[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
@@ -140,7 +142,8 @@ benchmark:
     # in process_agentic_result.py instead. TP must match sglang_config tp-size.
     IS_MULTINODE: "false"
     TP: "4"
-    # Bind each conversation through the supported Dynamo session header.
+    # 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:"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1024-mtp-kvoffload.yaml → benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml

@@ -1,7 +1,7 @@
-name: "disagg-gb300-12p4d-dep8-dep16-c1024-mtp-kvoffload"
+name: "disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (12P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 1024.
+# (12P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 1536.
 #
 # Uses the flat single-variant srtctl schema the agentic CI flow expects;
 # resources + backend (prefill/decode env + sglang_config) are normalized
@@ -15,7 +15,7 @@ model:
 
 dynamo:
   install: true
-  wheel: "1.3.0.dev1"
+  wheel: "1.3.0.dev20260718"
 
 slurm:
   time_limit: "8:00:00"
@@ -41,18 +41,17 @@ 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"
     # AgentX warmup can legitimately keep the single wide decode worker busy
     # for longer than Dynamo's 10-second TCP request-plane default.
     DYN_TCP_REQUEST_TIMEOUT: "60"
     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"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c8-mtp-kvoffload.yaml → benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml

@@ -1,7 +1,7 @@
-name: "disagg-gb300-2p1d-dep8-tp4-c8-mtp-kvoffload"
+name: "disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (2P x DEP8 / 1D x TP4, MTP + hierarchical-cache KV offload), tuned for concurrency 8.
+# (2P x DEP8 / 1D x TP4, MTP + hierarchical-cache KV offload), tuned for concurrency 80.
 #
 # Uses the flat single-variant srtctl schema the agentic CI flow expects;
 # resources + backend (prefill/decode env + sglang_config) are normalized
@@ -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"

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"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c256-mtp-kvoffload.yaml → benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml

@@ -1,7 +1,7 @@
-name: "disagg-gb300-4p4d-dep8-dep16-c256-mtp-kvoffload"
+name: "disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated recipe for DeepSeek-V4-Pro on GB300
-# (4P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 256.
+# (4P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 512.
 #
 # DEP8-prefill variant aligned with the measured Pareto point: prefill uses
 # tp/dp/ep 8, four nodes, and SGLANG_DSV4_MHC_PREWARM=1.
@@ -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"

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c512-mtp-kvoffload.yaml → benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml

@@ -1,7 +1,7 @@
-name: "disagg-gb300-6p4d-dep8-dep16-c512-mtp-kvoffload"
+name: "disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (6P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 512.
+# (6P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 768.
 #
 # Uses the flat single-variant srtctl schema the agentic CI flow expects;
 # resources + backend (prefill/decode env + sglang_config) are normalized
@@ -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"

configs/nvidia-master.yaml

@@ -7419,21 +7419,21 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
           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:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
       - spec-decoding: mtp
-        conc-list: [1, 4, 8, 16]
+        conc-list: [2, 4, 8, 16]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7453,13 +7453,13 @@ 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
@@ -7468,7 +7468,7 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
     - dram-utilization: 0.80
       search-space:
       - spec-decoding: mtp
-        conc-list: [8]
+        conc-list: [80]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7477,7 +7477,7 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c8-mtp-kvoffload.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 4
@@ -7500,7 +7500,7 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 16
           dp-attn: true
       - spec-decoding: mtp
-        conc-list: [256]
+        conc-list: [512]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7509,14 +7509,14 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c256-mtp-kvoffload.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16
           ep: 16
           dp-attn: true
       - spec-decoding: mtp
-        conc-list: [512]
+        conc-list: [768]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7525,14 +7525,14 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c512-mtp-kvoffload.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16
           ep: 16
           dp-attn: true
       - spec-decoding: mtp
-        conc-list: [1024]
+        conc-list: [1536]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7541,7 +7541,7 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1024-mtp-kvoffload.yaml"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16

perf-changelog.yaml

@@ -5709,3 +5709,14 @@
     - "Enable SGLang metrics on every aggregate, prefill, and decode engine."
     - "Use supported header-based Dynamo session routing with the in-repo AIPerf build."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2520
+
+- 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"
+    - "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"
+    - "Restore the pre-watchdog AgentX concurrency grid and use AIPerf's default 300-second per-trajectory idle-gap cap"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2319