From 4899e76b37866b8e66ca07c0088880f71c582020 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 31 Jul 2026 09:43:08 -0400 Subject: [PATCH 1/3] =?UTF-8?q?kimik3-fp4-b300-vllm-agentic-dspark:=20appe?= =?UTF-8?q?nd=20perf-changelog=20entry=20/=20=E8=BF=BD=E5=8A=A0=20perf-cha?= =?UTF-8?q?ngelog=20=E6=9D=A1=E7=9B=AE?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- perf-changelog.yaml | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 6eebd9707..ce642737d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5355,3 +5355,9 @@ - "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: + - kimik3-fp4-b300-vllm-agentic-dspark + description: + - "Update DSpark recipe: K=7 (AL 3.84), native offload, GPU-resident conc [1,2,4,8,12,14,16,18], offload conc [14,16,18,20,22,24], prefix-match-unit 64 (GPU-resident only), gsm8k eval." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2419 From 91804e88a0038084e506165269468275405fef12 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 31 Jul 2026 09:45:47 -0400 Subject: [PATCH 2/3] restore mtp recipes, switch 2xDEP2 prefill to 1xDEP4 --- .../8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml | 94 ++++++++++++++++ .../8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml | 96 +++++++++++++++++ .../8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml | 94 ++++++++++++++++ .../8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml | 94 ++++++++++++++++ .../8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml | 94 ++++++++++++++++ .../8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml | 94 ++++++++++++++++ configs/nvidia-master.yaml | 101 ++++++++++++++++++ 7 files changed, 667 insertions(+) create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml new file mode 100644 index 000000000..8cc390fbe --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml @@ -0,0 +1,94 @@ +name: "minimax-m3-vllm-disagg-b200-1p1d-dep2-tp4-fp4-8k1k-eagle3" + +model: + path: "nvidia/MiniMax-M3-NVFP4" + container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" + precision: fp4 + +resources: + gpu_type: b200 + gpus_per_node: 8 + prefill_nodes: 1 + decode_nodes: 0 + prefill_workers: 1 + decode_workers: 1 + gpus_per_prefill: 2 + gpus_per_decode: 4 + +dynamo: + install: true + version: 1.3.0.dev20260710 + +frontend: + type: dynamo + enable_multiple_frontends: false + +backend: + type: vllm + connector: null + allow_prefill_decode_colocation: true + allow_prefill_decode_colocation_across_nodes: true + + prefill_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + decode_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + vllm_config: + prefill: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 2 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-cudagraph-capture-size: 2048 + max-num-batched-tokens: 16384 + + decode: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 4 + pipeline-parallel-size: 1 + enable-expert-parallel: false + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-num-seqs: 1024 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 2048 + +health_check: + max_attempts: 360 + interval_seconds: 10 + +benchmark: + type: sa-bench + isl: 8192 + osl: 1024 + req_rate: inf + num_warmup_mult: 0 + random_range_ratio: 0.8 + use_chat_template: true + concurrencies: "64" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml new file mode 100644 index 000000000..e8487333d --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml @@ -0,0 +1,96 @@ +name: "minimax-m3-vllm-disagg-b200-1p1d-dep4-dep4-fp4-8k1k-eagle3" + +model: + path: "nvidia/MiniMax-M3-NVFP4" + container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" + precision: fp4 + +resources: + gpu_type: b200 + gpus_per_node: 8 + prefill_nodes: 1 + decode_nodes: 0 + prefill_workers: 1 + decode_workers: 1 + gpus_per_prefill: 4 + gpus_per_decode: 4 + +dynamo: + install: true + version: 1.3.0.dev20260710 + +frontend: + type: dynamo + enable_multiple_frontends: false + +backend: + type: vllm + connector: null + allow_prefill_decode_colocation: true + allow_prefill_decode_colocation_across_nodes: true + + prefill_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + decode_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + vllm_config: + prefill: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 4 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-cudagraph-capture-size: 2048 + max-num-batched-tokens: 16384 + + decode: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 4 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.9 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-num-seqs: 1024 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 2048 + +health_check: + max_attempts: 360 + interval_seconds: 10 + +benchmark: + type: sa-bench + isl: 8192 + osl: 1024 + req_rate: inf + num_warmup_mult: 0 + random_range_ratio: 0.8 + use_chat_template: true + concurrencies: "256x512" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml new file mode 100644 index 000000000..c77a46a32 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml @@ -0,0 +1,94 @@ +name: "minimax-m3-vllm-disagg-b200-1p2d-dep2-tp4-fp4-8k1k-eagle3" + +model: + path: "nvidia/MiniMax-M3-NVFP4" + container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" + precision: fp4 + +resources: + gpu_type: b200 + gpus_per_node: 8 + prefill_nodes: 1 + decode_nodes: 1 + prefill_workers: 1 + decode_workers: 2 + gpus_per_prefill: 2 + gpus_per_decode: 4 + +dynamo: + install: true + version: 1.3.0.dev20260710 + +frontend: + type: dynamo + enable_multiple_frontends: false + +backend: + type: vllm + connector: null + allow_prefill_decode_colocation: true + allow_prefill_decode_colocation_across_nodes: true + + prefill_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + decode_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + vllm_config: + prefill: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 2 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-cudagraph-capture-size: 2048 + max-num-batched-tokens: 16384 + + decode: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 4 + pipeline-parallel-size: 1 + enable-expert-parallel: false + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-num-seqs: 1024 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 2048 + +health_check: + max_attempts: 360 + interval_seconds: 10 + +benchmark: + type: sa-bench + isl: 8192 + osl: 1024 + req_rate: inf + num_warmup_mult: 0 + random_range_ratio: 0.8 + use_chat_template: true + concurrencies: "4x32x64" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml new file mode 100644 index 000000000..c011f00f8 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml @@ -0,0 +1,94 @@ +name: "minimax-m3-vllm-disagg-b200-1p3d-dep4-tp4-fp4-8k1k-eagle3" + +model: + path: "nvidia/MiniMax-M3-NVFP4" + container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" + precision: fp4 + +resources: + gpu_type: b200 + gpus_per_node: 8 + prefill_nodes: 1 + decode_nodes: 1 + prefill_workers: 1 + decode_workers: 3 + gpus_per_prefill: 4 + gpus_per_decode: 4 + +dynamo: + install: true + version: 1.3.0.dev20260710 + +frontend: + type: dynamo + enable_multiple_frontends: false + +backend: + type: vllm + connector: null + allow_prefill_decode_colocation: true + allow_prefill_decode_colocation_across_nodes: true + + prefill_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + decode_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + vllm_config: + prefill: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 4 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-cudagraph-capture-size: 2048 + max-num-batched-tokens: 16384 + + decode: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 4 + pipeline-parallel-size: 1 + enable-expert-parallel: false + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-num-seqs: 1024 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 2048 + +health_check: + max_attempts: 360 + interval_seconds: 10 + +benchmark: + type: sa-bench + isl: 8192 + osl: 1024 + req_rate: inf + num_warmup_mult: 0 + random_range_ratio: 0.8 + use_chat_template: true + concurrencies: "192" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml new file mode 100644 index 000000000..0347f11a0 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml @@ -0,0 +1,94 @@ +name: "minimax-m3-vllm-disagg-b200-1p4d-dep2-tp4-fp4-8k1k-eagle3" + +model: + path: "nvidia/MiniMax-M3-NVFP4" + container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" + precision: fp4 + +resources: + gpu_type: b200 + gpus_per_node: 8 + prefill_nodes: 1 + decode_nodes: 2 + prefill_workers: 1 + decode_workers: 4 + gpus_per_prefill: 2 + gpus_per_decode: 4 + +dynamo: + install: true + version: 1.3.0.dev20260710 + +frontend: + type: dynamo + enable_multiple_frontends: false + +backend: + type: vllm + connector: null + allow_prefill_decode_colocation: true + allow_prefill_decode_colocation_across_nodes: true + + prefill_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + decode_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + vllm_config: + prefill: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 2 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-cudagraph-capture-size: 2048 + max-num-batched-tokens: 16384 + + decode: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 4 + pipeline-parallel-size: 1 + enable-expert-parallel: false + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-num-seqs: 1024 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 2048 + +health_check: + max_attempts: 360 + interval_seconds: 10 + +benchmark: + type: sa-bench + isl: 8192 + osl: 1024 + req_rate: inf + num_warmup_mult: 0 + random_range_ratio: 0.8 + use_chat_template: true + concurrencies: "4x8x64" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml new file mode 100644 index 000000000..72c05f717 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml @@ -0,0 +1,94 @@ +name: "minimax-m3-vllm-disagg-b200-1p6d-dep2-tp4-fp4-8k1k-eagle3" + +model: + path: "nvidia/MiniMax-M3-NVFP4" + container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" + precision: fp4 + +resources: + gpu_type: b200 + gpus_per_node: 8 + prefill_nodes: 1 + decode_nodes: 3 + prefill_workers: 1 + decode_workers: 6 + gpus_per_prefill: 2 + gpus_per_decode: 4 + +dynamo: + install: true + version: 1.3.0.dev20260710 + +frontend: + type: dynamo + enable_multiple_frontends: false + +backend: + type: vllm + connector: null + allow_prefill_decode_colocation: true + allow_prefill_decode_colocation_across_nodes: true + + prefill_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + decode_environment: + VLLM_FLOAT32_MATMUL_PRECISION: high + VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm + UCX_TLS: cuda_copy,cuda_ipc,rc + + vllm_config: + prefill: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 1 + pipeline-parallel-size: 1 + data-parallel-size: 2 + data-parallel-rpc-port: 13345 + enable-expert-parallel: true + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-cudagraph-capture-size: 2048 + max-num-batched-tokens: 16384 + + decode: + no-enable-flashinfer-autotune: true + tensor-parallel-size: 4 + pipeline-parallel-size: 1 + enable-expert-parallel: false + 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"}' + block-size: 128 + gpu-memory-utilization: 0.95 + max-model-len: 9472 + language-model-only: true + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + stream-interval: 32 + max-num-seqs: 1024 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 2048 + +health_check: + max_attempts: 360 + interval_seconds: 10 + +benchmark: + type: sa-bench + isl: 8192 + osl: 1024 + req_rate: inf + num_warmup_mult: 0 + random_range_ratio: 0.8 + use_chat_template: true + concurrencies: "24x48" diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index a30af5964..587993e12 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7481,6 +7481,107 @@ minimaxm3-fp4-b200-dynamo-vllm: ep: 1 dp-attn: false +minimaxm3-fp4-b200-dynamo-vllm-mtp: + image: vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de + model: nvidia/MiniMax-M3-NVFP4 + model-prefix: minimaxm3 + runner: b200-multinode + precision: fp4 + framework: dynamo-vllm + router: { name: dynamo-router, version: "1.3.0.dev20260710" } + kv-p2p-transfer: nixl + multinode: true + disagg: true + scenarios: + fixed-seq-len: + - isl: 8192 + osl: 1024 + search-space: + - spec-decoding: "mtp" + conc-list: [24, 48] + prefill: + num-worker: 1 + tp: 2 + ep: 2 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml" + decode: + num-worker: 6 + tp: 4 + ep: 1 + dp-attn: false + - spec-decoding: "mtp" + conc-list: [4, 8, 64] + prefill: + num-worker: 1 + tp: 2 + ep: 2 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml" + decode: + num-worker: 4 + tp: 4 + ep: 1 + dp-attn: false + - spec-decoding: "mtp" + conc-list: [4, 32, 64] + prefill: + num-worker: 1 + tp: 2 + ep: 2 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml" + decode: + num-worker: 2 + tp: 4 + ep: 1 + dp-attn: false + - spec-decoding: "mtp" + conc-list: [64] + prefill: + num-worker: 1 + tp: 2 + ep: 2 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml" + decode: + num-worker: 1 + tp: 4 + ep: 1 + dp-attn: false + - spec-decoding: "mtp" + conc-list: [192] + prefill: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml" + decode: + num-worker: 3 + tp: 4 + ep: 1 + dp-attn: false + - spec-decoding: "mtp" + conc-list: [256, 512] + prefill: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml" + decode: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: true + # MiniMax-M3 NVFP4 disagg sweep on the same B300 topology matrix as the MXFP8 # baseline above. The image includes vLLM PR #46380, so no runtime patch is # needed. From 179538b4936be2b7b252ed01a4c634945688aeb5 Mon Sep 17 00:00:00 2001 From: Xin Li Date: Fri, 31 Jul 2026 09:48:05 -0400 Subject: [PATCH 3/3] Revert "restore mtp recipes, switch 2xDEP2 prefill to 1xDEP4" This reverts commit 91804e88a0038084e506165269468275405fef12. --- .../8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml | 94 ---------------- .../8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml | 96 ----------------- .../8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml | 94 ---------------- .../8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml | 94 ---------------- .../8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml | 94 ---------------- .../8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml | 94 ---------------- configs/nvidia-master.yaml | 101 ------------------ 7 files changed, 667 deletions(-) delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml deleted file mode 100644 index 8cc390fbe..000000000 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml +++ /dev/null @@ -1,94 +0,0 @@ -name: "minimax-m3-vllm-disagg-b200-1p1d-dep2-tp4-fp4-8k1k-eagle3" - -model: - path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" - precision: fp4 - -resources: - gpu_type: b200 - gpus_per_node: 8 - prefill_nodes: 1 - decode_nodes: 0 - prefill_workers: 1 - decode_workers: 1 - gpus_per_prefill: 2 - gpus_per_decode: 4 - -dynamo: - install: true - version: 1.3.0.dev20260710 - -frontend: - type: dynamo - enable_multiple_frontends: false - -backend: - type: vllm - connector: null - allow_prefill_decode_colocation: true - allow_prefill_decode_colocation_across_nodes: true - - prefill_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - decode_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - vllm_config: - prefill: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 2 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-cudagraph-capture-size: 2048 - max-num-batched-tokens: 16384 - - decode: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 4 - pipeline-parallel-size: 1 - enable-expert-parallel: false - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-num-seqs: 1024 - max-num-batched-tokens: 16384 - max-cudagraph-capture-size: 2048 - -health_check: - max_attempts: 360 - interval_seconds: 10 - -benchmark: - type: sa-bench - isl: 8192 - osl: 1024 - req_rate: inf - num_warmup_mult: 0 - random_range_ratio: 0.8 - use_chat_template: true - concurrencies: "64" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml deleted file mode 100644 index e8487333d..000000000 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml +++ /dev/null @@ -1,96 +0,0 @@ -name: "minimax-m3-vllm-disagg-b200-1p1d-dep4-dep4-fp4-8k1k-eagle3" - -model: - path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" - precision: fp4 - -resources: - gpu_type: b200 - gpus_per_node: 8 - prefill_nodes: 1 - decode_nodes: 0 - prefill_workers: 1 - decode_workers: 1 - gpus_per_prefill: 4 - gpus_per_decode: 4 - -dynamo: - install: true - version: 1.3.0.dev20260710 - -frontend: - type: dynamo - enable_multiple_frontends: false - -backend: - type: vllm - connector: null - allow_prefill_decode_colocation: true - allow_prefill_decode_colocation_across_nodes: true - - prefill_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - decode_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - vllm_config: - prefill: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 4 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-cudagraph-capture-size: 2048 - max-num-batched-tokens: 16384 - - decode: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 4 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.9 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-num-seqs: 1024 - max-num-batched-tokens: 16384 - max-cudagraph-capture-size: 2048 - -health_check: - max_attempts: 360 - interval_seconds: 10 - -benchmark: - type: sa-bench - isl: 8192 - osl: 1024 - req_rate: inf - num_warmup_mult: 0 - random_range_ratio: 0.8 - use_chat_template: true - concurrencies: "256x512" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml deleted file mode 100644 index c77a46a32..000000000 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml +++ /dev/null @@ -1,94 +0,0 @@ -name: "minimax-m3-vllm-disagg-b200-1p2d-dep2-tp4-fp4-8k1k-eagle3" - -model: - path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" - precision: fp4 - -resources: - gpu_type: b200 - gpus_per_node: 8 - prefill_nodes: 1 - decode_nodes: 1 - prefill_workers: 1 - decode_workers: 2 - gpus_per_prefill: 2 - gpus_per_decode: 4 - -dynamo: - install: true - version: 1.3.0.dev20260710 - -frontend: - type: dynamo - enable_multiple_frontends: false - -backend: - type: vllm - connector: null - allow_prefill_decode_colocation: true - allow_prefill_decode_colocation_across_nodes: true - - prefill_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - decode_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - vllm_config: - prefill: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 2 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-cudagraph-capture-size: 2048 - max-num-batched-tokens: 16384 - - decode: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 4 - pipeline-parallel-size: 1 - enable-expert-parallel: false - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-num-seqs: 1024 - max-num-batched-tokens: 16384 - max-cudagraph-capture-size: 2048 - -health_check: - max_attempts: 360 - interval_seconds: 10 - -benchmark: - type: sa-bench - isl: 8192 - osl: 1024 - req_rate: inf - num_warmup_mult: 0 - random_range_ratio: 0.8 - use_chat_template: true - concurrencies: "4x32x64" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml deleted file mode 100644 index c011f00f8..000000000 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml +++ /dev/null @@ -1,94 +0,0 @@ -name: "minimax-m3-vllm-disagg-b200-1p3d-dep4-tp4-fp4-8k1k-eagle3" - -model: - path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" - precision: fp4 - -resources: - gpu_type: b200 - gpus_per_node: 8 - prefill_nodes: 1 - decode_nodes: 1 - prefill_workers: 1 - decode_workers: 3 - gpus_per_prefill: 4 - gpus_per_decode: 4 - -dynamo: - install: true - version: 1.3.0.dev20260710 - -frontend: - type: dynamo - enable_multiple_frontends: false - -backend: - type: vllm - connector: null - allow_prefill_decode_colocation: true - allow_prefill_decode_colocation_across_nodes: true - - prefill_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - decode_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - vllm_config: - prefill: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 4 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-cudagraph-capture-size: 2048 - max-num-batched-tokens: 16384 - - decode: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 4 - pipeline-parallel-size: 1 - enable-expert-parallel: false - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-num-seqs: 1024 - max-num-batched-tokens: 16384 - max-cudagraph-capture-size: 2048 - -health_check: - max_attempts: 360 - interval_seconds: 10 - -benchmark: - type: sa-bench - isl: 8192 - osl: 1024 - req_rate: inf - num_warmup_mult: 0 - random_range_ratio: 0.8 - use_chat_template: true - concurrencies: "192" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml deleted file mode 100644 index 0347f11a0..000000000 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml +++ /dev/null @@ -1,94 +0,0 @@ -name: "minimax-m3-vllm-disagg-b200-1p4d-dep2-tp4-fp4-8k1k-eagle3" - -model: - path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" - precision: fp4 - -resources: - gpu_type: b200 - gpus_per_node: 8 - prefill_nodes: 1 - decode_nodes: 2 - prefill_workers: 1 - decode_workers: 4 - gpus_per_prefill: 2 - gpus_per_decode: 4 - -dynamo: - install: true - version: 1.3.0.dev20260710 - -frontend: - type: dynamo - enable_multiple_frontends: false - -backend: - type: vllm - connector: null - allow_prefill_decode_colocation: true - allow_prefill_decode_colocation_across_nodes: true - - prefill_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - decode_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - vllm_config: - prefill: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 2 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-cudagraph-capture-size: 2048 - max-num-batched-tokens: 16384 - - decode: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 4 - pipeline-parallel-size: 1 - enable-expert-parallel: false - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-num-seqs: 1024 - max-num-batched-tokens: 16384 - max-cudagraph-capture-size: 2048 - -health_check: - max_attempts: 360 - interval_seconds: 10 - -benchmark: - type: sa-bench - isl: 8192 - osl: 1024 - req_rate: inf - num_warmup_mult: 0 - random_range_ratio: 0.8 - use_chat_template: true - concurrencies: "4x8x64" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml deleted file mode 100644 index 72c05f717..000000000 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml +++ /dev/null @@ -1,94 +0,0 @@ -name: "minimax-m3-vllm-disagg-b200-1p6d-dep2-tp4-fp4-8k1k-eagle3" - -model: - path: "nvidia/MiniMax-M3-NVFP4" - container: "vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de" - precision: fp4 - -resources: - gpu_type: b200 - gpus_per_node: 8 - prefill_nodes: 1 - decode_nodes: 3 - prefill_workers: 1 - decode_workers: 6 - gpus_per_prefill: 2 - gpus_per_decode: 4 - -dynamo: - install: true - version: 1.3.0.dev20260710 - -frontend: - type: dynamo - enable_multiple_frontends: false - -backend: - type: vllm - connector: null - allow_prefill_decode_colocation: true - allow_prefill_decode_colocation_across_nodes: true - - prefill_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - decode_environment: - VLLM_FLOAT32_MATMUL_PRECISION: high - VLLM_FLASHINFER_ALLREDUCE_BACKEND: trtllm - UCX_TLS: cuda_copy,cuda_ipc,rc - - vllm_config: - prefill: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 1 - pipeline-parallel-size: 1 - data-parallel-size: 2 - data-parallel-rpc-port: 13345 - enable-expert-parallel: true - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":1,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-cudagraph-capture-size: 2048 - max-num-batched-tokens: 16384 - - decode: - no-enable-flashinfer-autotune: true - tensor-parallel-size: 4 - pipeline-parallel-size: 1 - enable-expert-parallel: false - 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"}' - block-size: 128 - gpu-memory-utilization: 0.95 - max-model-len: 9472 - language-model-only: true - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' - stream-interval: 32 - max-num-seqs: 1024 - max-num-batched-tokens: 16384 - max-cudagraph-capture-size: 2048 - -health_check: - max_attempts: 360 - interval_seconds: 10 - -benchmark: - type: sa-bench - isl: 8192 - osl: 1024 - req_rate: inf - num_warmup_mult: 0 - random_range_ratio: 0.8 - use_chat_template: true - concurrencies: "24x48" diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 587993e12..a30af5964 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7481,107 +7481,6 @@ minimaxm3-fp4-b200-dynamo-vllm: ep: 1 dp-attn: false -minimaxm3-fp4-b200-dynamo-vllm-mtp: - image: vllm/vllm-openai:nightly-6f91edf96d3f3272945809c04702380053bff4de - model: nvidia/MiniMax-M3-NVFP4 - model-prefix: minimaxm3 - runner: b200-multinode - precision: fp4 - framework: dynamo-vllm - router: { name: dynamo-router, version: "1.3.0.dev20260710" } - kv-p2p-transfer: nixl - multinode: true - disagg: true - scenarios: - fixed-seq-len: - - isl: 8192 - osl: 1024 - search-space: - - spec-decoding: "mtp" - conc-list: [24, 48] - prefill: - num-worker: 1 - tp: 2 - ep: 2 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml" - decode: - num-worker: 6 - tp: 4 - ep: 1 - dp-attn: false - - spec-decoding: "mtp" - conc-list: [4, 8, 64] - prefill: - num-worker: 1 - tp: 2 - ep: 2 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml" - decode: - num-worker: 4 - tp: 4 - ep: 1 - dp-attn: false - - spec-decoding: "mtp" - conc-list: [4, 32, 64] - prefill: - num-worker: 1 - tp: 2 - ep: 2 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml" - decode: - num-worker: 2 - tp: 4 - ep: 1 - dp-attn: false - - spec-decoding: "mtp" - conc-list: [64] - prefill: - num-worker: 1 - tp: 2 - ep: 2 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml" - decode: - num-worker: 1 - tp: 4 - ep: 1 - dp-attn: false - - spec-decoding: "mtp" - conc-list: [192] - prefill: - num-worker: 1 - tp: 4 - ep: 4 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p3d-dep4-tp4-eagle3-8k1k.yaml" - decode: - num-worker: 3 - tp: 4 - ep: 1 - dp-attn: false - - spec-decoding: "mtp" - conc-list: [256, 512] - prefill: - num-worker: 1 - tp: 4 - ep: 4 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/b200-fp4/8k1k/mtp/1p1d-dep4-dep4-eagle3-8k1k.yaml" - decode: - num-worker: 1 - tp: 4 - ep: 4 - dp-attn: true - # MiniMax-M3 NVFP4 disagg sweep on the same B300 topology matrix as the MXFP8 # baseline above. The image includes vLLM PR #46380, so no runtime patch is # needed.