From 7a4c813ba69f0dd0753f3e24b8af36d7741e936d Mon Sep 17 00:00:00 2001 From: Xin Li Date: Mon, 3 Aug 2026 10:13:26 -0400 Subject: [PATCH 1/5] feat(minimaxm3-fp4-b300-dynamo-vllm-mtp): bump vLLM nightly to 5e35a6f4, enable cutlass MSA decode backend MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 中文：将 minimaxm3-fp4-b300-dynamo-vllm-mtp vLLM 镜像升级至 nightly-5e35a6f4f9bbc217c599692157ca985c894373f7，更新 B300 FP4 MTP recipe YAML（排除 legacy-dep4 专用配方），并启用 VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass Co-Authored-By: Claude Sonnet 4.6 (1M context) --- .../b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml | 4 +++- .../b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml | 4 +++- .../b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml | 4 +++- .../b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml | 4 +++- .../b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml | 4 +++- configs/nvidia-master.yaml | 2 +- 6 files changed, 16 insertions(+), 6 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml index 165301c53c..b0bec777dc 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml @@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p1d-dep2-tp4-fp4-8k1k-eagle3" model: path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9" + container: "vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7" precision: fp4 resources: @@ -32,10 +32,12 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml index b32295e325..18df59e860 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml @@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p2d-dep2-tp4-fp4-8k1k-eagle3" model: path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9" + container: "vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7" precision: fp4 resources: @@ -32,10 +32,12 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml index ff1b0c73c7..a98301918f 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml @@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p4d-dep2-tp4-fp4-8k1k-eagle3" model: path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9" + container: "vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7" precision: fp4 resources: @@ -32,10 +32,12 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml index c5f096d64d..46e1156417 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml @@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-1p6d-dep2-tp4-fp4-8k1k-eagle3" model: path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9" + container: "vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7" precision: fp4 resources: @@ -32,10 +32,12 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml index 790e6759a4..29ec988507 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml @@ -2,7 +2,7 @@ name: "minimax-m3-vllm-disagg-b300-2p3d-dep2-tp4-fp4-8k1k-eagle3" model: path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9" + container: "vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7" precision: fp4 resources: @@ -32,10 +32,12 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc + VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index a30af59641..ce3c6c119b 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7768,7 +7768,7 @@ minimaxm3-fp4-b300-dynamo-vllm-8k1k-legacy-max-tput: dp-attn: false minimaxm3-fp4-b300-dynamo-vllm-mtp: - image: vllm/vllm-openai:nightly-4080263bb2c5d10deac17aaeb88e0823bc35bca9 + image: vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 model: nvidia/MiniMax-M3-NVFP4 model-prefix: minimaxm3 runner: b300 From 8037b9ea2a0a62b5be1da41cfe997c239538b26a Mon Sep 17 00:00:00 2001 From: Xin Li Date: Mon, 3 Aug 2026 10:13:59 -0400 Subject: [PATCH 2/5] chore(perf-changelog): add minimaxm3-fp4-b300-dynamo-vllm-mtp nightly bump entry Co-Authored-By: Claude Sonnet 4.6 (1M context) --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 6eebd97070..446189a9ac 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5355,3 +5355,10 @@ - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path" - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 + +- config-keys: + - minimaxm3-fp4-b300-dynamo-vllm-mtp + description: + - "Bump vLLM nightly image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (fixed-len MiniMax-M3 fix); update B300 FP4 MTP recipe YAML container fields to match (excluding 2p1d-dep2-dep4 used by legacy-dep4)" + - "Enable VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass in prefill and decode environments" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2471 From 9534f06476952f15c6645428e2bb3074741911a9 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Mon, 3 Aug 2026 10:42:33 -0400 Subject: [PATCH 3/5] fix(minimaxm3-fp4-b300-dynamo-vllm-mtp): remove cutlass MSA env var from prefill (no cudagraph) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 中文：从预填充环境中移除 VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass（预填充工作进程不使用 CUDA graph） --- .../minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml | 1 - .../minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml | 1 - .../minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml | 1 - .../minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml | 1 - .../minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml | 1 - 5 files changed, 5 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml index b0bec777dc..f62103b0e4 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml @@ -32,7 +32,6 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml index 18df59e860..8338306055 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml @@ -32,7 +32,6 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml index a98301918f..03389625c9 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml @@ -32,7 +32,6 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml index 46e1156417..04823c9a1f 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml @@ -32,7 +32,6 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml index 29ec988507..0bffff9ea9 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml @@ -32,7 +32,6 @@ backend: prefill_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high From b6828c9e00656bc5539c227f96b3aa029bd77d0a Mon Sep 17 00:00:00 2001 From: Xin Li Date: Mon, 3 Aug 2026 11:35:31 -0400 Subject: [PATCH 4/5] fix(minimaxm3-fp4-b300-dynamo-vllm-mtp): use attention_config field for cutlass MSA decode, not env var MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 中文：将 decode_environment 中的 VLLM_MINIMAX_M3_MSA_DECODE_BACKEND 替换为解码侧 attention-config JSON 中的 minimax_m3_msa_decode_backend 字段 --- .../b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml | 3 +-- .../b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml | 3 +-- .../b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml | 3 +-- .../b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml | 3 +-- .../b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml | 3 +-- 5 files changed, 5 insertions(+), 10 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml index f62103b0e4..1fc5400781 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml @@ -36,7 +36,6 @@ backend: decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: @@ -67,7 +66,7 @@ backend: trust-remote-code: true no-enable-prefix-caching: true kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' - attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' + attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 gpu-memory-utilization: 0.95 max-model-len: 9472 diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml index 8338306055..9b308b5ed6 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml @@ -36,7 +36,6 @@ backend: decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: @@ -67,7 +66,7 @@ backend: trust-remote-code: true no-enable-prefix-caching: true kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' - attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' + attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 gpu-memory-utilization: 0.95 max-model-len: 9472 diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml index 03389625c9..daba06f113 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml @@ -36,7 +36,6 @@ backend: decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: @@ -67,7 +66,7 @@ backend: trust-remote-code: true no-enable-prefix-caching: true kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' - attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' + attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 gpu-memory-utilization: 0.95 max-model-len: 9472 diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml index 04823c9a1f..effedff045 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml @@ -36,7 +36,6 @@ backend: decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: @@ -67,7 +66,7 @@ backend: trust-remote-code: true no-enable-prefix-caching: true kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' - attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' + attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 gpu-memory-utilization: 0.95 max-model-len: 9472 diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml index 0bffff9ea9..87deb9a4fe 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml @@ -36,7 +36,6 @@ backend: decode_environment: VLLM_FLOAT32_MATMUL_PRECISION: high UCX_TLS: cuda_copy,cuda_ipc,rc - VLLM_MINIMAX_M3_MSA_DECODE_BACKEND: cutlass vllm_config: prefill: @@ -67,7 +66,7 @@ backend: trust-remote-code: true no-enable-prefix-caching: true kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' - attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' + attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 gpu-memory-utilization: 0.95 max-model-len: 9472 From 2f3c570c4c6ea2f93edc913a94b01652610a0880 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Mon, 3 Aug 2026 13:11:16 -0400 Subject: [PATCH 5/5] fix(minimaxm3-fp4-b300-dynamo-vllm-mtp): add kv-cache-dtype fp8 to MTP recipe YAMLs MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 中文：在 B300 FP4 MTP recipe YAML 的预填充与解码配置中补充 kv-cache-dtype: fp8 Co-Authored-By: Claude Sonnet 4.6 (1M context) --- .../minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml | 2 ++ .../minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml | 2 ++ .../minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml | 2 ++ .../minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml | 2 ++ .../minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml | 2 ++ 5 files changed, 10 insertions(+) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml index 1fc5400781..4269a63c8d 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml @@ -50,6 +50,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true @@ -68,6 +69,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml index 9b308b5ed6..7cedc791e0 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml @@ -50,6 +50,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true @@ -68,6 +69,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml index daba06f113..843269fa26 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml @@ -50,6 +50,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true @@ -68,6 +69,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml index effedff045..a7101ec777 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml @@ -50,6 +50,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true @@ -68,6 +69,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml index 87deb9a4fe..376503040a 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml @@ -50,6 +50,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true @@ -68,6 +69,7 @@ backend: kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}' attention-config: '{"backend": "FLASHINFER", "use_trtllm_attention": true, "indexer_kv_dtype": "fp8", "minimax_m3_msa_decode_backend": "cutlass"}' block-size: 128 + kv-cache-dtype: fp8 gpu-memory-utilization: 0.95 max-model-len: 9472 language-model-only: true