config: add GB200 DSpark AgentX recipes / 添加 GB200 DSpark AgentX 配置 (#3416)

* config: add GB200 DSpark AgentX recipes

新增 GB200 DSpark AgentX 配置，包含 C32 和 C64。

* chore: link GB200 DSpark changelog

为 GB200 DSpark 变更记录添加 PR 链接。

* config: include DSpark throughput runs

为 GB200 DSpark AgentX 配置添加吞吐量测试。

* config: refresh GB200 DSpark recipes

适配当前目录和启动脚本，并在 TP8 C1/C4/C8 使用 DSpark K6。

* chore: refresh PR #3416 for sweep reuse [skip-sweep]

---------

Co-authored-by: Ankur-singh <ankusingh@nvidia.com>

Author: xinli-sw

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb200-fp4/agentx/agg-tp8-c1-dspark6.yaml

@@ -0,0 +1,135 @@
+schema: 2
+name: dsv4-vllm-agg-gb200-tp8-c1-dspark6-agentic
+identity:
+  model:
+    revision: 72e1d3230f6c080a530b0a1d46f8eb4602340597
+  frameworks:
+    dynamo: 1.3.0.dev20260720
+dynamo:
+  install: true
+  source:
+    wheel: 1.3.0.dev20260720
+environment:
+  PYTHONHASHSEED: '0'
+  WITH_NVIDIA_PEERMEM: '0'
+  EVAL_MAX_MODEL_LEN: '1048576'
+  DYNAMO_WHEEL_DIRS: /srtctl-wheels
+  ETCD_LEASE_TTL: '7200'
+slurm:
+  time_limit: '2:00:00'
+health_check:
+  max_attempts: 540
+  interval_seconds: 10
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+services:
+- name: etcd
+  type: etcd
+  placement:
+    node: infra
+- name: nats
+  type: nats
+  placement:
+    node: infra
+  options:
+    max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  options:
+    store_config:
+      metadata_server: P2PHANDSHAKE
+      global_segment_size: 150GB
+      local_buffer_size: 4GB
+      protocol: rdma
+      device_name: "@fabric.rdma-devices"
+      mode: embedded
+      enable_offload: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: kv
+    router-temperature: 0
+    router-queue-threshold: 65536
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+    tokenizer: fastokens
+engine:
+  type: vllm
+  connector: null
+roles:
+  agg:
+    nodes: 2
+    workers: 1
+    gpus: 8
+    env:
+      HF_HUB_CACHE: /hf_hub_cache
+      HUGGINGFACE_HUB_CACHE: /hf_hub_cache
+      TRANSFORMERS_CACHE: /hf_hub_cache
+      VLLM_ENGINE_READY_TIMEOUT_S: '3600'
+      VLLM_RPC_TIMEOUT: '600000'
+      VLLM_LOG_STATS_INTERVAL: '1'
+      VLLM_V2_WARMUP_MAX_NUM_SEQS: '16'
+      TILELANG_CLEANUP_TEMP_FILES: '1'
+      VLLM_USE_NCCL_SYMM_MEM: '0'
+      TORCH_SYMMMEM: NVSHMEM
+      NCCL_CUMEM_ENABLE: '1'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_NVLS_ENABLE: '1'
+      VLLM_SERVER_DEV_MODE: '1'
+      VLLM_USE_V2_MODEL_RUNNER: '1'
+      VLLM_USE_RUST_FRONTEND: '1'
+      VLLM_ALLREDUCE_USE_FLASHINFER: '1'
+      VLLM_FLASHINFER_ALLREDUCE_BACKEND: auto
+      VLLM_MOONCAKE_LOAD_RECV_THREADS: '20'
+      VLLM_PREFIX_CACHE_RETENTION_INTERVAL: '32768'
+      VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: '1024'
+      VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: '1800'
+      UCX_MEMTYPE_CACHE: n
+      UCX_NET_DEVICES: "@fabric.rdma-device-ports"
+      UCX_TLS: rc,cuda_copy
+      NCCL_IB_HCA: "@fabric.rdma-devices"
+      NCCL_P2P_LEVEL: NVL
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      MC_STORE_CLIENT_METRIC: '1'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_TE_METRIC: '0'
+      DG_JIT_CACHE_DIR: /tmp/dg-cache-dsv4-gb200-dspark-agg-{job_id}
+      VLLM_USE_BREAKABLE_CUDAGRAPH: '0'
+    args:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      disable-custom-all-reduce: true
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
+      max-model-len: 1048576
+      max-num-seqs: 16
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      cudagraph-metrics: true
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[7,14,21,28,35,42,49,56,63,70,77,84,91,98,105,112],"mode":0}'
+      speculative-config: '{"method":"dspark","num_speculative_tokens":6,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      gpu-memory-utilization: 0.94
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: deepseek_v4
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+srun_options:
+  container-remap-root: ''
+  cpu-bind: none
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    PORT: '8000'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    AIPERF_DATASET_MMAP_CACHE_ENABLED: 'false'

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/vllm/gb200-fp4/agentx/agg-tp8-c4-dspark6.yaml

@@ -0,0 +1,135 @@
+schema: 2
+name: dsv4-vllm-agg-gb200-tp8-c4-dspark6-agentic
+identity:
+  model:
+    revision: 72e1d3230f6c080a530b0a1d46f8eb4602340597
+  frameworks:
+    dynamo: 1.3.0.dev20260720
+dynamo:
+  install: true
+  source:
+    wheel: 1.3.0.dev20260720
+environment:
+  PYTHONHASHSEED: '0'
+  WITH_NVIDIA_PEERMEM: '0'
+  EVAL_MAX_MODEL_LEN: '1048576'
+  DYNAMO_WHEEL_DIRS: /srtctl-wheels
+  ETCD_LEASE_TTL: '7200'
+slurm:
+  time_limit: '2:00:00'
+health_check:
+  max_attempts: 540
+  interval_seconds: 10
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+services:
+- name: etcd
+  type: etcd
+  placement:
+    node: infra
+- name: nats
+  type: nats
+  placement:
+    node: infra
+  options:
+    max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  options:
+    store_config:
+      metadata_server: P2PHANDSHAKE
+      global_segment_size: 150GB
+      local_buffer_size: 4GB
+      protocol: rdma
+      device_name: "@fabric.rdma-devices"
+      mode: embedded
+      enable_offload: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: kv
+    router-temperature: 0
+    router-queue-threshold: 65536
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+    tokenizer: fastokens
+engine:
+  type: vllm
+  connector: null
+roles:
+  agg:
+    nodes: 2
+    workers: 1
+    gpus: 8
+    env:
+      HF_HUB_CACHE: /hf_hub_cache
+      HUGGINGFACE_HUB_CACHE: /hf_hub_cache
+      TRANSFORMERS_CACHE: /hf_hub_cache
+      VLLM_ENGINE_READY_TIMEOUT_S: '3600'
+      VLLM_RPC_TIMEOUT: '600000'
+      VLLM_LOG_STATS_INTERVAL: '1'
+      VLLM_V2_WARMUP_MAX_NUM_SEQS: '16'
+      TILELANG_CLEANUP_TEMP_FILES: '1'
+      VLLM_USE_NCCL_SYMM_MEM: '0'
+      TORCH_SYMMMEM: NVSHMEM
+      NCCL_CUMEM_ENABLE: '1'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_NVLS_ENABLE: '1'
+      VLLM_SERVER_DEV_MODE: '1'
+      VLLM_USE_V2_MODEL_RUNNER: '1'
+      VLLM_USE_RUST_FRONTEND: '1'
+      VLLM_ALLREDUCE_USE_FLASHINFER: '1'
+      VLLM_FLASHINFER_ALLREDUCE_BACKEND: auto
+      VLLM_MOONCAKE_LOAD_RECV_THREADS: '20'
+      VLLM_PREFIX_CACHE_RETENTION_INTERVAL: '32768'
+      VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: '1024'
+      VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: '1800'
+      UCX_MEMTYPE_CACHE: n
+      UCX_NET_DEVICES: "@fabric.rdma-device-ports"
+      UCX_TLS: rc,cuda_copy
+      NCCL_IB_HCA: "@fabric.rdma-devices"
+      NCCL_P2P_LEVEL: NVL
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      MC_STORE_CLIENT_METRIC: '1'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_TE_METRIC: '0'
+      DG_JIT_CACHE_DIR: /tmp/dg-cache-dsv4-gb200-dspark-agg-{job_id}
+      VLLM_USE_BREAKABLE_CUDAGRAPH: '0'
+    args:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      disable-custom-all-reduce: true
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"indexer_kv_dtype":"mxfp4"}'
+      max-model-len: 1048576
+      max-num-seqs: 16
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      cudagraph-metrics: true
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[7,14,21,28,35,42,49,56,63,70,77,84,91,98,105,112],"mode":0}'
+      speculative-config: '{"method":"dspark","num_speculative_tokens":6,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      gpu-memory-utilization: 0.94
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: deepseek_v4
+sbatch_directives:
+  cpus-per-task: '144'
+  mem: '0'
+srun_options:
+  container-remap-root: ''
+  cpu-bind: none
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    PORT: '8000'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    AIPERF_DATASET_MMAP_CACHE_ENABLED: 'false'