34+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:04:49+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:04:49+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.192 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.61 GB, active+cache 1.00 GB, phase peak 0.87 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.31 GB, active+cache 0.69 GB, phase peak 1.47 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          68.65 ms (7458.21 tok/s)
+  Decode:           463.13 ms (276.38 tok/s)
+  MLX peak memory:  1.47 GB
+rocm_gpu_guard: 2026-10-09T02:04:52+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:04:52+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..d92cd768f
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:03:49+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:03:49+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:04:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:04:04+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.211 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.26 GB, active+cache 0.64 GB, phase peak 0.60 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.26 GB, active+cache 0.65 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          59.82 ms (8559.12 tok/s)
+  Decode:           467.00 ms (274.09 tok/s)
+  Decode path:      engine
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-09T02:04:06+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:04:06+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..504582ada
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:23:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:23:43+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:23:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:23:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.195 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.56 GB, active+cache 0.95 GB, phase peak 0.83 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.31 GB, active+cache 0.69 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          65.70 ms (7793.54 tok/s)
+  Decode:           462.61 ms (276.69 tok/s)
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-09T02:24:01+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:24:01+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..ae61fa0ba
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:24:06+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:24:06+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:24:20+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:24:20+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.198 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.55 GB, active+cache 0.93 GB, phase peak 0.82 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.31 GB, active+cache 0.69 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          68.96 ms (7424.31 tok/s)
+  Decode:           465.01 ms (275.26 tok/s)
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-09T02:24:23+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:24:23+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..3639461b2
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-0.6B-4bit_pp512_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:24:28+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:24:28+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:24:43+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:24:43+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-0.6B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.194 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 0.39 GB, cache 0.30 GB, active+cache 0.69 GB, phase peak 0.83 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 0.39 GB, cache 0.31 GB, active+cache 0.69 GB, phase peak 1.46 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          68.81 ms (7440.48 tok/s)
+  Decode:           465.67 ms (274.88 tok/s)
+  Decode path:      engine
+  MLX peak memory:  1.46 GB
+rocm_gpu_guard: 2026-10-09T02:24:46+0900 monitor: 2 samples
+rocm_gpu_guard: 2026-10-09T02:24:46+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log
new file mode 100644
index 000000000..7c2287d83
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:53:41+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:53:41+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:53:56+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:53:56+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.212 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.91 GB, active+cache 19.17 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.37 GB, active+cache 18.64 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7150.03 ms (286.43 tok/s)
+  Decode:           2200.86 ms (58.16 tok/s)
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T01:54:21+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T01:54:21+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base.log
new file mode 100644
index 000000000..c90caa743
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:52:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:52:12+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:52:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:52:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.210 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.91 GB, active+cache 19.17 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.54 GB, active+cache 18.81 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7161.06 ms (285.99 tok/s)
+  Decode:           2205.06 ms (58.05 tok/s)
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T01:52:52+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T01:52:52+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_new.log
new file mode 100644
index 000000000..5130d3887
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r1_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:52:57+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:52:57+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:53:11+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:53:11+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.207 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.37 GB, active+cache 18.64 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.54 GB, active+cache 18.81 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7249.95 ms (282.48 tok/s)
+  Decode:           2196.16 ms (58.28 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T01:53:36+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T01:53:36+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log
new file mode 100644
index 000000000..e00939b08
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:12:50+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:12:50+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:13:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:13:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.217 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.90 GB, active+cache 19.17 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.54 GB, active+cache 18.81 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7272.66 ms (281.60 tok/s)
+  Decode:           2200.73 ms (58.16 tok/s)
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T02:13:28+0900 monitor: 16 samples
+rocm_gpu_guard: 2026-10-09T02:13:28+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base.log
new file mode 100644
index 000000000..7fc69a5b3
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:13:33+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:13:33+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:13:48+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:13:48+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.251 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.90 GB, active+cache 19.17 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.54 GB, active+cache 18.81 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7144.89 ms (286.64 tok/s)
+  Decode:           2202.88 ms (58.11 tok/s)
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T02:14:13+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T02:14:13+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_new.log
new file mode 100644
index 000000000..7cf1a0ea4
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:11:54+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:11:54+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:12:09+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:12:09+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.221 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.37 GB, active+cache 18.64 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.37 GB, active+cache 18.64 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7053.70 ms (290.34 tok/s)
+  Decode:           2201.71 ms (58.14 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T02:12:44+0900 monitor: 24 samples
+rocm_gpu_guard: 2026-10-09T02:12:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log
new file mode 100644
index 000000000..d8f0ae959
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:31:48+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:31:48+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:32:03+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:32:03+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.217 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.90 GB, active+cache 19.17 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.54 GB, active+cache 18.81 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7198.39 ms (284.51 tok/s)
+  Decode:           2208.97 ms (57.95 tok/s)
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T02:32:28+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T02:32:28+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base.log
new file mode 100644
index 000000000..1822eec3f
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:32:33+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:32:33+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:32:47+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:32:47+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.215 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.90 GB, active+cache 19.17 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.54 GB, active+cache 18.81 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7080.17 ms (289.26 tok/s)
+  Decode:           2192.55 ms (58.38 tok/s)
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T02:33:13+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T02:33:13+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_new.log
new file mode 100644
index 000000000..e7c91f7b0
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp2048_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:33:18+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:33:18+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:33:32+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:33:32+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.215 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 1.37 GB, active+cache 18.63 GB, phase peak 18.80 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 1.37 GB, active+cache 18.64 GB, phase peak 18.91 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          7148.04 ms (286.51 tok/s)
+  Decode:           2204.57 ms (58.06 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.91 GB
+rocm_gpu_guard: 2026-10-09T02:33:57+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T02:33:57+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..a7916b02e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:51:37+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:51:37+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:51:52+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:51:52+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.209 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.65 GB, active+cache 17.92 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.77 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1806.26 ms (283.46 tok/s)
+  Decode:           2042.81 ms (62.66 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T01:52:07+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T01:52:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..20dccea75
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:50:11+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:50:11+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:50:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:50:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.216 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.65 GB, active+cache 17.92 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.77 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          2292.18 ms (223.37 tok/s)
+  Decode:           2046.99 ms (62.53 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T01:50:58+0900 monitor: 21 samples
+rocm_gpu_guard: 2026-10-09T01:50:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..87bc3e5b4
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r1_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:51:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:51:03+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:51:17+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:51:17+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.216 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.38 GB, active+cache 17.65 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.77 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1889.81 ms (270.93 tok/s)
+  Decode:           2047.51 ms (62.52 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T01:51:32+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T01:51:32+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..0f456150f
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:10:44+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:10:44+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:10:59+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:10:59+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.208 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.65 GB, active+cache 17.92 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.77 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1605.19 ms (318.97 tok/s)
+  Decode:           2048.39 ms (62.49 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:11:14+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:11:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..634199821
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:11:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:11:19+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:11:34+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:11:34+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.213 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.65 GB, active+cache 17.92 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.76 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1736.22 ms (294.89 tok/s)
+  Decode:           2055.99 ms (62.26 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:11:49+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:11:49+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..484b481ff
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:10:10+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:10:10+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:10:25+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:10:25+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.220 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.38 GB, active+cache 17.65 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.39 GB, active+cache 17.66 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1964.26 ms (260.66 tok/s)
+  Decode:           2050.47 ms (62.42 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:10:39+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:10:39+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..47da1dbfb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:30:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:30:03+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:30:18+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:30:18+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.213 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.65 GB, active+cache 17.92 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.76 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1982.61 ms (258.25 tok/s)
+  Decode:           2039.94 ms (62.75 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:30:34+0900 monitor: 11 samples
+rocm_gpu_guard: 2026-10-09T02:30:34+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..f63e12679
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:30:39+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:30:39+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:30:54+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:30:54+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.220 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.65 GB, active+cache 17.92 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.76 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1807.65 ms (283.24 tok/s)
+  Decode:           2039.27 ms (62.77 tok/s)
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:31:09+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:31:09+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..f268c4011
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/Qwen3-30B-A3B-4bit_pp512_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:31:14+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:31:14+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:31:28+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:31:28+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/Qwen3-30B-A3B-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.223 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 17.27 GB, cache 0.38 GB, active+cache 17.65 GB, phase peak 17.66 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(40960), reserved 128 for generation)
+[Memory] after measured pass: active 17.27 GB, cache 0.50 GB, active+cache 17.76 GB, phase peak 18.66 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1903.10 ms (269.03 tok/s)
+  Decode:           2041.74 ms (62.69 tok/s)
+  Decode path:      engine
+  MLX peak memory:  18.66 GB
+rocm_gpu_guard: 2026-10-09T02:31:43+0900 monitor: 10 samples
+rocm_gpu_guard: 2026-10-09T02:31:43+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base-null.log
new file mode 100644
index 000000000..90c19449e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:49:44+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:49:44+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:49:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:49:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.630 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 2.47 GB, active+cache 5.28 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.97 GB, cache 1.04 GB, active+cache 4.02 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          702.59 ms (2914.94 tok/s)
+  Decode:           2046.94 ms (62.53 tok/s)
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T01:50:06+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T01:50:06+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base.log
new file mode 100644
index 000000000..06a2380fd
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:48:50+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:48:50+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:49:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:49:04+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.633 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 2.47 GB, active+cache 5.27 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.97 GB, cache 1.21 GB, active+cache 4.18 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          720.27 ms (2843.38 tok/s)
+  Decode:           2037.11 ms (62.83 tok/s)
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T01:49:12+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T01:49:12+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_new.log
new file mode 100644
index 000000000..813cdab26
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r1_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T01:49:17+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:49:17+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:49:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:49:31+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.639 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 1.24 GB, active+cache 4.04 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.80 GB, cache 1.33 GB, active+cache 4.14 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          702.60 ms (2914.89 tok/s)
+  Decode:           2061.91 ms (62.08 tok/s)
+  Decode path:      engine
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T01:49:39+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T01:49:39+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base-null.log
new file mode 100644
index 000000000..310273323
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:09:16+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:09:16+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:09:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:09:31+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.668 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 2.47 GB, active+cache 5.28 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.97 GB, cache 1.21 GB, active+cache 4.18 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          703.54 ms (2911.00 tok/s)
+  Decode:           2046.37 ms (62.55 tok/s)
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T02:09:38+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:09:38+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base.log
new file mode 100644
index 000000000..a1da50093
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:09:43+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:09:43+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:09:58+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:09:58+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.662 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 2.47 GB, active+cache 5.28 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.97 GB, cache 1.04 GB, active+cache 4.02 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          706.78 ms (2897.64 tok/s)
+  Decode:           2045.58 ms (62.57 tok/s)
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T02:10:05+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:10:05+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_new.log
new file mode 100644
index 000000000..d77e36575
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r2_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T02:08:49+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:08:49+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:09:04+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:09:04+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.649 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 1.24 GB, active+cache 4.04 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.80 GB, cache 1.42 GB, active+cache 4.22 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          709.08 ms (2888.25 tok/s)
+  Decode:           2074.55 ms (61.70 tok/s)
+  Decode path:      engine
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T02:09:11+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:09:11+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base-null.log
new file mode 100644
index 000000000..76c7e1a9e
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:28:42+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:28:42+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:28:57+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:28:57+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.648 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 2.47 GB, active+cache 5.27 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.97 GB, cache 1.21 GB, active+cache 4.18 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          703.00 ms (2913.23 tok/s)
+  Decode:           2043.34 ms (62.64 tok/s)
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T02:29:04+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:29:04+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base.log
new file mode 100644
index 000000000..9fd239a73
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:29:09+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:29:09+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:29:24+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:29:24+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.648 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 2.47 GB, active+cache 5.28 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.97 GB, cache 1.21 GB, active+cache 4.18 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          706.11 ms (2900.41 tok/s)
+  Decode:           2042.94 ms (62.65 tok/s)
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T02:29:31+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:29:31+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_new.log
new file mode 100644
index 000000000..5bed9a721
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp2048_r3_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T02:29:36+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:29:36+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:29:51+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:29:51+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 2048
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.641 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 1.24 GB, active+cache 4.04 GB, phase peak 4.27 GB
+[long-prompt] target=2048 tokens -> using 2048 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.80 GB, cache 1.25 GB, active+cache 4.05 GB, phase peak 4.52 GB
+[Profile Results]
+  Prompt tokens:    2048
+  Generated tokens: 128
+  Prefill:          701.77 ms (2918.35 tok/s)
+  Decode:           2060.93 ms (62.11 tok/s)
+  Decode path:      engine
+  MLX peak memory:  4.52 GB
+rocm_gpu_guard: 2026-10-09T02:29:58+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:29:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..69eba7246
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:48:24+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:48:24+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:48:39+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:48:39+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.623 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.53 GB, active+cache 3.33 GB, phase peak 3.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.91 GB, cache 0.25 GB, active+cache 3.16 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          228.61 ms (2239.65 tok/s)
+  Decode:           1856.04 ms (68.96 tok/s)
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T01:48:45+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T01:48:45+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..5002b4b18
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:47:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:47:31+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:47:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:47:46+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.666 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.52 GB, active+cache 3.33 GB, phase peak 3.09 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.91 GB, cache 0.24 GB, active+cache 3.15 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          480.68 ms (1065.16 tok/s)
+  Decode:           1853.53 ms (69.06 tok/s)
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T01:47:53+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T01:47:53+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..b5b637887
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r1_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T01:47:58+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:47:58+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:48:13+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:48:13+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.660 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.38 GB, active+cache 3.18 GB, phase peak 3.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.80 GB, cache 0.37 GB, active+cache 3.18 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          231.93 ms (2207.57 tok/s)
+  Decode:           1852.54 ms (69.09 tok/s)
+  Decode path:      engine
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T01:48:19+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T01:48:19+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..76a3c6c70
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:07:58+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:07:58+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:08:12+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:08:12+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.661 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.53 GB, active+cache 3.33 GB, phase peak 3.09 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.91 GB, cache 0.28 GB, active+cache 3.19 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          236.92 ms (2161.07 tok/s)
+  Decode:           1852.86 ms (69.08 tok/s)
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T02:08:18+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:08:18+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..390c92703
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:08:23+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:08:23+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:08:38+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:08:38+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.662 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.53 GB, active+cache 3.33 GB, phase peak 3.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.91 GB, cache 0.29 GB, active+cache 3.20 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          237.18 ms (2158.68 tok/s)
+  Decode:           1847.37 ms (69.29 tok/s)
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T02:08:44+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:08:44+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..43ebfe808
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r2_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T02:07:32+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:07:32+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:07:47+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:07:47+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.658 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.34 GB, active+cache 3.14 GB, phase peak 3.09 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.80 GB, cache 0.39 GB, active+cache 3.20 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          234.46 ms (2183.76 tok/s)
+  Decode:           1857.67 ms (68.90 tok/s)
+  Decode path:      engine
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T02:07:53+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:07:53+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..6fb346d2a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:27:25+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:27:25+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:27:40+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:27:40+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.671 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.53 GB, active+cache 3.33 GB, phase peak 3.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.91 GB, cache 0.25 GB, active+cache 3.16 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          229.13 ms (2234.52 tok/s)
+  Decode:           1834.46 ms (69.78 tok/s)
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T02:27:46+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:27:46+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..29eb6b311
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:27:51+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:27:51+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:28:05+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:28:05+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.648 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.53 GB, active+cache 3.33 GB, phase peak 3.10 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.91 GB, cache 0.29 GB, active+cache 3.20 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          237.50 ms (2155.78 tok/s)
+  Decode:           1849.93 ms (69.19 tok/s)
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T02:28:11+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:28:11+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..6e00f6967
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gemma-3-4b-it-4bit_pp512_r3_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T02:28:16+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:28:16+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:28:31+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:28:31+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gemma-3-4b-it-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Warning: model.safetensors.index.json in models/mlx/gemma-3-4b-it-4bit references shards that don't match the on-disk files (likely a repackaged mlx-community quant). Falling back to all *.safetensors files in the directory.
+[Load] wall: 0.638 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 2.80 GB, cache 0.38 GB, active+cache 3.18 GB, phase peak 3.09 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=None, reserved 128 for generation)
+[Memory] after measured pass: active 2.80 GB, cache 0.39 GB, active+cache 3.20 GB, phase peak 3.92 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          229.49 ms (2231.04 tok/s)
+  Decode:           1865.00 ms (68.63 tok/s)
+  Decode path:      engine
+  MLX peak memory:  3.92 GB
+rocm_gpu_guard: 2026-10-09T02:28:37+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:28:37+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log
new file mode 100644
index 000000000..3c51ef697
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:59:12+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:59:12+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:59:26+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:59:26+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.489 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.26 GB, cache 1.90 GB, active+cache 13.16 GB, phase peak 13.67 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.84 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          968.89 ms (528.44 tok/s)
+  Decode:           15015.86 ms (8.52 tok/s)
+  MLX peak memory:  13.84 GB
+rocm_gpu_guard: 2026-10-09T01:59:48+0900 monitor: 14 samples
+rocm_gpu_guard: 2026-10-09T01:59:48+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log
new file mode 100644
index 000000000..56373792d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:57:31+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:57:31+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:57:46+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:57:46+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.570 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.26 GB, cache 1.82 GB, active+cache 13.08 GB, phase peak 12.88 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.84 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          969.43 ms (528.15 tok/s)
+  Decode:           14985.87 ms (8.54 tok/s)
+  MLX peak memory:  13.84 GB
+rocm_gpu_guard: 2026-10-09T01:58:22+0900 monitor: 24 samples
+rocm_gpu_guard: 2026-10-09T01:58:22+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log
new file mode 100644
index 000000000..1d75c1fda
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r1_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T01:58:27+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:58:27+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:58:42+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:58:42+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.566 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.67 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.30 GB, cache 0.28 GB, active+cache 11.57 GB, phase peak 13.60 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1004.76 ms (509.57 tok/s)
+  Decode:           15016.75 ms (8.52 tok/s)
+  Decode path:      engine
+  MLX peak memory:  13.67 GB
+rocm_gpu_guard: 2026-10-09T01:59:07+0900 monitor: 16 samples
+rocm_gpu_guard: 2026-10-09T01:59:07+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log
new file mode 100644
index 000000000..aba3ec3cd
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:18:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:18:19+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:18:33+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:18:33+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.555 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.26 GB, cache 1.90 GB, active+cache 13.16 GB, phase peak 13.67 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.84 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          973.88 ms (525.73 tok/s)
+  Decode:           15085.56 ms (8.48 tok/s)
+  MLX peak memory:  13.84 GB
+rocm_gpu_guard: 2026-10-09T02:18:58+0900 monitor: 16 samples
+rocm_gpu_guard: 2026-10-09T02:18:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log
new file mode 100644
index 000000000..111adbc8a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:19:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:19:03+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:19:17+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:19:17+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.502 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.26 GB, cache 1.90 GB, active+cache 13.16 GB, phase peak 13.67 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.84 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          971.68 ms (526.92 tok/s)
+  Decode:           15123.50 ms (8.46 tok/s)
+  MLX peak memory:  13.84 GB
+rocm_gpu_guard: 2026-10-09T02:19:40+0900 monitor: 15 samples
+rocm_gpu_guard: 2026-10-09T02:19:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log
new file mode 100644
index 000000000..c3d0c9d91
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r2_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T02:17:21+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:17:21+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:17:36+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:17:36+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.579 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 12.88 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.30 GB, cache 0.28 GB, active+cache 11.57 GB, phase peak 13.60 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1007.14 ms (508.37 tok/s)
+  Decode:           15144.73 ms (8.45 tok/s)
+  Decode path:      engine
+  MLX peak memory:  13.60 GB
+rocm_gpu_guard: 2026-10-09T02:18:14+0900 monitor: 25 samples
+rocm_gpu_guard: 2026-10-09T02:18:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log
new file mode 100644
index 000000000..4e6a5e281
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base-null.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:37:06+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:37:06+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:37:21+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:37:21+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.615 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.26 GB, cache 1.83 GB, active+cache 13.09 GB, phase peak 12.88 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.84 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          971.76 ms (526.88 tok/s)
+  Decode:           15159.36 ms (8.44 tok/s)
+  MLX peak memory:  13.84 GB
+rocm_gpu_guard: 2026-10-09T02:37:58+0900 monitor: 25 samples
+rocm_gpu_guard: 2026-10-09T02:37:58+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log
new file mode 100644
index 000000000..315b97ec5
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_base.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:38:03+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:38:03+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:38:18+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:38:18+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.546 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.26 GB, cache 1.90 GB, active+cache 13.16 GB, phase peak 13.67 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.84 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1097.23 ms (466.63 tok/s)
+  Decode:           15250.46 ms (8.39 tok/s)
+  MLX peak memory:  13.84 GB
+rocm_gpu_guard: 2026-10-09T02:38:43+0900 monitor: 17 samples
+rocm_gpu_guard: 2026-10-09T02:38:43+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log
new file mode 100644
index 000000000..ad6ea9f63
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/gpt-oss-20b-MXFP4-Q4_pp512_r3_new.log
@@ -0,0 +1,24 @@
+rocm_gpu_guard: 2026-10-09T02:38:48+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:38:48+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:39:03+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:39:03+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/gpt-oss-20b-MXFP4-Q4 -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+Quantization mode mxfp4: running on native ROCm kernels, no load-time conversion (backend quantization capability table)
+[Load] wall: 0.499 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 11.28 GB, cache 0.27 GB, active+cache 11.55 GB, phase peak 13.67 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 11.30 GB, cache 0.28 GB, active+cache 11.57 GB, phase peak 13.60 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1002.98 ms (510.48 tok/s)
+  Decode:           15075.33 ms (8.49 tok/s)
+  Decode path:      engine
+  MLX peak memory:  13.67 GB
+rocm_gpu_guard: 2026-10-09T02:39:25+0900 monitor: 15 samples
+rocm_gpu_guard: 2026-10-09T02:39:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log
new file mode 100644
index 000000000..c46dd80b3
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:55:19+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:55:19+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:55:34+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:55:34+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.118 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 3.76 GB, active+cache 7.71 GB, phase peak 7.22 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.02 GB, cache 1.78 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          865.72 ms (591.41 tok/s)
+  Decode:           1435.18 ms (89.19 tok/s)
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T01:55:40+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T01:55:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base.log
new file mode 100644
index 000000000..726c5040c
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T01:54:26+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:54:26+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:54:40+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:54:40+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.137 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 1.85 GB, active+cache 5.81 GB, phase peak 6.61 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.02 GB, cache 1.78 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1141.73 ms (448.44 tok/s)
+  Decode:           1426.12 ms (89.75 tok/s)
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T01:54:49+0900 monitor: 6 samples
+rocm_gpu_guard: 2026-10-09T01:54:49+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_new.log
new file mode 100644
index 000000000..059b91b1d
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r1_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T01:54:54+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T01:54:54+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T01:55:08+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T01:55:08+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.122 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 1.85 GB, active+cache 5.81 GB, phase peak 7.22 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 3.96 GB, cache 1.85 GB, active+cache 5.81 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          869.89 ms (588.58 tok/s)
+  Decode:           1444.81 ms (88.59 tok/s)
+  Decode path:      engine
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T01:55:14+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T01:55:14+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log
new file mode 100644
index 000000000..4b892dd4a
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:14:45+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:14:45+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:15:00+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:15:00+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.126 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 3.76 GB, active+cache 7.71 GB, phase peak 7.22 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.02 GB, cache 1.78 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          894.81 ms (572.19 tok/s)
+  Decode:           1429.58 ms (89.54 tok/s)
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T02:15:06+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:15:06+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base.log
new file mode 100644
index 000000000..3fcd771ef
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:15:11+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:15:11+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:15:25+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:15:25+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.121 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 3.76 GB, active+cache 7.71 GB, phase peak 7.22 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.02 GB, cache 1.78 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          894.08 ms (572.66 tok/s)
+  Decode:           1429.87 ms (89.52 tok/s)
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T02:15:31+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:15:31+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_new.log
new file mode 100644
index 000000000..7f9a607db
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r2_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:14:18+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:14:18+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:14:33+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:14:33+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.131 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 1.85 GB, active+cache 5.81 GB, phase peak 6.61 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 3.96 GB, cache 1.84 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1128.64 ms (453.64 tok/s)
+  Decode:           1438.29 ms (88.99 tok/s)
+  Decode path:      engine
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T02:14:40+0900 monitor: 5 samples
+rocm_gpu_guard: 2026-10-09T02:14:40+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log
new file mode 100644
index 000000000..d82645bb6
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base-null.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:34:02+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:34:02+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:34:17+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:34:17+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.131 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 1.85 GB, active+cache 5.81 GB, phase peak 6.61 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.02 GB, cache 1.78 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          1331.37 ms (384.57 tok/s)
+  Decode:           1427.41 ms (89.67 tok/s)
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T02:34:25+0900 monitor: 6 samples
+rocm_gpu_guard: 2026-10-09T02:34:25+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base.log
new file mode 100644
index 000000000..c855ae2bb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_base.log
@@ -0,0 +1,22 @@
+rocm_gpu_guard: 2026-10-09T02:34:30+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:34:30+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:34:45+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:34:45+0900 attempt 1/10: start: /tmp/claude-1001/-home-devops-Development-mlxcel/66a03ced-36c6-4e0a-9e39-155a3eb57353/scratchpad/issue-2192/pre-epic/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.123 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 3.20 GB, active+cache 7.16 GB, phase peak 7.22 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 4.02 GB, cache 1.78 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          868.14 ms (589.77 tok/s)
+  Decode:           1431.48 ms (89.42 tok/s)
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T02:34:51+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:34:51+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_new.log b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_new.log
new file mode 100644
index 000000000..ab486a6fb
--- /dev/null
+++ b/docs/benchmark_results/data/rocm-unified-engine-gfx1151-2026-10-08/rounds/raw/granite-4.0-h-tiny-4bit_pp512_r3_new.log
@@ -0,0 +1,23 @@
+rocm_gpu_guard: 2026-10-09T02:34:56+0900 lock held by outer guard 2263193: not taking /tmp/mlxcel-rocm-gpu-guard.lock
+rocm_gpu_guard: 2026-10-09T02:34:56+0900 waiting for 10s of idle GPU and no compiler
+rocm_gpu_guard: 2026-10-09T02:35:10+0900 idle for 10s: kfd proc empty and no compiler in every 1 Hz sample
+rocm_gpu_guard: 2026-10-09T02:35:10+0900 attempt 1/10: start: /home/devops/Development/wt-20261006T1212-issue-2192/target/release/mlxcel-bench-decode -m models/mlx/granite-4.0-h-tiny-4bit -p x -n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens 512
+[mlx-rocm] bound HIP device 0: gfx1151 (AMD Radeon 8060S Graphics) cus=20 warp=32 lds=64KB
+[Memory] cache limit: 2.15 GB
+[Load] wall: 0.121 s  MLX peak: 0.00 GB
+[Memory] after load: active 0.00 GB, cache 0.00 GB, active+cache 0.00 GB, phase peak 0.00 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[mlx-rocm] gfx1151 native_wmma=1 cus=20 warp=32 low_cu=0 host_WARP_SIZE=32
+[hipBLASLt caps] device 0 (gfx1151): bf16=1 fp8_e4m3=0 fp8_e5m2=0 int8=1
+[Memory] after warmup (prefill + warmup tokens): active 3.96 GB, cache 1.85 GB, active+cache 5.81 GB, phase peak 7.22 GB
+[long-prompt] target=512 tokens -> using 512 tokens (max_context=Some(131072), reserved 128 for generation)
+[Memory] after measured pass: active 3.96 GB, cache 1.84 GB, active+cache 5.80 GB, phase peak 7.31 GB
+[Profile Results]
+  Prompt tokens:    512
+  Generated tokens: 128
+  Prefill:          855.19 ms (598.70 tok/s)
+  Decode:           1428.92 ms (89.58 tok/s)
+  Decode path:      engine
+  MLX peak memory:  7.31 GB
+rocm_gpu_guard: 2026-10-09T02:35:16+0900 monitor: 4 samples
+rocm_gpu_guard: 2026-10-09T02:35:16+0900 attempt 1: CLEAN, exit 0
diff --git a/docs/benchmark_results/rocm-unified-engine-gfx1151-2026-10-08.md b/docs/benchmark_results/rocm-unified-engine-gfx1151-2026-10-08.md
index f54f96dfa..6ea0d56d1 100644
--- a/docs/benchmark_results/rocm-unified-engine-gfx1151-2026-10-08.md
+++ b/docs/benchmark_results/rocm-unified-engine-gfx1151-2026-10-08.md
@@ -126,7 +126,7 @@ ADR 0007's ROCm paragraph: the CLI and the server engine with dense storage must
 
 ### Method
 
-`data/.../harness/bench_rounds.sh`: for each (model, prompt length) cell, three interleaved rounds of three arms, each a fresh `mlxcel-bench-decode` process at `scripts/bench_decode.sh`'s shape (`-n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens N`, greedy): `base` is the pre-epic binary (`4c44e317`, `CxxGenerator`), `new` is `main` (`DirectEngine`, `Decode path: engine`), and `base-null` is the pre-epic binary again, the method's noise floor. The arm order rotates each round. Every run was its own `scripts/rocm_gpu_guard.sh` command inside one host-wide lock hold (`data/.../harness/full_session.sh`): 10 s of idle GPU and no compiler before each run, a 1 Hz monitor during it, and all 99 runs were CLEAN on the first attempt (`data/.../guard/guard-logs.tar.gz`). The session ran 2026-10-09 01:19 to 02:43 KST. `data/.../harness/summarize_rounds.py` computes the medians and the paired per-round deltas below (median and range over the three rounds, percent of `base`). The threshold is ADR 0007's 1.0 percent on the paired decode delta.
+`data/.../harness/bench_rounds.sh`: for each (model, prompt length) cell, three interleaved rounds of three arms, each a fresh `mlxcel-bench-decode` process at `scripts/bench_decode.sh`'s shape (`-n 128 --warmup-tokens 20 --ignore-eos --prompt-tokens N`, greedy): `base` is the pre-epic binary (`4c44e317`, `CxxGenerator`), `new` is `main` (`DirectEngine`, `Decode path: engine`), and `base-null` is the pre-epic binary again, the method's noise floor. The arm order rotates each round. Every run was its own `scripts/rocm_gpu_guard.sh` command inside one host-wide lock hold (`data/.../harness/full_session.sh`): 10 s of idle GPU and no compiler before each run, a 1 Hz monitor during it, and all 99 runs were CLEAN on the first attempt (`data/.../guard/logs/`). The session ran 2026-10-09 01:19 to 02:43 KST. `data/.../harness/summarize_rounds.py` computes the medians and the paired per-round deltas below (median and range over the three rounds, percent of `base`). The threshold is ADR 0007's 1.0 percent on the paired decode delta.
 
 ### Decode, `mlxcel generate` (DirectEngine) vs `CxxGenerator`