Add B300 Dynamo+SGLang AgentX configs for DeepSeek V4 with aggregate DEP8 c384 / 添加含聚合式 DEP8 c384 的 DeepSeek V4 B300 Dynamo+SGLang AgentX 配置

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh

@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+python3 -m pip uninstall --break-system-packages -y mooncake-transfer-engine-cuda13 mooncake-transfer-engine-efa-cuda13
+python3 -m pip install --break-system-packages --no-deps mooncake-transfer-engine-efa-cuda13==0.3.13.post1

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-dep8-c384-mtp-kvoffload.yaml

@@ -0,0 +1,163 @@
+schema: 2
+name: "agg-b300-dep8-c384-mtp-kvoffload"
+
+# AgentX aggregate topology: one DEP8 worker (attention DP8 + EP8 Mega-MoE +
+# FP4 indexer) occupies one eight-GPU B300 node and serves both prefill and
+# decode with DSpark K=6 and a HiCache DRAM tier, tuned for concurrency 384.
+# Engine settings follow the single-node SGLang DEP8 c384 point
+# (benchmarks/single_node/srt-slurm-recipes/dsv4/sglang/b300-fp4-mtp/agentic.yaml,
+# override_dep8_c384). The Dynamo KV router replaces the SGLang router: each
+# DP rank publishes KV events, and sessions stay sticky via X-Dynamo-Session-ID.
+# Concurrency is exported by the master config.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+
+dynamo:
+  install: false
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+    # The 5 s default request-plane ack timeout expires while the worker
+    # ingests the c384 warmup burst; match the disaggregated recipes.
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_DEFAULT_THINKING: "1"
+      SGLANG_DSV4_REASONING_EFFORT: high
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      PYTHONNOUSERSITE: "1"
+      TORCH_CUDA_ARCH_LIST: "10.0"
+      # Triton compiles with the image's CUDA ptxas.
+      TRITON_PTXAS_PATH: /usr/local/cuda/bin/ptxas
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+      SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+      SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+      SGLANG_OPT_USE_JIT_NORM: "1"
+      SGLANG_OPT_USE_TOPK_V2: "True"
+      # Covers the 8192-token per-rank prefill budget (65536 / 8 DP ranks).
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: "8320"
+
+    args:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      watchdog-timeout: 1000000
+      allow-auto-truncate: true
+      attention-backend: dsv4
+      page-size: 256
+      disable-shared-experts-fusion: true
+      disable-flashinfer-autotune: true
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      enable-dp-attention-local-control-broadcast: true
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-prefill-delayer: true
+      prefill-decode-interval: 20
+      incremental-streaming-output: true
+      stream-interval: 20
+      # Mega-MoE's transient workspace sits outside the static pool.
+      mem-fraction-static: 0.88
+      swa-full-tokens-ratio: 0.075
+      chunked-prefill-size: 65536
+      max-running-requests: 768
+      # Per-rank graph limit: max-running-requests / dp-size = 768 / 8.
+      cuda-graph-max-bs-decode: 96
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      # HiCache capacity is a host/device ratio; 3 keeps the tier near 2 TB.
+      enable-hierarchical-cache: true
+      hicache-ratio: 3
+      hicache-write-policy: write_back
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+sbatch_directives:
+  mem: "0"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "false"
+    TP: "8"
+    EP_SIZE: "8"
+    DP_ATTENTION: "true"
+    PP_SIZE: "1"
+    PCP_SIZE: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    # Warmup can leave bytes unacknowledged past AIPerf's 30 s default.
+    AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c4-mtp.yaml

@@ -0,0 +1,128 @@
+schema: 2
+name: "agg-b300-tp4-c4-mtp"
+
+# AgentX aggregate topology: one TP4 worker occupies half of one eight-GPU
+# B300 node and serves both prefill and decode with DSpark K=6.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+
+dynamo:
+  install: false
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+    gpus: 4
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DEFAULT_THINKING: "1"
+      SGLANG_DSV4_REASONING_EFFORT: high
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+      SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+      SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+      SGLANG_OPT_USE_JIT_NORM: "1"
+      SGLANG_OPT_USE_TOPK_V2: "True"
+
+    args:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.90
+      page-size: 256
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      enable-deepseek-v4-fp4-indexer: true
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 8
+      cuda-graph-max-bs-decode: 8
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 4
+      ep-size: 1
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+
+sbatch_directives:
+  mem: "0"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "false"
+    TP: "4"
+    PP_SIZE: "1"
+    PCP_SIZE: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"