From 15cb539dafe32a2b8eed76108ff3d70f5579a316 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 10:19:51 -0700 Subject: [PATCH 1/3] config(minimaxm3): add GB200 disaggregated recipes MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 添加 MiniMax-M3 GB200 Dynamo-vLLM 解耦配方，配置 Mooncake 主机 KV 存储、NIXL UCX 传输和已提交的 thinking-on EAGLE3-GQA 接受长度。 --- .../disagg-1p1d-tp4-tp4-c24-agentic.yaml | 159 +++++++++++++++++ .../disagg-1p1d-tp4-tp8-c1-agentic.yaml | 162 ++++++++++++++++++ .../disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml | 159 +++++++++++++++++ configs/nvidia-master.yaml | 52 ++++-- perf-changelog.yaml | 9 + 5 files changed, 530 insertions(+), 11 deletions(-) create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml new file mode 100644 index 0000000000..b2df1bdc18 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml @@ -0,0 +1,159 @@ +name: "minimax-m3-vllm-disagg-gb200-1p1d-tp4-tp4-c24-agentic" + +model: + path: "minimax-m3-nvfp4" + container: &container "vllm/vllm-openai@sha256:b9104b7ef3048e42f79fba9ab5da06e5aff8164aca9968692ec2f569aaaf34c6" + precision: "fp4" + +identity: + model: { repo: "nvidia/MiniMax-M3-NVFP4" } + container: { image: *container } + frameworks: { dynamo: "1.5.0.dev20260819" } + +dynamo: { wheel: "1.5.0.dev20260819", install: true } +environment: { PYTHONHASHSEED: "0" } +health_check: { max_attempts: 2160, interval_seconds: 10 } + +resources: + gpu_type: "gb200" + gpus_per_node: 4 + prefill_nodes: 1 + decode_nodes: 1 + prefill_workers: 1 + decode_workers: 1 + gpus_per_prefill: 4 + gpus_per_decode: 4 + +infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 } + +frontend: + type: dynamo + enable_multiple_frontends: false + args: + dyn-chat-processor: "vllm" + trust-remote-code: true + tool-call-parser: "minimax_m3" + reasoning-parser: "minimax_m3" + enable-auto-tool-choice: true + router-mode: "least-loaded" + router-session-affinity-ttl-secs: 3600 + env: + DYN_LOG: "info" + DYN_TCP_CONNECT_TIMEOUT: "120" + DYN_TOKENIZER: "fastokens" + DYN_TOKENIZER_CACHE: "1" + +backend: + type: vllm + connector: null + dp_launch_mode: per_gpu + mooncake_kv_store: + env: + MC_ENABLE_DEST_DEVICE_AFFINITY: "1" + MC_STORE_CLIENT_METRIC: "1" + MC_STORE_CLIENT_METRIC_INTERVAL: "5" + MC_TE_METRIC: "1" + store_config: + metadata_server: "P2PHANDSHAKE" + global_segment_size: "150GB" + local_buffer_size: "4GB" + protocol: "rdma" + device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3" + mode: "embedded" + enable_offload: true + prefill_environment: &worker_environment + HF_HOME: "/hf_hub_cache" + HF_HUB_CACHE: "/hf_hub_cache" + HUGGINGFACE_HUB_CACHE: "/hf_hub_cache" + TRANSFORMERS_CACHE: "/hf_hub_cache" + DYN_TCP_CONNECT_TIMEOUT: "120" + VLLM_ENGINE_READY_TIMEOUT_S: "3600" + VLLM_FLOAT32_MATMUL_PRECISION: "high" + VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm" + VLLM_MOONCAKE_LOAD_RECV_THREADS: "20" + VLLM_USE_NCCL_SYMM_MEM: "0" + VLLM_ALLREDUCE_USE_SYMM_MEM: "0" + NCCL_CUMEM_ENABLE: "1" + NCCL_MNNVL_ENABLE: "1" + NCCL_NVLS_ENABLE: "1" + NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3" + UCX_CUDA_IPC_ENABLE_MNNVL: "y" + UCX_MEMTYPE_CACHE: "n" + UCX_MODULE_DIR: "/usr/local/lib/python3.12/dist-packages/nixl_cu13.libs/ucx" + UCX_RCACHE_MAX_UNRELEASED: "1024" + UCX_RNDV_PIPELINE_ERROR_HANDLING: "y" + UCX_TLS: "tcp,cuda_ipc,cuda_copy" + WITH_NVIDIA_PEERMEM: "0" + decode_environment: *worker_environment + vllm_config: + prefill: + served-model-name: "nvidia/MiniMax-M3-NVFP4" + tensor-parallel-size: 4 + enable-expert-parallel: true + all2all-backend: "flashinfer_nvlink_one_sided" + trust-remote-code: true + enable-prefix-caching: true + language-model-only: true + kv-cache-dtype: "fp8" + block-size: 128 + gpu-memory-utilization: 0.9 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 512 + stream-interval: 20 + no-enable-flashinfer-autotune: true + enable-cumem-allocator: true + numa-bind: true + numa-bind-nodes: &numa_nodes [0, 0, 1, 1] + attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"lookup_async":false,"enable_lookup":false}}]}}' + reasoning-parser: "minimax_m3" + dyn-tool-call-parser: "minimax_m3" + dyn-reasoning-parser: "minimax_m3" + decode: + served-model-name: "nvidia/MiniMax-M3-NVFP4" + tensor-parallel-size: 4 + enable-expert-parallel: true + all2all-backend: "flashinfer_nvlink_one_sided" + trust-remote-code: true + enable-prefix-caching: true + language-model-only: true + kv-cache-dtype: "fp8" + block-size: 128 + gpu-memory-utilization: 0.9 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 512 + stream-interval: 20 + no-enable-flashinfer-autotune: true + enable-cumem-allocator: true + numa-bind: true + numa-bind-nodes: *numa_nodes + attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"enable_lookup":false}}]}}' + reasoning-parser: "minimax_m3" + dyn-tool-call-parser: "minimax_m3" + dyn-reasoning-parser: "minimax_m3" + dyn-default-thinking-mode: "enabled" + +sbatch_directives: { cpus-per-task: "144", mem: "0" } +srun_options: { container-remap-root: "" } + +benchmark: + type: custom + command: "bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh" + env: + INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace" + RESULT_DIR: "/logs/agentic" + PORT: "8000" + IS_MULTINODE: "true" + AGENTIC_WARMUP_GRACE_PERIOD: "1800" + AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" + AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" + AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "3600" + AIPERF_EXTRA_INPUTS: "thinking:true" + AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:" + AIPERF_SERVER_METRICS_COLLECTION_INTERVAL: "1.0" + AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache" + HF_HUB_CACHE: "/hf_hub_cache" + WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml new file mode 100644 index 0000000000..b99e681677 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml @@ -0,0 +1,162 @@ +name: "minimax-m3-vllm-disagg-gb200-1p1d-tp4-tp8-c1-agentic" + +model: + path: "minimax-m3-nvfp4" + container: &container "vllm/vllm-openai@sha256:b9104b7ef3048e42f79fba9ab5da06e5aff8164aca9968692ec2f569aaaf34c6" + precision: "fp4" + +identity: + model: { repo: "nvidia/MiniMax-M3-NVFP4" } + container: { image: *container } + frameworks: { dynamo: "1.5.0.dev20260819" } + +dynamo: { wheel: "1.5.0.dev20260819", install: true } +environment: { PYTHONHASHSEED: "0" } +health_check: { max_attempts: 2160, interval_seconds: 10 } + +resources: + gpu_type: "gb200" + gpus_per_node: 4 + prefill_nodes: 1 + decode_nodes: 2 + prefill_workers: 1 + decode_workers: 1 + gpus_per_prefill: 4 + gpus_per_decode: 8 + +infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 } + +frontend: + type: dynamo + enable_multiple_frontends: false + args: + dyn-chat-processor: "vllm" + trust-remote-code: true + tool-call-parser: "minimax_m3" + reasoning-parser: "minimax_m3" + enable-auto-tool-choice: true + router-mode: "least-loaded" + router-session-affinity-ttl-secs: 3600 + env: + DYN_LOG: "info" + DYN_TCP_CONNECT_TIMEOUT: "120" + DYN_TOKENIZER: "fastokens" + DYN_TOKENIZER_CACHE: "1" + +backend: + type: vllm + connector: null + dp_launch_mode: per_gpu + mooncake_kv_store: + env: + MC_ENABLE_DEST_DEVICE_AFFINITY: "1" + MC_STORE_CLIENT_METRIC: "1" + MC_STORE_CLIENT_METRIC_INTERVAL: "5" + MC_TE_METRIC: "1" + store_config: + metadata_server: "P2PHANDSHAKE" + global_segment_size: "150GB" + local_buffer_size: "4GB" + protocol: "rdma" + device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3" + mode: "embedded" + enable_offload: true + prefill_environment: &worker_environment + HF_HOME: "/hf_hub_cache" + HF_HUB_CACHE: "/hf_hub_cache" + HUGGINGFACE_HUB_CACHE: "/hf_hub_cache" + TRANSFORMERS_CACHE: "/hf_hub_cache" + DYN_TCP_CONNECT_TIMEOUT: "120" + VLLM_ENGINE_READY_TIMEOUT_S: "3600" + VLLM_FLOAT32_MATMUL_PRECISION: "high" + VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm" + VLLM_MOONCAKE_LOAD_RECV_THREADS: "20" + VLLM_USE_NCCL_SYMM_MEM: "0" + VLLM_ALLREDUCE_USE_SYMM_MEM: "0" + NCCL_CUMEM_ENABLE: "1" + NCCL_MNNVL_ENABLE: "1" + NCCL_NVLS_ENABLE: "1" + NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3" + UCX_CUDA_IPC_ENABLE_MNNVL: "y" + UCX_MEMTYPE_CACHE: "n" + UCX_MODULE_DIR: "/usr/local/lib/python3.12/dist-packages/nixl_cu13.libs/ucx" + UCX_RCACHE_MAX_UNRELEASED: "1024" + UCX_RNDV_PIPELINE_ERROR_HANDLING: "y" + UCX_TLS: "tcp,cuda_ipc,cuda_copy" + WITH_NVIDIA_PEERMEM: "0" + decode_environment: + <<: *worker_environment + VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl" + vllm_config: + prefill: + served-model-name: "nvidia/MiniMax-M3-NVFP4" + tensor-parallel-size: 4 + enable-expert-parallel: true + all2all-backend: "flashinfer_nvlink_one_sided" + trust-remote-code: true + enable-prefix-caching: true + language-model-only: true + kv-cache-dtype: "fp8" + block-size: 128 + gpu-memory-utilization: 0.9 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 512 + stream-interval: 20 + no-enable-flashinfer-autotune: true + enable-cumem-allocator: true + numa-bind: true + numa-bind-nodes: &numa_nodes [0, 0, 1, 1] + attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"lookup_async":false,"enable_lookup":false}}]}}' + reasoning-parser: "minimax_m3" + dyn-tool-call-parser: "minimax_m3" + dyn-reasoning-parser: "minimax_m3" + decode: + served-model-name: "nvidia/MiniMax-M3-NVFP4" + tensor-parallel-size: 8 + enable-expert-parallel: true + all2all-backend: "flashinfer_nvlink_one_sided" + trust-remote-code: true + enable-prefix-caching: true + language-model-only: true + kv-cache-dtype: "fp8" + block-size: 128 + gpu-memory-utilization: 0.9 + max-num-batched-tokens: 16384 + max-num-seqs: 1 + max-cudagraph-capture-size: 512 + stream-interval: 20 + no-enable-flashinfer-autotune: true + enable-cumem-allocator: true + numa-bind: true + numa-bind-nodes: *numa_nodes + attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"enable_lookup":false}}]}}' + reasoning-parser: "minimax_m3" + dyn-tool-call-parser: "minimax_m3" + dyn-reasoning-parser: "minimax_m3" + dyn-default-thinking-mode: "enabled" + +sbatch_directives: { cpus-per-task: "144", mem: "0" } +srun_options: { container-remap-root: "" } + +benchmark: + type: custom + command: "bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh" + env: + INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace" + RESULT_DIR: "/logs/agentic" + PORT: "8000" + IS_MULTINODE: "true" + AGENTIC_WARMUP_GRACE_PERIOD: "1800" + AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" + AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" + AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "3600" + AIPERF_EXTRA_INPUTS: "thinking:true" + AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:" + AIPERF_SERVER_METRICS_COLLECTION_INTERVAL: "1.0" + AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache" + HF_HUB_CACHE: "/hf_hub_cache" + WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml new file mode 100644 index 0000000000..dac8ae85b9 --- /dev/null +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml @@ -0,0 +1,159 @@ +name: "minimax-m3-vllm-disagg-gb200-1p2d-tp4-tp4-c8-c16-agentic" + +model: + path: "minimax-m3-nvfp4" + container: &container "vllm/vllm-openai@sha256:b9104b7ef3048e42f79fba9ab5da06e5aff8164aca9968692ec2f569aaaf34c6" + precision: "fp4" + +identity: + model: { repo: "nvidia/MiniMax-M3-NVFP4" } + container: { image: *container } + frameworks: { dynamo: "1.5.0.dev20260819" } + +dynamo: { wheel: "1.5.0.dev20260819", install: true } +environment: { PYTHONHASHSEED: "0" } +health_check: { max_attempts: 2160, interval_seconds: 10 } + +resources: + gpu_type: "gb200" + gpus_per_node: 4 + prefill_nodes: 1 + decode_nodes: 2 + prefill_workers: 1 + decode_workers: 2 + gpus_per_prefill: 4 + gpus_per_decode: 4 + +infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 } + +frontend: + type: dynamo + enable_multiple_frontends: false + args: + dyn-chat-processor: "vllm" + trust-remote-code: true + tool-call-parser: "minimax_m3" + reasoning-parser: "minimax_m3" + enable-auto-tool-choice: true + router-mode: "least-loaded" + router-session-affinity-ttl-secs: 3600 + env: + DYN_LOG: "info" + DYN_TCP_CONNECT_TIMEOUT: "120" + DYN_TOKENIZER: "fastokens" + DYN_TOKENIZER_CACHE: "1" + +backend: + type: vllm + connector: null + dp_launch_mode: per_gpu + mooncake_kv_store: + env: + MC_ENABLE_DEST_DEVICE_AFFINITY: "1" + MC_STORE_CLIENT_METRIC: "1" + MC_STORE_CLIENT_METRIC_INTERVAL: "5" + MC_TE_METRIC: "1" + store_config: + metadata_server: "P2PHANDSHAKE" + global_segment_size: "150GB" + local_buffer_size: "4GB" + protocol: "rdma" + device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3" + mode: "embedded" + enable_offload: true + prefill_environment: &worker_environment + HF_HOME: "/hf_hub_cache" + HF_HUB_CACHE: "/hf_hub_cache" + HUGGINGFACE_HUB_CACHE: "/hf_hub_cache" + TRANSFORMERS_CACHE: "/hf_hub_cache" + DYN_TCP_CONNECT_TIMEOUT: "120" + VLLM_ENGINE_READY_TIMEOUT_S: "3600" + VLLM_FLOAT32_MATMUL_PRECISION: "high" + VLLM_FLASHINFER_ALLREDUCE_BACKEND: "trtllm" + VLLM_MOONCAKE_LOAD_RECV_THREADS: "20" + VLLM_USE_NCCL_SYMM_MEM: "0" + VLLM_ALLREDUCE_USE_SYMM_MEM: "0" + NCCL_CUMEM_ENABLE: "1" + NCCL_MNNVL_ENABLE: "1" + NCCL_NVLS_ENABLE: "1" + NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3" + UCX_CUDA_IPC_ENABLE_MNNVL: "y" + UCX_MEMTYPE_CACHE: "n" + UCX_MODULE_DIR: "/usr/local/lib/python3.12/dist-packages/nixl_cu13.libs/ucx" + UCX_RCACHE_MAX_UNRELEASED: "1024" + UCX_RNDV_PIPELINE_ERROR_HANDLING: "y" + UCX_TLS: "tcp,cuda_ipc,cuda_copy" + WITH_NVIDIA_PEERMEM: "0" + decode_environment: *worker_environment + vllm_config: + prefill: + served-model-name: "nvidia/MiniMax-M3-NVFP4" + tensor-parallel-size: 4 + enable-expert-parallel: true + all2all-backend: "flashinfer_nvlink_one_sided" + trust-remote-code: true + enable-prefix-caching: true + language-model-only: true + kv-cache-dtype: "fp8" + block-size: 128 + gpu-memory-utilization: 0.9 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 512 + stream-interval: 20 + no-enable-flashinfer-autotune: true + enable-cumem-allocator: true + numa-bind: true + numa-bind-nodes: &numa_nodes [0, 0, 1, 1] + attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"lookup_async":false,"enable_lookup":false}}]}}' + reasoning-parser: "minimax_m3" + dyn-tool-call-parser: "minimax_m3" + dyn-reasoning-parser: "minimax_m3" + decode: + served-model-name: "nvidia/MiniMax-M3-NVFP4" + tensor-parallel-size: 4 + enable-expert-parallel: true + all2all-backend: "flashinfer_nvlink_one_sided" + trust-remote-code: true + enable-prefix-caching: true + language-model-only: true + kv-cache-dtype: "fp8" + block-size: 128 + gpu-memory-utilization: 0.9 + max-num-batched-tokens: 16384 + max-cudagraph-capture-size: 512 + stream-interval: 20 + no-enable-flashinfer-autotune: true + enable-cumem-allocator: true + numa-bind: true + numa-bind-nodes: *numa_nodes + attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"enable_lookup":false}}]}}' + reasoning-parser: "minimax_m3" + dyn-tool-call-parser: "minimax_m3" + dyn-reasoning-parser: "minimax_m3" + dyn-default-thinking-mode: "enabled" + +sbatch_directives: { cpus-per-task: "144", mem: "0" } +srun_options: { container-remap-root: "" } + +benchmark: + type: custom + command: "bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh" + env: + INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace" + RESULT_DIR: "/logs/agentic" + PORT: "8000" + IS_MULTINODE: "true" + AGENTIC_WARMUP_GRACE_PERIOD: "1800" + AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true" + AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0" + AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "3600" + AIPERF_EXTRA_INPUTS: "thinking:true" + AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:" + AIPERF_SERVER_METRICS_COLLECTION_INTERVAL: "1.0" + AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache" + HF_HUB_CACHE: "/hf_hub_cache" + WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126" diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 0d5bd0c7d3..8018732cef 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7881,33 +7881,63 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-agg-mtp: - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: - image: vllm/vllm-openai:v0.27.1 + image: vllm/vllm-openai@sha256:b9104b7ef3048e42f79fba9ab5da06e5aff8164aca9968692ec2f569aaaf34c6 model: nvidia/MiniMax-M3-NVFP4 model-prefix: minimaxm3 runner: cluster:gb200-nv precision: fp4 framework: dynamo-vllm - router: { name: dynamo-router, version: "1.3.1" } + router: { name: dynamo-router, version: "1.5.0.dev20260819" } kv-p2p-transfer: nixl multinode: true disagg: true scenarios: agentic-coding: - - dram-utilization: 0.61 + # Each active four-rank prefill node reserves 150 GB/rank in Mooncake. + - dram-utilization: 0.665 search-space: - spec-decoding: mtp - kv-offloading: none - conc-list: [48] + kv-offloading: dram + kv-offload-backend: { name: mooncake, version: "0.3.12.post1" } + conc-list: [1] prefill: num-worker: 1 - tp: 8 - ep: 8 - dp-attn: true + tp: 4 + ep: 4 + dp-attn: false additional-settings: - - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml" + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" - decode: { num-worker: 1, tp: 4, ep: 4, dp-attn: true } + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + decode: { num-worker: 1, tp: 8, ep: 8, dp-attn: false } + - spec-decoding: mtp + kv-offloading: dram + kv-offload-backend: { name: mooncake, version: "0.3.12.post1" } + conc-list: [8, 16] + prefill: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: false + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + decode: { num-worker: 2, tp: 4, ep: 4, dp-attn: false } + - spec-decoding: mtp + kv-offloading: dram + kv-offload-backend: { name: mooncake, version: "0.3.12.post1" } + conc-list: [24] + prefill: + num-worker: 1 + tp: 4 + ep: 4 + dp-attn: false + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + decode: { num-worker: 1, tp: 4, ep: 4, dp-attn: false } dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-agg: image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 4eb0a2caa6..4b1faa237e 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-disagg-mtp + scenario-type: + - agentic-coding + description: + - "Add GB200 Dynamo-vLLM disaggregated AgentX recipes with TP4 prefill and TP4 or TP8 decode workers." + - "Configure Mooncake host KV storage, NIXL over UCX, Dynamo 1.5.0.dev20260819, explicit thinking mode, and the committed MiniMax-M3 EAGLE3-GQA acceptance target." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2807 From 9581d3388cfd4408c396a9646f1744b71481b7a9 Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 10:49:18 -0700 Subject: [PATCH 2/3] fix(minimaxm3): align disaggregated GQA acceptance target MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 将 MiniMax-M3 GB200 分离式 AgentX 配置与仓库中已提交的 EAGLE3-GQA thinking-on 接受长度保持一致。 --- .../gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml | 4 ++-- .../gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml | 4 ++-- .../agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml | 4 ++-- configs/nvidia-master.yaml | 6 +++--- 4 files changed, 9 insertions(+), 9 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml index b2df1bdc18..19c6748e90 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml @@ -105,7 +105,7 @@ backend: numa-bind: true numa-bind-nodes: &numa_nodes [0, 0, 1, 1] attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}' - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"lookup_async":false,"enable_lookup":false}}]}}' reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" @@ -129,7 +129,7 @@ backend: numa-bind: true numa-bind-nodes: *numa_nodes attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}' - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"enable_lookup":false}}]}}' reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml index b99e681677..7ab2a4506a 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml @@ -107,7 +107,7 @@ backend: numa-bind: true numa-bind-nodes: &numa_nodes [0, 0, 1, 1] attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}' - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"lookup_async":false,"enable_lookup":false}}]}}' reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" @@ -132,7 +132,7 @@ backend: numa-bind: true numa-bind-nodes: *numa_nodes attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}' - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"enable_lookup":false}}]}}' reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml index dac8ae85b9..562a44fcf2 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml @@ -105,7 +105,7 @@ backend: numa-bind: true numa-bind-nodes: &numa_nodes [0, 0, 1, 1] attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}' - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"lookup_async":false,"enable_lookup":false}}]}}' reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" @@ -129,7 +129,7 @@ backend: numa-bind: true numa-bind-nodes: *numa_nodes attention-config: '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8","minimax_m3_msa_decode_backend":"cutlass"}' - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.83}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_connector_extra_config":{"backends":["UCX"],"num_threads":8}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_consumer","kv_connector_extra_config":{"enable_lookup":false}}]}}' reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 8018732cef..1c8fd72207 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7908,7 +7908,7 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: additional-settings: - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp8-c1-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" decode: { num-worker: 1, tp: 8, ep: 8, dp-attn: false } - spec-decoding: mtp kv-offloading: dram @@ -7922,7 +7922,7 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: additional-settings: - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p2d-tp4-tp4-c8-c16-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" decode: { num-worker: 2, tp: 4, ep: 4, dp-attn: false } - spec-decoding: mtp kv-offloading: dram @@ -7936,7 +7936,7 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: additional-settings: - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-tp4-tp4-c24-agentic.yaml" - "SYNTHETIC_ACCEPTANCE=true" - - "SYNTHETIC_ACCEPTANCE_LENGTH=2.83" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" decode: { num-worker: 1, tp: 4, ep: 4, dp-attn: false } dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-agg: From 38157cded915d881c5a4181b08bb78db836b9e2e Mon Sep 17 00:00:00 2001 From: Rohit Pujar Nagraj Date: Wed, 2 Sep 2026 14:44:55 -0700 Subject: [PATCH 3/3] fix(minimaxm3): restore full disaggregated sweep MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 恢复 MiniMax-M3 GB200 解耦 AgentX 的完整配置矩阵，使完整扫描覆盖现有及新增的所有拓扑点。 --- .../disagg-1p1d-dep8-dep4-agentic.yaml | 22 +++++++++++++------ configs/nvidia-master.yaml | 15 +++++++++++++ perf-changelog.yaml | 2 +- 3 files changed, 31 insertions(+), 8 deletions(-) diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml index 1a798085b9..80b5dc4e0e 100644 --- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml +++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml @@ -1,13 +1,15 @@ name: "minimax-m3-vllm-disagg-gb200-1p1d-dep8-dep4-agentic" -model: { path: "minimax-m3-nvfp4", container: "vllm/vllm-openai:v0.27.1", precision: "fp4" } +model: + path: "minimax-m3-nvfp4" + container: &container "vllm/vllm-openai@sha256:b9104b7ef3048e42f79fba9ab5da06e5aff8164aca9968692ec2f569aaaf34c6" + precision: "fp4" identity: model: { repo: "nvidia/MiniMax-M3-NVFP4" } - container: { image: "vllm/vllm-openai:v0.27.1" } - frameworks: { dynamo: "1.3.1" } -dynamo: { version: "1.3.1", install: true } + container: { image: *container } + frameworks: { dynamo: "1.5.0.dev20260819" } +dynamo: { wheel: "1.5.0.dev20260819", install: true } environment: { ETCD_LEASE_TTL: "7200" } -slurm: { time_limit: "12:00:00" } health_check: { max_attempts: 2160, interval_seconds: 10 } resources: gpu_type: "gb200" @@ -41,6 +43,10 @@ backend: dp_launch_mode: per_node kv_events_config: true prefill_environment: &worker_env + HF_HOME: "/hf_hub_cache" + 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_NIXL_ABORT_REQUEST_TIMEOUT: "300" @@ -73,7 +79,7 @@ backend: max-model-len: 1048576 language-model-only: true kv-cache-dtype: "fp8" - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' compilation-config: '{"cudagraph_mode":"PIECEWISE"}' stream-interval: 20 max-cudagraph-capture-size: 512 @@ -99,7 +105,7 @@ backend: max-model-len: 1048576 language-model-only: true kv-cache-dtype: "fp8" - speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}' + speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":2.78}' compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY"}' stream-interval: 20 max-cudagraph-capture-size: 512 @@ -108,6 +114,7 @@ backend: reasoning-parser: "minimax_m3" dyn-tool-call-parser: "minimax_m3" dyn-reasoning-parser: "minimax_m3" + dyn-default-thinking-mode: "enabled" sbatch_directives: { cpus-per-task: "144", mem: "0" } srun_options: { container-remap-root: "" } benchmark: @@ -123,6 +130,7 @@ benchmark: AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400" AIPERF_EXTRA_INPUTS: "thinking:true" AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:" + AIPERF_SERVER_METRICS_COLLECTION_INTERVAL: "1.0" AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache" HF_HUB_CACHE: "/hf_hub_cache" WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126" diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 1c8fd72207..35ab17a1da 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7893,6 +7893,21 @@ minimaxm3-fp4-gb200-dynamo-vllm-agentic-disagg-mtp: disagg: true scenarios: agentic-coding: + - dram-utilization: 0.61 + search-space: + - spec-decoding: mtp + kv-offloading: none + conc-list: [48] + prefill: + num-worker: 1 + tp: 8 + ep: 8 + dp-attn: true + additional-settings: + - "CONFIG_FILE=recipes/vllm/minimax-m3/gb200-fp4/agentic/disagg-1p1d-dep8-dep4-agentic.yaml" + - "SYNTHETIC_ACCEPTANCE=true" + - "SYNTHETIC_ACCEPTANCE_LENGTH=2.78" + decode: { num-worker: 1, tp: 4, ep: 4, dp-attn: true } # Each active four-rank prefill node reserves 150 GB/rank in Mooncake. - dram-utilization: 0.665 search-space: diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 4b1faa237e..d724c59fcc 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6827,6 +6827,6 @@ scenario-type: - agentic-coding description: - - "Add GB200 Dynamo-vLLM disaggregated AgentX recipes with TP4 prefill and TP4 or TP8 decode workers." + - "Refresh the full GB200 Dynamo-vLLM disaggregated AgentX configuration across the configured topology points." - "Configure Mooncake host KV storage, NIXL over UCX, Dynamo 1.5.0.dev20260819, explicit thinking mode, and the committed MiniMax-M3 EAGLE3-GQA acceptance target." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2807