From d84aefcdc17204192b2e0c160d1a38bc5a95e868 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 12:44:54 +0000 Subject: [PATCH 1/2] config(amd): bump dsv41flash-fp4-mi325x-vllm-agentic-dspark to vLLM ROCm nightly dee37d89 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Update the MI325X DeepSeek-V4.1-Flash vLLM DSpark AgentX image from vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657 (2026-09-12) to vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 (2026-09-18, digest sha256:5550994c1874ef331c6aed3ea27ae1efb7f4cfd17a2f18f5466c2837b9bf5467). The v0.29.0 release predates DeepSeek V4.1 support, so the family stays on the nightly channel. Recipe, TP8, DSpark settings, concurrency 1-32 and evals are unchanged. 将 MI325X DeepSeek-V4.1-Flash vLLM DSpark AgentX 镜像从 2026-09-12 的 ROCm nightly 更新为 2026-09-18 的 nightly-dee37d89115db4c94a820a79a78a7828e141c910；v0.29.0 发行版尚不支持 DeepSeek V4.1，因此继续使用 nightly 渠道。配方、TP8、DSpark 设置、 并发 1-32 与 eval 保持不变。 Co-Authored-By: Claude Fable 5.1 --- configs/amd-master.yaml | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index ed79e8c39..6080c3ba4 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -1194,14 +1194,17 @@ minimaxm3-fp8-mi325x-vllm-agentic-mtp: - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 10, 12, 14, 16, 18] } # DeepSeek-V4.1-Flash AgentX on MI325X (gfx942), a copy of the validated MI355X vLLM -# arm (run 34710937012) on the same ROCm nightly. gfx942 is not in the upstream -# hardware table (h100, h200, b200, gb200, gb300, mi350x); TP8 rather than TP4 -# because a 256 GB card must hold its share of the 511 GB checkpoint, the -# GPU-resident Engram tables and a 1M-context KV pool. Native five-token DSpark, +# arm (run 34710937012), originally on the same 2026-09-12 ROCm nightly. gfx942 is +# not in the upstream hardware table (h100, h200, b200, gb200, gb300, mi350x); TP8 +# rather than TP4 because a 256 GB card must hold its share of the 511 GB checkpoint, +# the GPU-resident Engram tables and a 1M-context KV pool. Native five-token DSpark, # golden AL 3.51 for throughput, real block verification for evals, adaptive # verification off on ROCm. +# Refreshed to the 2026-09-18 upstream ROCm nightly (vllm-project/vllm@dee37d89, +# Docker Hub digest sha256:5550994c1874ef331c6aed3ea27ae1efb7f4cfd17a2f18f5466c2837b9bf5467) +# on the same ROCm base image; the v0.29.0 release predates DeepSeek V4.1 support. dsv41flash-fp4-mi325x-vllm-agentic-dspark: - image: vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657 + image: vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 model: deepseek-ai/DeepSeek-V4.1-Flash model-prefix: dsv41flash runner: cluster:mi325x-amds From 99b06b5ce09f7cbfd6f4bcfd1d69e80e88d4b4f0 Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Fri, 18 Sep 2026 13:57:10 +0000 Subject: [PATCH 2/2] perf-changelog: record the dsv41flash-fp4-mi325x-vllm-agentic-dspark image refresh MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Append the changelog entry for moving the MI325X DeepSeek-V4.1-Flash vLLM DSpark AgentX family to vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 after the smoke throughput point and gsm8k eval passed in run 35346379012. 为 dsv41flash-fp4-mi325x-vllm-agentic-dspark 镜像更新追加 perf-changelog 条目； smoke 吞吐点与 gsm8k eval 已在运行 35346379012 中通过。 Co-Authored-By: Claude Fable 5.1 --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index a112ba56d..46cd92fa4 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -8082,3 +8082,10 @@ - "Capture only full decode CUDA graphs (VLLM_USE_BREAKABLE_CUDAGRAPH=0 with cudagraph_mode FULL_DECODE_ONLY, as the MiniMax-M3 gfx942 arm does) and restore --moe-backend aiter: on gfx942 every worker segfaulted during piecewise graph capture with both the Triton W4A16 MoE kernel (run 35305045778) and the auto-selected unfused Triton kernel (run 35306398350), so the capture mode rather than the MoE kernel is the failing piece; prefill runs eagerly" - "仅捕获完整的 decode CUDA graph（VLLM_USE_BREAKABLE_CUDAGRAPH=0 并设置 cudagraph_mode FULL_DECODE_ONLY，与 MiniMax-M3 gfx942 配方一致）并恢复 --moe-backend aiter：在 gfx942 上，无论使用 Triton W4A16 MoE 内核（运行 35305045778）还是自动选择的未融合 Triton 内核（运行 35306398350），所有 worker 都在 piecewise graph 捕获期间段错误，说明问题在于捕获模式而非 MoE 内核；prefill 以 eager 方式运行" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3247 + +- config-keys: + - dsv41flash-fp4-mi325x-vllm-agentic-dspark + description: + - "Update the MI325X DeepSeek-V4.1-Flash vLLM DSpark AgentX image from vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657 (2026-09-12) to vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910 (2026-09-18, Docker Hub digest sha256:5550994c1874ef331c6aed3ea27ae1efb7f4cfd17a2f18f5466c2837b9bf5467, same base-nightly ROCm layer). v0.29.0 predates DeepSeek V4.1 support, so the family stays on the upstream nightly channel; recipe flags, TP8, five-token DSpark with golden AL 3.51, concurrency 1-32 and the gsm8k eval are unchanged." + - "将 MI325X DeepSeek-V4.1-Flash vLLM DSpark AgentX 镜像从 vllm/vllm-openai-rocm:nightly-eed1f3d0c6043bd494424a22443ee198dd56f657（2026-09-12）更新为 vllm/vllm-openai-rocm:nightly-dee37d89115db4c94a820a79a78a7828e141c910（2026-09-18，Docker Hub 摘要 sha256:5550994c1874ef331c6aed3ea27ae1efb7f4cfd17a2f18f5466c2837b9bf5467，base-nightly ROCm 层相同）。v0.29.0 早于 DeepSeek V4.1 支持，因此该配置族继续使用上游 nightly 渠道；配方参数、TP8、五 token DSpark（黄金 AL 3.51）、并发 1-32 与 gsm8k eval 保持不变。" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3258