From a762d65270965e3d46e79897605ed63af02b97aa Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Mon, 7 Sep 2026 18:57:29 -0400 Subject: [PATCH 1/2] [Klaud Cold] Update minimaxm3-fp4-b200-vllm-agentic-mtp vLLM image to nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 Co-Authored-By: Claude Fable 5.1 --- configs/nvidia-master.yaml | 2 +- perf-changelog.yaml | 8 ++++++++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 0a6c86df6..afab7bd10 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7777,7 +7777,7 @@ minimaxm3-fp4-b300-trtllm-agentic-mtp: # the same 3 TB AgentX ceiling before the proportional-GPU rule is applied. # GPU-resident points receive a zero budget. minimaxm3-fp4-b200-vllm-agentic-mtp: - image: vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45 + image: vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 model: nvidia/MiniMax-M3-NVFP4 model-prefix: minimaxm3 runner: cluster:b200-nscale diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 6b249ea15..64ea4600e 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6921,3 +6921,11 @@ - "Pick up the latest automatic ROCm DeepSeek-V4 optimizations, including fused mHC post/pre plus RMSNorm, gfx950 C4A top-k dispatch, fused C4 compressor GEMMs, fused SWA q/kv RMSNorm plus q FP8 quantization, and medium-batch cooperative top-k tuning." - "Keep the existing VLLM_ROCM_USE_AITER=1, VLLM_ROCM_USE_AITER_MOE=1, and --moe-backend aiter settings, and explicitly add VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 plus VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 to both STP and MTP paths. The current checkpoint's shared-expert path does not satisfy the latest vLLM fusion conditions, so that fusion flag self-disables while preserving recipe parity." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2792 + +- config-keys: + - minimaxm3-fp4-b200-vllm-agentic-mtp + scenario-type: + - agentic-coding + description: + - "Update vLLM image from vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45 (2026-08-30 nightly) to vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 nightly, digest sha256:254eebf919e8b7b0d530d97fccc606c36f380ff6724d64bc190951bec1aee838, tag commit vllm-project/vllm@d9105ea8). Docker Hub last pushed the tag at 2026-09-07T06:16:01Z." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2860 From 50e127a82fc177c56f007f6366b612c44ebd01ed Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Mon, 7 Sep 2026 23:06:30 -0400 Subject: [PATCH 2/2] Re-pin to the 2026-09-04 vLLM nightly: the 09-05 FlashAttention sync breaks the FA4 CuTe fp8-KV descale path used by the EAGLE3 draft on Blackwell Co-Authored-By: Claude Fable 5.1 --- configs/nvidia-master.yaml | 2 +- perf-changelog.yaml | 1 + 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index afab7bd10..0413e97b0 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7777,7 +7777,7 @@ minimaxm3-fp4-b300-trtllm-agentic-mtp: # the same 3 TB AgentX ceiling before the proportional-GPU rule is applied. # GPU-resident points receive a zero budget. minimaxm3-fp4-b200-vllm-agentic-mtp: - image: vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 + image: vllm/vllm-openai:nightly-8a728663c1c3eeace834a95f5654fa653cc1998c model: nvidia/MiniMax-M3-NVFP4 model-prefix: minimaxm3 runner: cluster:b200-nscale diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 64ea4600e..4de994d12 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6928,4 +6928,5 @@ - agentic-coding description: - "Update vLLM image from vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45 (2026-08-30 nightly) to vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 nightly, digest sha256:254eebf919e8b7b0d530d97fccc606c36f380ff6724d64bc190951bec1aee838, tag commit vllm-project/vllm@d9105ea8). Docker Hub last pushed the tag at 2026-09-07T06:16:01Z." + - "Re-pin from vllm/vllm-openai:nightly-d9105ea8001e0a6d77a96327d17515bb5791fb36 (2026-09-07 nightly) to vllm/vllm-openai:nightly-8a728663c1c3eeace834a95f5654fa653cc1998c (2026-09-04 nightly, digest sha256:f5df5cc3302b5f404848c4eca88d7bf7ed5226e151c056da22816d7734644d67; Docker Hub last pushed 2026-09-04T06:18:14Z, tag commit vllm-project/vllm@8a728663). Nightlies from 2026-09-05 onward carry vllm-project/vllm@4ee259551 (Sync FA with upstream, #54819), which moves vllm-flash-attn to 506341a1; on Blackwell the EAGLE3 draft runs FLASH_ATTN through the FA4 CuTe path and, with the fp8 KV cache, its descale tensors now fail to_cute_tensor with RuntimeError: Expected strides[leading_dim] == 1, but got 0 during CUDA-graph memory profiling, so the engine never starts (B200 run 34168437161). The 2026-09-04 nightly is the newest build still on vllm-flash-attn 06bdd47c, the same FA pin as the recipe's previous 2026-08-30 image, and the same vllm commit the MI355X MiniMax-M3 vLLM AgentX recipe already runs (nightly-8a728663). No vllm fix has landed on main as of 2026-09-08T03:00Z." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2860