[NV] GB300 DeepSeek-V4-Pro AgentX Dynamo-SGLang session-routing upgrade / [NV] GB300 DeepSeek-V4-Pro AgentX Dynamo-SGLang 会话路由升级

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml

@@ -16,7 +16,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -44,19 +44,18 @@ infra:
 frontend:
   type: dynamo
   nginx_session_affinity: true
-  nginx_session_affinity_header: X-Correlation-ID
+  nginx_session_affinity_header: X-Dynamo-Session-ID
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
-    DYN_ROUTER_TEMPERATURE: "10000000"
     # 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-reset-states: true
+    router-session-affinity-ttl-secs: "3600"
     active-decode-blocks-threshold: "None"
     active-prefill-tokens-threshold: "None"
     active-prefill-tokens-threshold-frac: "None"
@@ -66,6 +65,9 @@ backend:
 
   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
@@ -139,16 +141,13 @@ benchmark:
     # in process_agentic_result.py instead. TP must match sglang_config tp-size.
     IS_MULTINODE: "false"
     TP: "4"
-    # Enable dynamo conv-aware routing (per-session affinity so multi-turn
-    # KV cache is reused across turns). Previously opted out because the
-    # frontend 400'd on aiperf's nvext.session_control actions; re-enabled
-    # to test with the current build.
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    # 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"
     # 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"
-

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1024-mtp-kvoffload.yaml

@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -41,15 +41,14 @@ infra:
 frontend:
   type: dynamo
   nginx_session_affinity: true
-  nginx_session_affinity_header: X-Correlation-ID
+  nginx_session_affinity_header: X-Dynamo-Session-ID
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
-    DYN_ROUTER_TEMPERATURE: "10000000"
     PIP_BREAK_SYSTEM_PACKAGES: "1"
   args:
     router-mode: "kv"
-    router-reset-states: true
+    router-session-affinity-ttl-secs: "3600"
     active-decode-blocks-threshold: "None"
     active-prefill-tokens-threshold: "None"
     active-prefill-tokens-threshold-frac: "None"
@@ -206,7 +205,6 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
-

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c8-mtp-kvoffload.yaml

@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -41,15 +41,14 @@ infra:
 frontend:
   type: dynamo
   nginx_session_affinity: true
-  nginx_session_affinity_header: X-Correlation-ID
+  nginx_session_affinity_header: X-Dynamo-Session-ID
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
-    DYN_ROUTER_TEMPERATURE: "10000000"
     PIP_BREAK_SYSTEM_PACKAGES: "1"
   args:
     router-mode: "kv"
-    router-reset-states: true
+    router-session-affinity-ttl-secs: "3600"
     active-decode-blocks-threshold: "None"
     active-prefill-tokens-threshold: "None"
     active-prefill-tokens-threshold-frac: "None"
@@ -92,6 +91,9 @@ backend:
 
   decode_environment:
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+    # 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"
@@ -205,7 +207,6 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
-

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.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -41,15 +41,14 @@ infra:
 frontend:
   type: dynamo
   nginx_session_affinity: true
-  nginx_session_affinity_header: X-Correlation-ID
+  nginx_session_affinity_header: X-Dynamo-Session-ID
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
-    DYN_ROUTER_TEMPERATURE: "10000000"
     PIP_BREAK_SYSTEM_PACKAGES: "1"
   args:
     router-mode: "kv"
-    router-reset-states: true
+    router-session-affinity-ttl-secs: "3600"
     active-decode-blocks-threshold: "None"
     active-prefill-tokens-threshold: "None"
     active-prefill-tokens-threshold-frac: "None"
@@ -206,7 +205,6 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
-

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c256-mtp-kvoffload.yaml

@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -40,15 +40,14 @@ infra:
 frontend:
   type: dynamo
   nginx_session_affinity: true
-  nginx_session_affinity_header: X-Correlation-ID
+  nginx_session_affinity_header: X-Dynamo-Session-ID
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
-    DYN_ROUTER_TEMPERATURE: "10000000"
     PIP_BREAK_SYSTEM_PACKAGES: "1"
   args:
     router-mode: "kv"
-    router-reset-states: true
+    router-session-affinity-ttl-secs: "3600"
     active-decode-blocks-threshold: "None"
     active-prefill-tokens-threshold: "None"
     active-prefill-tokens-threshold-frac: "None"
@@ -206,7 +205,6 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
-

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c512-mtp-kvoffload.yaml

@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -41,15 +41,14 @@ infra:
 frontend:
   type: dynamo
   nginx_session_affinity: true
-  nginx_session_affinity_header: X-Correlation-ID
+  nginx_session_affinity_header: X-Dynamo-Session-ID
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
-    DYN_ROUTER_TEMPERATURE: "10000000"
     PIP_BREAK_SYSTEM_PACKAGES: "1"
   args:
     router-mode: "kv"
-    router-reset-states: true
+    router-session-affinity-ttl-secs: "3600"
     active-decode-blocks-threshold: "None"
     active-prefill-tokens-threshold: "None"
     active-prefill-tokens-threshold-frac: "None"
@@ -206,7 +205,6 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
-

configs/nvidia-master.yaml

@@ -8836,13 +8836,13 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
           ep: 8
           dp-attn: true
 dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260711-7de33ce8
+  image: lmsysorg/sglang:nightly-dev-cu13-20260719-99f5a6f4
   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.dev1" }
+  router: { name: dynamo-router, version: "1.3.0.dev20260718" }
   multinode: true
   disagg: false
   scenarios:
@@ -8870,13 +8870,13 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
           dp-attn: false
 
 dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260711-7de33ce8
+  image: lmsysorg/sglang:nightly-dev-cu13-20260719-99f5a6f4
   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.dev1" }
+  router: { name: dynamo-router, version: "1.3.0.dev20260718" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true

perf-changelog.yaml

@@ -5355,3 +5355,13 @@
     - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path"
     - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+  description:
+    - "Upgrade Dynamo from 1.3.0.dev1 to 1.3.0.dev20260718 so router queueing is disabled by default at high concurrency"
+    - "Use AIPerf f5ae8134 (including the X-Dynamo-Session-ID session-routing support merged in AIPerf #19) and migrate AgentX session affinity from nvext.session_control to X-Dynamo-Session-ID headers with a 3600-second router TTL"
+    - "Remove the DYN_ROUTER_TEMPERATURE=10000000 override and restore Dynamo's deterministic 0.0 default"
+    - "Upgrade SGLang from nightly-dev-cu13-20260711-7de33ce8 to nightly-dev-cu13-20260719-99f5a6f4, the latest nightly before the scheduler WAR-barrier regression in SGLang #31687"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2319