From 5f9b9920c0c1e497b55c150d426eaf74891e6e91 Mon Sep 17 00:00:00 2001 From: "Yang, JiaRong" Date: Wed, 9 Sep 2026 05:25:48 +0000 Subject: [PATCH 1/2] [AMD][MI35X] Bump glm5.2-fp4-mi355x-sglang-agentic-mtp image to v0.5.19-rocm10-20260908 Bump image from lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728 to lmsysorg/sglang-rocm:v0.5.19-rocm10-mi35x-20260908. Recipe unchanged. Co-authored-by: Cursor --- configs/amd-master.yaml | 2 +- perf-changelog.yaml | 8 ++++++++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 0940c2495a..93f0a8e4be 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -1709,7 +1709,7 @@ minimaxm3-fp4-mi355x-vllm-agentic-mtp: # c6 and c8 removed after sweep validation: dominated by TP4/EP4 arm. # SA selects the Pareto-optimal arm per concurrency point. glm5.2-fp4-mi355x-sglang-agentic-mtp: - image: lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728 + image: lmsysorg/sglang-rocm:v0.5.19-rocm10-mi35x-20260908 model: amd/GLM-5.2-MXFP4 model-prefix: glm5.2 runner: cluster:mi355x-amds diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 9dfeb4a02a..eae837f436 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6984,3 +6984,11 @@ - "Export OMP_NUM_THREADS=1 from all three scripts. vLLM clamps torch threads itself, but only after the weights are loaded: the 'Reducing Torch threads from 16 to 1 for serving' line is emitted after 'Model loading took', so the checkpoint is read with 16 torch threads per worker across 8 workers. Setting it externally caps threads from process start, which is what that log line recommends." - "Weight loading currently costs 5760-5946 s (96-99 min) per job on this runner, measured across the c1, c4, c56 and c64 AgentX jobs of run 34087124574 - roughly half of each ~3 h job. It does not affect any measured number: loading completes before warmup and the profiling phase. The effect is also only observable on this runner. A local EXT4 box serves the checkpoint from page cache and loads in ~40 s with or without the cap, so it never exercises the cold-read path. CI reads cold from WEKAFS, where each of the 8 ranks streams the full 805 GiB to slice its own shards - roughly 6.4 TB of aggregate network reads per job - and that is where oversubscribed reader threads cost real time." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2851 + +- config-keys: + - glm5.2-fp4-mi355x-sglang-agentic-mtp + scenario-type: + - agentic-coding + description: + - "Bump image from lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728 to lmsysorg/sglang-rocm:v0.5.19-rocm10-mi35x-20260908" + pr-link: XXX From 3b5875a043090343dc55df1d052be3f3ee4ec1a2 Mon Sep 17 00:00:00 2001 From: "Yang, JiaRong" Date: Wed, 9 Sep 2026 10:26:27 +0000 Subject: [PATCH 2/2] chore: set canonical pr-link for #2917 Co-authored-by: Cursor --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index eae837f436..6890f17240 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6991,4 +6991,4 @@ - agentic-coding description: - "Bump image from lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728 to lmsysorg/sglang-rocm:v0.5.19-rocm10-mi35x-20260908" - pr-link: XXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2917