From 4e52736885b61250765966a2b10d6cfa9d82d16f Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Sat, 30 May 2026 18:11:57 -0400 Subject: [PATCH] [Klaud Cold] Update dsv4-fp4-mi355x-vllm vLLM ROCm image to v0.22.0 Co-Authored-By: Claude Opus 4.8 (1M context) --- .github/configs/amd-master.yaml | 2 +- perf-changelog.yaml | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml index da9db6ed6..b667bd38f 100644 --- a/.github/configs/amd-master.yaml +++ b/.github/configs/amd-master.yaml @@ -2135,7 +2135,7 @@ dsv4-fp4-mi355x-sglang: # 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-4f940896a32c9e2a0eba7f50d521bf5f6b4de458 + image: vllm/vllm-openai-rocm:v0.22.0 model: deepseek-ai/DeepSeek-V4-Pro model-prefix: dsv4 runner: mi355x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index ad93a24b4..ec611d7f0 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3239,3 +3239,9 @@ - "Add --compilation-config '{\"mode\":3,\"cudagraph_mode\":\"PIECEWISE\"}' to vllm serve, mirroring `model.base_args` from the upstream recipe. `pass_config.fuse_minimax_qk_norm` from the recipe is intentionally omitted — it triggers an upstream NameError on ROCm because vllm/compilation/passes/pass_manager.py imports MiniMaxQKNormPass under `is_cuda()` (NVIDIA-only) while using it unconditionally" - "Conditionally enable VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1 per (TP, EP, CONC) — on for shapes where the AITER ASM paged-attention kernel exists in the gfx942 heuristic table (TP=2 EP=1 CONC<=16, TP=8 EP=8 CONC<=64), off otherwise. Above the thresholds vllm/v1/attention/backends/rocm_aiter_fa.py routes decode through aiter pa_fwd_asm and crashes with `RuntimeError: get_heuristic_kernel: cannot get heuristic kernel!` for MiniMax-M2.5's attention shape (gqa=6 block_size=32 qTile=0); below them the ASM auto-dispatch is the perf win the recipe wants. Thresholds confirmed across 17 bench cells + 3 eval cells in PR #1594 sweep run 26692603804. Mirrors the per-shape toggle pattern in benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh; can collapse to unconditional SHUFFLE=1 once AITER registers the missing kernel on gfx942" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1594 + +- config-keys: + - dsv4-fp4-mi355x-vllm + description: + - "Update vLLM ROCm image from nightly-4f940896a32c9e2a0eba7f50d521bf5f6b4de458 to v0.22.0" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1624