From c7b277d57ea472b01c99031786a9af1c5255addf Mon Sep 17 00:00:00 2001 From: Fangzhou-Ai Date: Mon, 7 Sep 2026 02:54:39 +0000 Subject: [PATCH] [AMD][DSV4] Repin MI355X vLLM entries to the nightly shipping FHMoE Fuse the DeepSeek-V4 native-FP8 shared expert into the MXFP4 routed-expert AITER kernel by moving the 8K/1K STP, 8K/1K MTP, and AgentX entries onto the first ROCm nightly containing vllm-project/vllm#53161 (vllm@de69e821). VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 was already requested by the AgentX script, but every earlier nightly rejected this mixed FP4+FP8 checkpoint in vLLM's eligibility check and self-disabled the fusion. Co-authored-by: Cursor Agent Signed-off-by: Fangzhou Ai --- configs/amd-master.yaml | 6 +++--- perf-changelog.yaml | 13 +++++++++++++ 2 files changed, 16 insertions(+), 3 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 7fa80d605..4151d14b0 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -1202,7 +1202,7 @@ dsv4-fp4-mi355x-sglang-mtp: # gpu-mem-util=0.6. TP8 sweeps conc 4-64; DEP8 has a single conc=64 # probe to validate the ROCm DP+EP path. dsv4-fp4-mi355x-vllm: - image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa + image: vllm/vllm-openai-rocm:nightly-de69e821b7c834bcf1ec96325ec8c410ac183680 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: mi355x @@ -1225,7 +1225,7 @@ dsv4-fp4-mi355x-vllm: # above ~conc32 (-37% @ conc32). Image reuses the base entry's v0.22.0 ROCm # build, which already contains the MTP commit. dsv4-fp4-mi355x-vllm-mtp: - image: vllm/vllm-openai-rocm:nightly-09663abde0f50944a8d5ea30120666024b503faa + image: vllm/vllm-openai-rocm:nightly-de69e821b7c834bcf1ec96325ec8c410ac183680 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: mi355x @@ -1316,7 +1316,7 @@ qwen3.5-fp8-mi325x-sglang-mtp: - { tp: 8, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } dsv4-fp4-mi355x-vllm-agentic-mtp: - image: vllm/vllm-openai-rocm:nightly-311b3513af33bc29b4acb2fde2e9313e5e9966a0 + image: vllm/vllm-openai-rocm:nightly-de69e821b7c834bcf1ec96325ec8c410ac183680 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: cluster:mi355x-amds diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 912728b2e..03333720c 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6902,3 +6902,16 @@ - "Lower mem-fraction-static from 0.89 to 0.86 on every arm: TP4, TP8 and TP8 with DP attention. swa-full-tokens-ratio becomes per-arm: 0.10 on the tensor-parallel arms as before, 0.15 under DP attention." - "Drop concurrency 2 and 10 from the TP4 arm, leaving [1, 4, 8], and drop concurrency 16 from the TP8 hicache arm, leaving [32, 48]. Concurrency 16 remains on the TP8 no-offload arm." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2800 + +- config-keys: + - dsv4-fp4-mi355x-vllm + - dsv4-fp4-mi355x-vllm-mtp + - dsv4-fp4-mi355x-vllm-agentic-mtp + scenario-type: + - fixed-seq-len + - agentic-coding + description: + - "Repin all three DSv4 MI355X vLLM entries to the first ROCm nightly that ships vllm-project/vllm#53161 (vllm-project/vllm@de69e821), which fuses the DeepSeek-V4 native-FP8 shared expert into the MXFP4 routed-expert AITER kernel (FHMoE)." + - "VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 was already exported by the AgentX script and is added to the 8K/1K STP and MTP scripts by #2792, but on every nightly before de69e821 the request was inert: vLLM's eligibility check rejected this mixed FP4+FP8 checkpoint and the fusion self-disabled at startup. #53161 is what makes the flag take effect." + - "Fusion eligibility is gated on gfx950, tensor parallelism 8, data parallelism 1, no expert parallelism and no EPLB, --moe-backend aiter, BF16 model dtype, and the DeepSeek-V4 block-128 FP8 shared-expert quantization config. All three entries satisfy it; the AgentX DP-attention arm does not and keeps the unfused path." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/PENDING