From a4f98a581143b2eb502ebda0ac821787d67214eb Mon Sep 17 00:00:00 2001 From: avininjamay8 Date: Mon, 10 Aug 2026 18:37:07 +0000 Subject: [PATCH 1/2] [CI][AMD][Disagg] Fix Kimi K2.5/K2.6 MXFP4 MLA backends on ROCm nightly for accuracy eval Use TRITON_MLA decode with FLASH_ATTN MLA prefill instead of ROCM_AITER_MLA so P/D GSM8K passes on current vllm-openai-rocm:nightly (avoids aiter fmha). Co-authored-by: Cursor --- .buildkite/amd-disagg/models.yaml | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/.buildkite/amd-disagg/models.yaml b/.buildkite/amd-disagg/models.yaml index e919475db85e..a6bb47a465bb 100644 --- a/.buildkite/amd-disagg/models.yaml +++ b/.buildkite/amd-disagg/models.yaml @@ -76,7 +76,7 @@ models: VLLM_ROCM_USE_AITER: "1" VLLM_ROCM_USE_AITER_MOE: "1" VLLM_ROCM_USE_AITER_RMSNORM: "1" - base_flags: "--trust-remote-code" + base_flags: "--trust-remote-code --attention-backend TRITON_MLA -ac.mla_prefill_backend=FLASH_ATTN --disable-custom-all-reduce" prefill: tp: "--gpu-memory-utilization 0.85" decode: @@ -86,11 +86,10 @@ models: env: VLLM_ROCM_USE_AITER: "1" AMDGCN_USE_BUFFER_OPS: "1" - VLLM_ROCM_USE_AITER_MLA: "1" VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: "INT4" VLLM_ROCM_USE_SKINNY_GEMM: "0" VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1" - base_flags: "--trust-remote-code --kv-cache-dtype fp8 --mm-encoder-tp-mode data --block-size 1 --attention-backend ROCM_AITER_MLA" + base_flags: "--trust-remote-code --kv-cache-dtype fp8 --mm-encoder-tp-mode data --block-size 16 --attention-backend TRITON_MLA -ac.mla_prefill_backend=FLASH_ATTN --disable-custom-all-reduce" prefill: tp: "--gpu-memory-utilization 0.9" decode: From a151894c07fa492eb0ffb695f2104b84c0bccf31 Mon Sep 17 00:00:00 2001 From: avininjamay8 Date: Wed, 12 Aug 2026 09:31:25 +0000 Subject: [PATCH 2/2] Remove --disable-custom-all-reduce from Kimi disagg CI config vLLM #51635 restores TCPStore when AITER custom all-reduce is enabled, so the workaround is no longer needed on nightly (validated in spur job 4357). Co-authored-by: Cursor --- .buildkite/amd-disagg/models.yaml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.buildkite/amd-disagg/models.yaml b/.buildkite/amd-disagg/models.yaml index a6bb47a465bb..0d44dea6bc27 100644 --- a/.buildkite/amd-disagg/models.yaml +++ b/.buildkite/amd-disagg/models.yaml @@ -76,7 +76,7 @@ models: VLLM_ROCM_USE_AITER: "1" VLLM_ROCM_USE_AITER_MOE: "1" VLLM_ROCM_USE_AITER_RMSNORM: "1" - base_flags: "--trust-remote-code --attention-backend TRITON_MLA -ac.mla_prefill_backend=FLASH_ATTN --disable-custom-all-reduce" + base_flags: "--trust-remote-code --attention-backend TRITON_MLA -ac.mla_prefill_backend=FLASH_ATTN" prefill: tp: "--gpu-memory-utilization 0.85" decode: @@ -89,7 +89,7 @@ models: VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: "INT4" VLLM_ROCM_USE_SKINNY_GEMM: "0" VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1" - base_flags: "--trust-remote-code --kv-cache-dtype fp8 --mm-encoder-tp-mode data --block-size 16 --attention-backend TRITON_MLA -ac.mla_prefill_backend=FLASH_ATTN --disable-custom-all-reduce" + base_flags: "--trust-remote-code --kv-cache-dtype fp8 --mm-encoder-tp-mode data --block-size 16 --attention-backend TRITON_MLA -ac.mla_prefill_backend=FLASH_ATTN" prefill: tp: "--gpu-memory-utilization 0.9" decode: