From 80eb81acac90ee565ed0f2c45692e38e526ad7f3 Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Tue, 18 Aug 2026 14:41:39 +0800
Subject: [PATCH 1/6] perf(agentx): tune SGLang low-concurrency throughput
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add TP8 concurrency 1/4 and TP4 concurrency 8 aggregate profiles with phase-specific decode CUDA graph limits and throughput-only synthetic acceptance.

中文：优化 SGLang AgentX 低并发吞吐，新增 TP8 并发 1/4 和 TP4 并发 8 的聚合配置，并使用解码阶段 CUDA Graph 上限及仅吞吐测试启用的模拟验收率。
---
 .../agentic/agg-gb300-tp4-mtp-kvoffload.yaml  | 156 ------------------
 .../agentic/agg-gb300-tp4-mtp-lowlatency.yaml | 115 +++++++++++++
 .../agentic/agg-gb300-tp8-mtp-lowlatency.yaml | 115 +++++++++++++
 configs/nvidia-master.yaml                    |  35 ++--
 4 files changed, 254 insertions(+), 167 deletions(-)
 delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml
deleted file mode 100644
index 7bb82d2b5b..0000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml
+++ /dev/null
@@ -1,156 +0,0 @@
-name: "agg-gb300-tp4-mtp-kvoffload"
-
-# Agentic-coding SGLang aggregated recipe for DeepSeek-V4-Pro on GB300
-# (single aggregated worker, TP4, MTP + hierarchical-cache KV offload).
-#
-# Uses the flat single-variant schema the agentic CI flow expects. Concurrency is NOT
-# a recipe field here: the GHA matrix fans out one job per concurrency from
-# the master-config conc-list, exporting CONC into agentic_srt.sh. Modeled on
-# the flat vllm/deepseek-v4/agentic recipes (agentic infra + benchmark wiring)
-# and the flat sglang/deepseek-v4/8k1k recipes (sglang backend schema).
-
-model:
-  path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
-  precision: "fp4"
-
-dynamo:
-  install: true
-  wheel: "1.3.0.dev20260718"
-
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 1440
-  interval_seconds: 10
-
-resources:
-  gpu_type: "gb300"
-  gpus_per_node: 4
-  agg_nodes: 1
-  agg_workers: 1
-  gpus_per_agg: 4
-
-infra:
-  # Dedicated etcd/nats node — matches the vllm agentic recipes; under the
-  # dynamo router + multiple frontends the infra services need their own node.
-  etcd_nats_dedicated_node: true
-  # cc-traces prompts (~125K tokens) serialize to ~2.8MB; the 1 MiB NATS
-  # default drops them. 32 MiB matches the agentic vllm recipes (~10x headroom
-  # over the largest observed payload); schema recommends 24+ for long ISL.
-  nats_max_payload_mb: 32
-
-frontend:
-  type: dynamo
-  nginx_session_affinity: true
-  nginx_session_affinity_header: X-Dynamo-Session-ID
-  enable_multiple_frontends: true
-  num_additional_frontends: 4
-  env:
-    # The sglang image is PEP 668 externally-managed; the runtime dynamo
-    # install (dynamo_wheels.py / source build) runs pip and fails with
-    # "externally-managed-environment" without this. Lets pip install into
-    # the system env. Applies to both dynamo.hash (source) and dynamo.wheel.
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-  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"
-
-backend:
-  type: sglang
-
-  aggregated_environment:
-    SGLANG_DEFAULT_THINKING: '1'
-    # TP4/bs128 can spend over an hour capturing the DSV4 EAGLE draft-extend
-    # graph. Keep the target/decode graphs and use the upstream eager fallback.
-    SGLANG_DISABLE_DRAFT_EXTEND_CUDA_GRAPH: "1"
-    SGLANG_DSV4_REASONING_EFFORT: high
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
-    # Lets the runtime dynamo install's pip bypass PEP 668 on the sglang image.
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-    # from submission for B300
-    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "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"
-    # for kvcache offload
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
-
-  sglang_config:
-    aggregated:
-      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
-      enable-metrics: true
-      trust-remote-code: true
-      stream-interval: 50
-      watchdog-timeout: 1000000
-      mem-fraction-static: 0.90
-      chunked-prefill-size: 8192  # can only support 11000 for TP
-      moe-runner-backend: "flashinfer_mxfp4"
-      disable-flashinfer-autotune: true
-      swa-full-tokens-ratio: 0.1
-      max-running-requests: 128
-      cuda-graph-max-bs: 128
-
-      scheduler-recv-interval: 30
-      dp-size: 1
-      tp-size: 4
-      ep-size: 1
-      # kv-cache offload
-      enable-hierarchical-cache: true
-      hicache-ratio: 6
-      hicache-write-policy: write_through
-      hicache-io-backend: kernel
-      # mtp section
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
-      speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
-
-sbatch_directives:
-  mem: "0"
-  cpus-per-task: "144"
-
-srun_options:
-  mem: "0"
-  # gb300-nv: pyxis maps the calling user into the container as a non-root
-  # uid; agentic_srt.sh's apt-get install git step needs EUID 0. Remap to
-  # uid 0 inside the container. srt-slurm renders empty-string values as
-  # flag-only srun args.
-  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"
-    # Aggregated: one worker serves both prefill and decode, so GPU accounting
-    # is the single-worker form num_gpus = TP (not the disagg prefill+decode
-    # sum). The multinode post-processing path assumes disagg and would divide
-    # throughput by prefill_gpus + decode_gpus; force the single-node code path
-    # in process_agentic_result.py instead. TP must match sglang_config tp-size.
-    IS_MULTINODE: "false"
-    TP: "4"
-    # Stamp X-Dynamo-Session-ID on every turn so both nginx and Dynamo's
-    # router keep the session on the worker that owns its KV prefix.
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
-    # Container-side path of the aiperf mmap dataset cache; host-side mount
-    # wired via launch_gb300-*.sh srtslurm.yaml default_mounts. Without it,
-    # aiperf re-tokenizes + re-writes the dataset mmap on every run.
-    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
-    # Persistent HF hub cache (also via default_mounts) so the trace dataset
-    # isn't re-downloaded each run; overrides the workflow-level HF_HUB_CACHE.
-    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
new file mode 100644
index 0000000000..300080dca5
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
@@ -0,0 +1,115 @@
+name: "agg-gb300-tp4-mtp-lowlatency"
+
+# Low-latency AgentX aggregate topology: one TP4 worker occupies one
+# four-GPU GB300 node and serves both prefill and decode.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+
+dynamo:
+  install: true
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  agg_nodes: 1
+  agg_workers: 1
+  gpus_per_agg: 4
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_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"
+
+backend:
+  type: sglang
+  aggregated_environment:
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: high
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "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"
+
+  sglang_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      enable-metrics: true
+      trust-remote-code: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.94
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 32
+      cuda-graph-max-bs-decode: 32
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 4
+      ep-size: 1
+      speculative-algorithm: EAGLE
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "144"
+
+srun_options:
+  mem: "0"
+  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: "false"
+    TP: "4"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_WARMUP_REQUESTS_PER_LANE: "1"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
new file mode 100644
index 0000000000..dda7f1adad
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
@@ -0,0 +1,115 @@
+name: "agg-gb300-tp8-mtp-lowlatency"
+
+# Low-latency AgentX aggregate topology: one TP8 worker spans two
+# four-GPU GB300 nodes and serves both prefill and decode.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+
+dynamo:
+  install: true
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_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"
+
+backend:
+  type: sglang
+  aggregated_environment:
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: high
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "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"
+
+  sglang_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      enable-metrics: true
+      trust-remote-code: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.94
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 4
+      cuda-graph-max-bs-decode: 4
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 8
+      ep-size: 1
+      speculative-algorithm: EAGLE
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "144"
+
+srun_options:
+  mem: "0"
+  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: "false"
+    TP: "8"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_WARMUP_REQUESTS_PER_LANE: "1"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 2f73b1c64a..c208c27540 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7784,33 +7784,46 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
           dp-attn: true
 
 dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260719-99f5a6f4
+  image: lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "1.3.0.dev20260718" }
+  router: { name: dynamo-router, version: "7a4e47ead90ec6b479e397996f6479ce7d251510" }
   multinode: true
   disagg: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
+    - search-space:
       - spec-decoding: mtp
-        conc-list: [2, 4, 8, 16]
-        kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        conc-list: [1, 4]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 1
+          dp-attn: false
+    - search-space:
+      - spec-decoding: mtp
+        conc-list: [8]
         prefill:
           num-worker: 1
           tp: 4
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml"
-        # Aggregated worker: prefill's GPUs also serve decode, so decode has no
-        # separate allocation (num-worker: 0). Deployment is recipe-driven
-        # (agg_workers: 1); this only makes GPU accounting count 4, not 4+4.
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml"
         decode:
           num-worker: 0
           tp: 4

From fd9e000158026cabae0875fa1d8d653901bce74d Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Tue, 18 Aug 2026 14:42:33 +0800
Subject: [PATCH 2/6] chore(perf): register SGLang low-concurrency sweep
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Append the AgentX aggregate configuration to the performance changelog for PR #2644.

中文：在性能变更记录末尾登记 PR #2644 的 SGLang AgentX 聚合低并发配置。
---
 perf-changelog.yaml | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 085e9cb889..27abc88b75 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6049,3 +6049,12 @@
     - "Add GB200 GLM-5.2 FP4 Dynamo-SGLang AgentX with EAGLE MTP, HiCache DRAM offload, and KV-aware correlation-ID affinity."
     - "Measure the TP8 aggregate curve and tuned 1P1D TP4 HiCache c10/c12 disaggregated frontier with every logical SGLang metrics endpoint."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2620
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Tune the GB300 SGLang AgentX aggregate low-concurrency frontier with TP8 concurrency 1/4 and TP4 concurrency 8 profiles."
+    - "Use phase-specific decode CUDA graph limits and throughput-only synthetic MTP acceptance; eval-only runs retain real target verification."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2644

From ad7466708c0883042a4de0c6bb9661315026e42e Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Tue, 18 Aug 2026 14:53:51 +0800
Subject: [PATCH 3/6] chore(agentx): align SGLang and Dynamo versions
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use SGLang nightly-dev-cu13-20260811-d59c1ddf and the Dynamo 1.4.0 release across the GB300 aggregate and disaggregated recipe family.

中文：将 GB300 聚合与分离式配置族统一更新到 SGLang nightly-dev-cu13-20260811-d59c1ddf 和 Dynamo 1.4.0 正式版。
---
 .../deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml   | 4 +++-
 .../deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml   | 4 +++-
 .../disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml  | 2 +-
 .../disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml       | 2 +-
 .../disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml    | 2 +-
 .../disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml    | 2 +-
 .../disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml    | 2 +-
 configs/nvidia-master.yaml                                  | 6 +++---
 perf-changelog.yaml                                         | 2 ++
 9 files changed, 16 insertions(+), 10 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
index 300080dca5..bc59062e0b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
@@ -13,10 +13,12 @@ identity:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
     image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
 
 dynamo:
   install: true
-  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "4:00:00"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
index dda7f1adad..2d961e467c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
@@ -13,10 +13,12 @@ identity:
     repo: "deepseek-ai/DeepSeek-V4-Pro"
   container:
     image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
 
 dynamo:
   install: true
-  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "4:00:00"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
index 709962936f..ad3bb91780 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "8:00:00"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
index e414387655..e0d7392259 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "8:00:00"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
index 74b371e94e..01784bf425 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "8:00:00"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
index 6b585fba80..380b78f622 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "8:00:00"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
index b2cac21c78..c42153480d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  wheel: "1.4.0"
 
 slurm:
   time_limit: "8:00:00"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index c208c27540..90916fb20d 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7790,7 +7790,7 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "7a4e47ead90ec6b479e397996f6479ce7d251510" }
+  router: { name: dynamo-router, version: "1.4.0" }
   multinode: true
   disagg: false
   scenarios:
@@ -7831,13 +7831,13 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
           dp-attn: false
 
 dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260719-99f5a6f4
+  image: lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "1.3.0.dev20260718" }
+  router: { name: dynamo-router, version: "1.4.0" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 27abc88b75..aca4780724 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4885,6 +4885,7 @@
 - config-keys:
     - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
     - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
   description:
     - "Add GB300 DeepSeek-V4 dynamo-sglang agentic recipes (agg-tp4 + dep8 disagg pareto)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2157
@@ -6056,5 +6057,6 @@
     - agentic-coding
   description:
     - "Tune the GB300 SGLang AgentX aggregate low-concurrency frontier with TP8 concurrency 1/4 and TP4 concurrency 8 profiles."
+    - "Update the aggregate and disaggregated configurations to SGLang nightly-dev-cu13-20260811-d59c1ddf and the Dynamo 1.4.0 release."
     - "Use phase-specific decode CUDA graph limits and throughput-only synthetic MTP acceptance; eval-only runs retain real target verification."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2644

From b402af5ffacc8936e154b4ea39508e4d04efacc3 Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Tue, 18 Aug 2026 16:33:43 +0800
Subject: [PATCH 4/6] fix(perf): restore append-only changelog scope
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Keep the historical PR #2157 entry unchanged and select both aggregate and disaggregated keys in the PR #2644 entry.

中文：恢复性能变更记录的仅追加约束，保持历史 PR #2157 条目不变，并在 PR #2644 条目中同时选择聚合与分离式配置。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index aca4780724..ae1ed458c1 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4885,7 +4885,6 @@
 - config-keys:
     - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
     - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
-    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
   description:
     - "Add GB300 DeepSeek-V4 dynamo-sglang agentic recipes (agg-tp4 + dep8 disagg pareto)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2157
@@ -6053,6 +6052,7 @@
 
 - config-keys:
     - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
   scenario-type:
     - agentic-coding
   description:

From 589cf26ad04b8c79665f06024849fc4c43ce9cb3 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Tue, 18 Aug 2026 11:14:51 -0700
Subject: [PATCH 5/6] fix(agentx): complete GB300 SGLang sweep contract
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

完善 GB300 SGLang AgentX 配置：同步配方镜像与身份信息，将模拟验收率限定到吞吐任务，并统一 Slurm 账户与作业时限。
---
 .../agentic/agg-gb300-tp4-mtp-lowlatency.yaml    |  5 +----
 .../agentic/agg-gb300-tp8-mtp-lowlatency.yaml    |  5 +----
 ...300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml | 16 +++++++++-------
 ...gg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml | 16 +++++++++-------
 ...gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml | 16 +++++++++-------
 ...gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml | 16 +++++++++-------
 ...gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml | 16 +++++++++-------
 configs/nvidia-master.yaml                       | 10 ++++++++++
 perf-changelog.yaml                              |  1 +
 runners/launch_gb300-nv.sh                       | 10 +++++++---
 10 files changed, 65 insertions(+), 46 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
index bc59062e0b..b2aa4a88e1 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
@@ -5,7 +5,7 @@ name: "agg-gb300-tp4-mtp-lowlatency"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
 identity:
@@ -20,9 +20,6 @@ dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "4:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
index 2d961e467c..336b90ca56 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
@@ -5,7 +5,7 @@ name: "agg-gb300-tp8-mtp-lowlatency"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
 identity:
@@ -20,9 +20,6 @@ dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "4:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
index ad3bb91780..e305f86242 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
@@ -10,16 +10,21 @@ name: "disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
+
 dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "8:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -94,9 +99,6 @@ backend:
 
   decode_environment:
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
index e0d7392259..151fe388eb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
@@ -10,16 +10,21 @@ name: "disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
+
 dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "8:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -94,9 +99,6 @@ backend:
     # Match the TP4 aggregate worker: avoid the pathological DSV4 EAGLE
     # draft-extend graph capture while retaining the target/decode graphs.
     SGLANG_DISABLE_DRAFT_EXTEND_CUDA_GRAPH: "1"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
index 01784bf425..fc0ef0a480 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
@@ -10,16 +10,21 @@ name: "disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
+
 dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "8:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -91,9 +96,6 @@ backend:
 
   decode_environment:
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
index 380b78f622..c396796a27 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
@@ -10,16 +10,21 @@ name: "disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
+
 dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "8:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -90,9 +95,6 @@ backend:
 
   decode_environment:
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
index c42153480d..803150680c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
@@ -10,16 +10,21 @@ name: "disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload"
 
 model:
   path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260811-d59c1ddf"
+  frameworks:
+    dynamo: "1.4.0"
+
 dynamo:
   install: true
   wheel: "1.4.0"
 
-slurm:
-  time_limit: "8:00:00"
-
 health_check:
   max_attempts: 1440
   interval_seconds: 10
@@ -91,9 +96,6 @@ backend:
 
   decode_environment:
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index dfbdc7a807..381b7af38f 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8010,6 +8010,8 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
@@ -8026,6 +8028,8 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
@@ -8042,6 +8046,8 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
@@ -8058,6 +8064,8 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
@@ -8074,6 +8082,8 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
           - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a48d2462a9..71986ce11c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6085,4 +6085,5 @@
     - "Tune the GB300 SGLang AgentX aggregate low-concurrency frontier with TP8 concurrency 1/4 and TP4 concurrency 8 profiles."
     - "Update the aggregate and disaggregated configurations to SGLang nightly-dev-cu13-20260811-d59c1ddf and the Dynamo 1.4.0 release."
     - "Use phase-specific decode CUDA graph limits and throughput-only synthetic MTP acceptance; eval-only runs retain real target verification."
+    - "Use an eight-hour DSV4 AgentX allocation and pass the configured GB300 Slurm account to image-validation steps."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2644
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 88b964a8ad..b682e97d20 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -118,7 +118,7 @@ NGINX_SQUASH_FILE="/data/home/sa-shared/gharunners/squash/$(echo "$NGINX_IMAGE"
 import_squash() {
     local squash="$1" image="$2"
     local lock="${squash}.lock"
-    srun --partition=$SLURM_PARTITION --exclusive --time=180 bash -c "
+    srun --account="$SLURM_ACCOUNT" --partition="$SLURM_PARTITION" --exclusive --time=180 bash -c "
         exec 9>\"$lock\"
         flock -w 600 9 || { echo 'Failed to acquire lock for $squash' >&2; exit 1; }
         if unsquashfs -l \"$squash\" > /dev/null 2>&1; then
@@ -175,7 +175,7 @@ if [[ "$USES_DCGM_POWER" == "1" ]]; then
     # x86, nodes aarch64).
     import_squash "$DCGM_EXPORTER_SQSH" "$DCGM_EXPORTER_IMAGE"
     test -r "$DCGM_EXPORTER_SQSH" || { echo "Error: DCGM exporter squash not readable: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
-    srun --partition=$SLURM_PARTITION --exclusive --time=30 bash -c "unsquashfs -l \"$DCGM_EXPORTER_SQSH\" > /dev/null" || { echo "Error: DCGM exporter squash invalid: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
+    srun --account="$SLURM_ACCOUNT" --partition="$SLURM_PARTITION" --exclusive --time=30 bash -c "unsquashfs -l \"$DCGM_EXPORTER_SQSH\" > /dev/null" || { echo "Error: DCGM exporter squash invalid: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
     sha256sum "$DCGM_EXPORTER_SQSH" > "$GITHUB_WORKSPACE/exporter-image.sha256"
 fi
 
@@ -383,13 +383,17 @@ echo "Configs available at: $SRT_REPO_DIR/"
 # Create srtslurm.yaml for srtctl (used by both frameworks)
 SRTCTL_ROOT="${SRT_REPO_DIR}"
 echo "Creating srtslurm.yaml configuration..."
+SRT_DEFAULT_TIME_LIMIT="4:00:00"
+if [[ "$IS_AGENTIC" == "1" && "$FRAMEWORK" == "dynamo-sglang" && "$MODEL_PREFIX" == "dsv4" ]]; then
+    SRT_DEFAULT_TIME_LIMIT="8:00:00"
+fi
 cat > srtslurm.yaml <<EOF
 # SRT SLURM Configuration for GB300
 
 # Default SLURM settings
 default_account: "${SLURM_ACCOUNT}"
 default_partition: "${SLURM_PARTITION}"
-default_time_limit: "4:00:00"
+default_time_limit: "${SRT_DEFAULT_TIME_LIMIT}"
 
 # Resource defaults
 gpus_per_node: 4

From e6864f931f0e5774981b69839bc666eb262579c3 Mon Sep 17 00:00:00 2001
From: Rohit Pujar Nagraj <rpujarnagraj@nvidia.com>
Date: Tue, 18 Aug 2026 12:10:39 -0700
Subject: [PATCH 6/6] test(gb300): align power contract with Slurm account
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

更新 GB300 电源通道契约测试，使其验证显式的 Slurm 账户和分区参数。
---
 utils/test_gb300_power_official_contract.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/utils/test_gb300_power_official_contract.py b/utils/test_gb300_power_official_contract.py
index 6255b24a48..7cf017c6ce 100644
--- a/utils/test_gb300_power_official_contract.py
+++ b/utils/test_gb300_power_official_contract.py
@@ -44,7 +44,8 @@ def test_launcher_provisions_exporter_through_shared_squash_path():
     assert_exporter_provisioning(launcher)
     # No SQUASH_DIR var here; the /data/ mount avoids the /home NFS ELOOP bug.
     assert 'DCGM_EXPORTER_SQSH="/data/home/sa-shared/gharunners/squash/' in launcher
-    assert 'srun --partition=$SLURM_PARTITION --exclusive --time=30 bash -c "unsquashfs -l' in launcher
+    assert 'srun --account="$SLURM_ACCOUNT" --partition="$SLURM_PARTITION" --exclusive --time=180' in launcher
+    assert 'srun --account="$SLURM_ACCOUNT" --partition="$SLURM_PARTITION" --exclusive --time=30 bash -c "unsquashfs -l' in launcher
 
 
 def test_launcher_pins_power_producer():