From 0a6b0d03489503c817e180ffb5d43a31f24021ba Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Thu, 21 May 2026 13:42:37 -0400 Subject: [PATCH 1/2] [Klaud Cold] minimaxm2.5-fp8-mi300x: add VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1 + --attention-backend ROCM_AITER_FA Adopts the AITER recipe pattern (shuffle KV-cache layout + AITER FA attention backend) used by the mi355x sibling. Matches AMD's recommended recipe for MiniMax-M2.x on ROCm vLLM. Co-Authored-By: Claude Opus 4.7 (1M context) --- benchmarks/single_node/minimaxm2.5_fp8_mi300x.sh | 2 ++ perf-changelog.yaml | 6 ++++++ 2 files changed, 8 insertions(+) diff --git a/benchmarks/single_node/minimaxm2.5_fp8_mi300x.sh b/benchmarks/single_node/minimaxm2.5_fp8_mi300x.sh index 65cb8ee8e..21c44f2d8 100755 --- a/benchmarks/single_node/minimaxm2.5_fp8_mi300x.sh +++ b/benchmarks/single_node/minimaxm2.5_fp8_mi300x.sh @@ -24,6 +24,7 @@ if [ -n "$ROCR_VISIBLE_DEVICES" ]; then fi export VLLM_ROCM_USE_AITER=1 +export VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1 SERVER_LOG=/workspace/server.log PORT=${PORT:-8888} @@ -42,6 +43,7 @@ vllm serve $MODEL --port $PORT \ --max-model-len $MAX_MODEL_LEN \ --block-size=32 \ --no-enable-prefix-caching \ +--attention-backend "ROCM_AITER_FA" \ --trust-remote-code > $SERVER_LOG 2>&1 & SERVER_PID=$! diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 684d40dcc..471697332 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3068,3 +3068,9 @@ description: - "Bump image to rocm/sgl-dev:rocm720-mi35x-8c3b5aa-20260521-DSv4" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1548 + +- config-keys: + - minimaxm2.5-fp8-mi300x-vllm + description: + - "Add VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1 + --attention-backend ROCM_AITER_FA (match AMD-recommended AITER recipe pattern used on mi355x)" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX From 28a68fe93dae9cd010473998862bcf445001d34d Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Thu, 21 May 2026 13:42:44 -0400 Subject: [PATCH 2/2] chore: fill pr-link for #1550 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 471697332..22143c7f8 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -3073,4 +3073,4 @@ - minimaxm2.5-fp8-mi300x-vllm description: - "Add VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1 + --attention-backend ROCM_AITER_FA (match AMD-recommended AITER recipe pattern used on mi355x)" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1550