[Klaud Cold] Update GLM-5.2 FP4 GB200 Dynamo-SGLang AgentX image to nightly-dev-20261028-586355a8 / [Klaud Cold] 将 GLM-5.2 FP4 GB200 Dynamo-SGLang AgentX 镜像更新至 nightly-dev-20261028-586355a8

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/agg-variants.yaml

@@ -6,13 +6,17 @@ base:
     model: {}
     container: {}
     frameworks:
-      dynamo: 1.6.0.dev20261019
-      sglang: 0.0.0.dev1+gab580b2e0
+      dynamo: 1.6.0
+      sglang: 0.0.0.dev1+g586355a83
   resources:
     gpu_type: gb200
     gpus_per_node: 4
+  # Dynamo a789babbe37a made frontend /health omit instances, which srtctl's
+  # readiness check needs; pin its parent until srtctl asks for them.
   dynamo:
-    install: false
+    install: true
+    source:
+      rev: 68c35a1ce41b035d1a79fff4be97aa1c43de62c3
   # Tachometer's per-node exporters slow decode steps and add host memory;
   # InferenceX does not read its output.
   observability:
@@ -40,6 +44,11 @@ base:
         PIP_BREAK_SYSTEM_PACKAGES: '1'
         DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
         SGLANG_TIMEOUT_KEEP_ALIVE: '900'
+        # Dynamo 5df849a06a (#14369) caps each worker at max-running-requests plus a
+        # 2x queue and rejects the rest with 529; keep its pre-engine-report limits.
+        DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: '10000'
+        DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: '10000'
+        DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: '20000'
       args:
         served-model-name: nvidia/GLM-5.2-NVFP4
         trust-remote-code: true

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/glm5.2/sglang/gb200-fp4/agentx/disagg-variants.yaml

@@ -6,13 +6,17 @@ base:
     model: {}
     container: {}
     frameworks:
-      dynamo: 1.6.0.dev20261019
-      sglang: 0.0.0.dev1+gab580b2e0
+      dynamo: 1.6.0
+      sglang: 0.0.0.dev1+g586355a83
   resources:
     gpu_type: gb200
     gpus_per_node: 4
+  # Dynamo a789babbe37a made frontend /health omit instances, which srtctl's
+  # readiness check needs; pin its parent until srtctl asks for them.
   dynamo:
-    install: false
+    install: true
+    source:
+      rev: 68c35a1ce41b035d1a79fff4be97aa1c43de62c3
   # Tachometer's per-node exporters slow decode steps and add host memory;
   # InferenceX does not read its output.
   observability:
@@ -55,6 +59,11 @@ base:
         UCX_CUDA_IPC_ENABLE_MNNVL: 'y'
         # The default 9,9,10,10,10,10,10,10 split also gives the last stage the LM head and MTP draft layer.
         SGLANG_PP_LAYER_PARTITION: 10,10,10,10,10,10,10,8
+        # Dynamo 5df849a06a (#14369) caps each worker at max-running-requests plus a
+        # 2x queue and rejects the rest with 529; keep its pre-engine-report limits.
+        DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: '10000'
+        DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: '10000'
+        DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: '20000'
       args:
         served-model-name: nvidia/GLM-5.2-NVFP4
         trust-remote-code: true
@@ -107,6 +116,11 @@ base:
         NVSHMEM_REMOTE_TRANSPORT: none
         UCX_TLS: cuda_copy,cuda_ipc,sm,self,tcp
         UCX_CUDA_IPC_ENABLE_MNNVL: 'y'
+        # Dynamo 5df849a06a (#14369) caps each worker at max-running-requests plus a
+        # 2x queue and rejects the rest with 529; keep its pre-engine-report limits.
+        DYN_BACKEND_ADMISSION_ENGINE_REQUEST_LIMIT: '10000'
+        DYN_BACKEND_ADMISSION_ENGINE_REQUEST_WAITING_LIMIT: '10000'
+        DYN_BACKEND_ADMISSION_REQUEST_QUEUE_LIMIT: '20000'
       args:
         served-model-name: nvidia/GLM-5.2-NVFP4
         trust-remote-code: true

inferencex-e2e/configs/nvidia-master.yaml

@@ -7569,7 +7569,7 @@ glm5.2-fp4-b200-dynamo-sglang-agentic-disagg:
           dp-attn: true
 
 glm5.2-fp4-gb200-dynamo-sglang-agentic-agg:
-  image: lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576
+  image: lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90
   model: nvidia/GLM-5.2-NVFP4
   model-prefix: glm5.2
   runner: cluster:gb200-nv
@@ -7610,7 +7610,7 @@ glm5.2-fp4-gb200-dynamo-sglang-agentic-agg:
           ep: 1
           dp-attn: false
 glm5.2-fp4-gb200-dynamo-sglang-agentic-disagg:
-  image: lmsysorg/sglang:nightly-dev-20261020-ab580b2e@sha256:0af411bb320a95bb5e1c58b4c20c9852920a19aa9f12f148f69731b0dca3b576
+  image: lmsysorg/sglang:nightly-dev-20261028-586355a8@sha256:b4f09a2ccbcfcc262d49600c3cd7f8581a793ea7d104947257b6a95500953b90
   model: nvidia/GLM-5.2-NVFP4
   model-prefix: glm5.2
   runner: cluster:gb200-nv

inferencex-e2e/perf-changelog.yaml

@@ -11499,3 +11499,11 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4509
   scenario-type: [agentic-coding]
   append-only: true
+
+- config-keys:
+    - glm5.2-fp4-gb200-dynamo-sglang-agentic-agg
+    - glm5.2-fp4-gb200-dynamo-sglang-agentic-disagg
+  description:
+    - "Update the SGLang image from nightly-dev-20261020-ab580b2e to nightly-dev-20261028-586355a8 and pin Dynamo 68c35a1."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4611
+  scenario-type: [agentic-coding]