feat(agentx): refresh DSV4 GB300 SGLang configs without W4A4 MegaMoE / 更新 DSV4 GB300 SGLang 配置并移除 W4A4 MegaMoE

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb300-fp4/agentx/agg-variants.yaml

@@ -13,11 +13,11 @@ base:
     model:
       repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
     container:
-      image: "lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9"
+      image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
   dynamo:
     install: true
     source:
-      wheel: "1.5.0.dev20260910"
+      wheel: "1.5.0.dev20260914"
   slurm:
     time_limit: "4:00:00"
   health_check:
@@ -77,7 +77,7 @@ base:
         weight-loader-prefetch-checkpoints: true
         stream-interval: 10
         watchdog-timeout: 1000000
-        mem-fraction-static: 0.94
+        mem-fraction-static: 0.90
         page-size: 256
         chunked-prefill-size: 8192
         max-prefill-tokens: 8192
@@ -107,6 +107,8 @@ base:
       RESULT_DIR: "/logs/agentic"
       PORT: "8000"
       IS_MULTINODE: "false"
+      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:"
@@ -115,31 +117,31 @@ base:
 
 # Low-latency AgentX aggregate topology: one TP4 worker occupies one
 # four-GPU GB300 node and serves both prefill and decode with DSpark K=6.
-override_tp4:
-  name: "agg-gb300-tp4-mtp-lowlatency"
+override_tp4_c4:
+  name: "agg-gb300-tp4-c4-mtp-lowlatency"
   roles:
     agg:
       nodes: 1
       gpus: 4
       args:
-        max-running-requests: 32
-        cuda-graph-max-bs-decode: 32
+        max-running-requests: 8
+        cuda-graph-max-bs-decode: 8
         tp-size: 4
   benchmark:
     env:
       TP: "4"
 
 # Low-latency AgentX aggregate topology: one TP8 worker spans two
 # four-GPU GB300 nodes and serves both prefill and decode with DSpark K=6.
-override_tp8:
-  name: "agg-gb300-tp8-mtp-lowlatency"
+override_tp8_c1:
+  name: "agg-gb300-tp8-c1-mtp-lowlatency"
   roles:
     agg:
       nodes: 2
       gpus: 8
       args:
-        max-running-requests: 4
-        cuda-graph-max-bs-decode: 4
+        max-running-requests: 2
+        cuda-graph-max-bs-decode: 2
         tp-size: 8
   benchmark:
     env:

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb300-fp4/agentx/disagg-variants.yaml

@@ -13,11 +13,11 @@ base:
     model:
       repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
     container:
-      image: "lmsysorg/sglang:nightly-dev-cu13-20260829-89816a21"
+      image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
   dynamo:
     install: true
     source:
-      wheel: "1.5.0.dev20260902"
+      wheel: "1.5.0.dev20260914"
   slurm:
     time_limit: "8:00:00"
   health_check:
@@ -103,7 +103,6 @@ base:
         moe-dense-tp-size: 1
         moe-a2a-backend: megamoe
         enable-deepseek-v4-fp4-indexer: true
-        enable-w4a4-mxfp4-megamoe: true
         disaggregation-transfer-backend: mooncake
         disaggregation-mode: prefill
         load-balance-method: total_tokens
@@ -172,7 +171,6 @@ base:
         moe-dense-tp-size: 1
         moe-a2a-backend: megamoe
         enable-deepseek-v4-fp4-indexer: true
-        enable-w4a4-mxfp4-megamoe: true
         disaggregation-transfer-backend: mooncake
         disaggregation-mode: decode
         load-balance-method: total_tokens
@@ -208,27 +206,57 @@ base:
       AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
       HF_HUB_CACHE: "/hf_hub_cache"
 
-# Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (2P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 480.
-#
-# Uses the flat single-variant srtctl schema the agentic CI flow expects;
-# resources + backend (prefill/decode env + sglang_config) are normalized
-# from the Pareto run.
-# Concurrency is exported into srt_agentic.sh from the master-config conc-list.
-override_1p1d_c480:
-  name: "disagg-gb300-2p4d-dep8-dep16-c480-mtp-kvoffload"
+# One DEP8 prefill worker and one DEP16 decode worker at concurrency 64.
+override_1p1d_c64:
+  name: "disagg-gb300-1p1d-dep8-dep16-c64-mtp-kvoffload"
   roles:
     prefill:
       nodes: 2
       workers: 1
       gpus: 8
       args:
         max-running-requests: 256
-        cuda-graph-max-bs: 256
+        cuda-graph-max-bs-decode: 256
+        hicache-mem-layout: page_first_direct
+    decode:
+      gpus: 16
+      args:
+        max-running-requests: 128
+        cuda-graph-max-bs-decode: 256
+
+# One DEP8 prefill worker and one DEP16 decode worker at concurrency 240.
+override_1p1d_c240:
+  name: "disagg-gb300-1p1d-dep8-dep16-c240-mtp-kvoffload"
+  roles:
+    prefill:
+      nodes: 2
+      workers: 1
+      gpus: 8
+      args:
+        max-running-requests: 512
+        cuda-graph-max-bs-decode: 256
     decode:
       gpus: 16
       args:
-        cuda-graph-max-bs: 256
+        max-running-requests: 512
+        cuda-graph-max-bs-decode: 256
+
+# Two DEP8 prefill workers and one DEP16 decode worker at concurrency 480.
+override_2p1d_c480:
+  name: "disagg-gb300-2p1d-dep8-dep16-c480-mtp-kvoffload"
+  roles:
+    prefill:
+      nodes: 4
+      workers: 2
+      gpus: 8
+      args:
+        max-running-requests: 512
+        cuda-graph-max-bs-decode: 256
+    decode:
+      gpus: 16
+      args:
+        max-running-requests: 1024
+        cuda-graph-max-bs-decode: 256
 
 # Agentic-coding SGLang disaggregated recipe for DeepSeek-V4-Pro on GB300
 # (4P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 960.
@@ -247,13 +275,13 @@ override_2p1d_c960:
         OMP_NUM_THREADS: '1'
       args:
         max-running-requests: 256
-        cuda-graph-max-bs: 256
+        cuda-graph-max-bs-decode: 256
     decode:
       env:
         OMP_NUM_THREADS: '1'
         SGLANG_DSV4_MHC_PREWARM: '1'
       args:
-        cuda-graph-max-bs: 256
+        cuda-graph-max-bs-decode: 256
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
 # (6P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 1440.
@@ -271,11 +299,11 @@ override_3p1d_c1440:
       gpus: 8
       args:
         max-running-requests: 512
-        cuda-graph-max-bs: 512
+        cuda-graph-max-bs-decode: 512
     decode:
       gpus: 16
       args:
-        cuda-graph-max-bs: 512
+        cuda-graph-max-bs-decode: 512
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
 # (8P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 1920.
@@ -301,13 +329,13 @@ override_4p1d_c1920:
         SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "3600"
       args:
         max-running-requests: 1024
-        cuda-graph-max-bs: 1024
+        cuda-graph-max-bs-decode: 1024
     decode:
       gpus: 16
       env:
         SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "3600"
       args:
-        cuda-graph-max-bs: 192
+        cuda-graph-max-bs-decode: 192
   benchmark:
     env:
       AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"

inferencex-e2e/configs/nvidia-master.yaml

@@ -6639,47 +6639,47 @@ kimik3-fp4-gb200-dynamo-vllm-agentic-dspark-mooncake-tp8pp2:
           additional-settings:
           - "CONFIG_FILE=recipes/kimik3/vllm/gb200-fp4/agentx/agg-tp8pp2-mooncake-c96.yaml"
 dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
-  image: lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9
+  image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
   model: deepseek-ai/DeepSeek-V4-Pro-0813
   model-prefix: dsv4
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "1.5.0.dev20260910" }
+  router: { name: dynamo-router, version: "1.5.0.dev20260914" }
   multinode: true
   disagg: false
   scenarios:
     agentic-coding:
     - search-space:
       - spec-decoding: draft_model
-        conc-list: [1, 4]
+        conc-list: [1]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/agg-variants.yaml:override_tp8"
+          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/agg-variants.yaml:override_tp8_c1"
     - search-space:
       - spec-decoding: draft_model
-        conc-list: [8]
+        conc-list: [4]
         num-nodes: 1
         worker:
           num-worker: 1
           tp: 4
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/agg-variants.yaml:override_tp4"
+          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/agg-variants.yaml:override_tp4_c4"
 dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260829-89816a21
+  image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
   model: deepseek-ai/DeepSeek-V4-Pro-0813
   model-prefix: dsv4
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "1.5.0.dev20260902" }
+  router: { name: dynamo-router, version: "1.5.0.dev20260914" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -6688,7 +6688,7 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
     - dram-utilization: 0.80
       search-space:
       - spec-decoding: draft_model
-        conc-list: [480]
+        conc-list: [64]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -6697,7 +6697,39 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p1d_c480"
+          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p1d_c64"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      - spec-decoding: draft_model
+        conc-list: [240]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p1d_c240"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      - spec-decoding: draft_model
+        conc-list: [480]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_2p1d_c480"
         decode:
           num-worker: 1
           tp: 16

inferencex-e2e/perf-changelog.yaml

@@ -9211,3 +9211,17 @@
     - "Add --attention-config '{\"indexer_kv_dtype\":\"mxfp4\",\"indexer_sparse_logits\":true}' and --block-size 128 to enable the vllm-project/vllm#58671 ROCm paged MXFP4 sparse-logits indexer, replacing the dense fp8 indexer path. A live A/B test (TP2 c16, matched 900s window, vllm-project/vllm#58208 reverted via vllm-project/vllm#59125 so the dense fallback doesn't crash) measured +14.7/+14.9% p50/p90 interactivity and -9.5/-10.8% p50/p90 e2e latency over the dense path, with throughput/GPU unchanged."
     - "Drop c128 from both TP2 and TP4. Neither c128 point was on the Pareto frontier in #3555's run 36528242520: TP2 c64 dominated both (P90 E2EL 58 s against 185 s and 85 s, at 111k against 85k and 79k total tok/s/GPU)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3571
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Refresh DeepSeek-V4-Pro-0813 GB300 Dynamo+SGLang AgentX on nightly-dev-20260916-c9a8fba9 and Dynamo 1.5.0.dev20260914. Keep TP8 c1 and TP4 c4 aggregate points, and DEP8 prefill / DEP16 decode points at c64, c240, c480, c960, c1440 and c1920."
+    - "Replace the long-TTFT 1P1D c480 point with 1P1D c240 and 2P1D c480, retain DSpark K=6 on both serving stages, and use the current SGLang CUDA graph flags. Adapt the recipes and registry to the shared variant layout under inferencex-e2e/."
+    - "Drop enable-w4a4-mxfp4-megamoe from the disaggregated recipes, and set PP_SIZE and PCP_SIZE in the aggregate benchmark environment."
+    - "将 DeepSeek-V4-Pro-0813 GB300 Dynamo+SGLang AgentX 更新至 nightly-dev-20260916-c9a8fba9 和 Dynamo 1.5.0.dev20260914。聚合配置保留 TP8 c1 与 TP4 c4；DEP8 预填充、DEP16 解码配置覆盖 c64、c240、c480、c960、c1440 和 c1920。"
+    - "将首 token 延迟较长的 1P1D c480 替换为 1P1D c240 与 2P1D c480，在预填充和解码阶段均使用 DSpark K=6，并采用当前的 SGLang CUDA graph 参数。配方和注册项适配 inferencex-e2e/ 下的共享变体布局。"
+    - "分离式配方移除 enable-w4a4-mxfp4-megamoe；聚合配置的基准测试环境设置 PP_SIZE 与 PCP_SIZE。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3630