From e0c832ac01e3d54eb0be8888811a74d59519c4a7 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 10:19:44 -0700 Subject: [PATCH 1/4] config(minimaxm3): update GB200 aggregate recipes MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 更新 MiniMax-M3 GB200 聚合配置，仅保留 SimpleCPU 卸载配方，并使用已提交的 thinking-on EAGLE3-GQA 接受长度。 --- configs/nvidia-master.yaml | 62 +++----------------------------------- perf-changelog.yaml | 9 ++++++ 2 files changed, 13 insertions(+), 58 deletions(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 0d5bd0c7d..6a01b4413 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7796,8 +7796,8 @@ minimaxm3-fp4-b200-trtllm-agentic-mtp: search-space: - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [5, 10, 15, 20, 25, 30, 35, 40, 45] } - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [1, 5] } -# Preserve the B200 TP4 search space and add the GB200 Pareto candidates found -# by direct DEP and P/D tuning. +# Retain the GB200 aggregate SimpleCPU-offload configuration at concurrency 20 +# and 30. minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: image: vllm/vllm-openai:v0.27.1 model: nvidia/MiniMax-M3-NVFP4 @@ -7812,23 +7812,10 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: agentic-coding: - dram-utilization: 0.61 search-space: - - spec-decoding: mtp - kv-offloading: none - conc-list: [1, 2, 5, 8, 10, 12, 15, 20] - num-nodes: 1 - worker: - num-worker: 1 - tp: 4 - ep: 1 - dp-attn: false - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-agentic.yaml" - - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - spec-decoding: mtp kv-offloading: dram kv-offload-backend: { name: vllm-simple } - conc-list: [20, 30, 40] + conc-list: [20, 30] num-nodes: 1 worker: num-worker: 1 @@ -7838,48 +7825,7 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: additional-settings: - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - - spec-decoding: mtp - kv-offloading: none - conc-list: [4, 32] - num-nodes: 1 - worker: - num-worker: 1 - tp: 4 - ep: 4 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-agentic.yaml" - - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - - spec-decoding: mtp - kv-offloading: dram - kv-offload-backend: { name: vllm-simple } - conc-list: [32, 40] - num-nodes: 1 - worker: - num-worker: 1 - tp: 4 - ep: 4 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-vllm-simple-agentic.yaml" - - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - - spec-decoding: mtp - kv-offloading: none - conc-list: [48] - num-nodes: 2 - worker: - num-worker: 1 - tp: 8 - ep: 8 - dp-attn: true - additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep8-agentic.yaml" - - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: image: vllm/vllm-openai:v0.27.1 model: nvidia/MiniMax-M3-NVFP4 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 4eb0a2caa..2edb0c148 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6821,3 +6821,12 @@ description: - "Refresh to collect TensorRT-LLM server metrics." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2774 + +- config-keys: + - minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp + scenario-type: + - agentic-coding + description: + - "Limit the GB200 aggregate Dynamo-vLLM configuration to the SimpleCPU-offload concurrency 20 and 30 recipes." + - "Use the committed MiniMax-M3 thinking-on EAGLE3-GQA acceptance target." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2806 From 4a50379b236c070d88d94b084896c425f57893df Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 10:49:17 -0700 Subject: [PATCH 2/4] fix(minimaxm3): align aggregate GQA acceptance target MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 将 MiniMax-M3 GB200 聚合 AgentX 配置与仓库中已提交的 EAGLE3-GQA thinking-on 接受长度保持一致。 --- configs/nvidia-master.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 6a01b4413..0aac2d18e 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7825,7 +7825,7 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: additional-settings: - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: image: vllm/vllm-openai:v0.27.1 model: nvidia/MiniMax-M3-NVFP4 From f9dcf6c6e821488206cd344c5842f5da44a47058 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 11:19:31 -0700 Subject: [PATCH 3/4] fix(minimaxm3): share backend model cache MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 为 MiniMax-M3 GB200 聚合式 AgentX 后端启用已挂载的 Hugging Face 缓存，避免投机解码草稿模型重复下载。 --- .../gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml | 3 +++ 1 file changed, 3 insertions(+) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml index 41a183508..3536914f9 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml @@ -39,6 +39,9 @@ backend: connector: null kv_events_config: { aggregated: true } aggregated_environment: + HF_HUB_CACHE: "/hf_hub_cache" + HUGGINGFACE_HUB_CACHE: "/hf_hub_cache" + TRANSFORMERS_CACHE: "/hf_hub_cache" VLLM_ENGINE_READY_TIMEOUT_S: "7200" VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800" VLLM_FLOAT32_MATMUL_PRECISION: "high" From d8d4221974b9b23049c3e2bc25bbdddc7396ad11 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 14:33:22 -0700 Subject: [PATCH 4/4] fix(minimaxm3): restore full aggregate sweep MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 恢复 MiniMax-M3 GB200 聚合 AgentX 的完整配置矩阵，使完整扫描覆盖所有已配置的并发点。 --- configs/nvidia-master.yaml | 60 ++++++++++++++++++++++++++++++++++++-- perf-changelog.yaml | 3 +- 2 files changed, 59 insertions(+), 4 deletions(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 0aac2d18e..0d5bd0c7d 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7796,8 +7796,8 @@ minimaxm3-fp4-b200-trtllm-agentic-mtp: search-space: - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [5, 10, 15, 20, 25, 30, 35, 40, 45] } - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [1, 5] } -# Retain the GB200 aggregate SimpleCPU-offload configuration at concurrency 20 -# and 30. +# Preserve the B200 TP4 search space and add the GB200 Pareto candidates found +# by direct DEP and P/D tuning. minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: image: vllm/vllm-openai:v0.27.1 model: nvidia/MiniMax-M3-NVFP4 @@ -7812,10 +7812,23 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: agentic-coding: - dram-utilization: 0.61 search-space: + - spec-decoding: mtp + kv-offloading: none + conc-list: [1, 2, 5, 8, 10, 12, 15, 20] + num-nodes: 1 + worker: + num-worker: 1 + tp: 4 + ep: 1 + dp-attn: false + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - spec-decoding: mtp kv-offloading: dram kv-offload-backend: { name: vllm-simple } - conc-list: [20, 30] + conc-list: [20, 30, 40] num-nodes: 1 worker: num-worker: 1 @@ -7826,6 +7839,47 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-tp4-vllm-simple-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" + - spec-decoding: mtp + kv-offloading: none + conc-list: [4, 32] + num-nodes: 1 + worker: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" + - spec-decoding: mtp + kv-offloading: dram + kv-offload-backend: { name: vllm-simple } + conc-list: [32, 40] + num-nodes: 1 + worker: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep4-vllm-simple-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" + - spec-decoding: mtp + kv-offloading: none + conc-list: [48] + num-nodes: 2 + worker: + num-worker: 1 + tp: 8 + ep: 8 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/agg-dep8-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" + minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: image: vllm/vllm-openai:v0.27.1 model: nvidia/MiniMax-M3-NVFP4 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 2edb0c148..8e6c32b6f 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6827,6 +6827,7 @@ scenario-type: - agentic-coding description: - - "Limit the GB200 aggregate Dynamo-vLLM configuration to the SimpleCPU-offload concurrency 20 and 30 recipes." + - "Refresh the full GB200 aggregate Dynamo-vLLM AgentX configuration." + - "Use the mounted Hugging Face cache for the TP4 SimpleCPU-offload backend." - "Use the committed MiniMax-M3 thinking-on EAGLE3-GQA acceptance target." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2806