[NVIDIA][AgentX] Refresh B300 DeepSeek-V4.1-Flash SGLang aggregate curve / [NVIDIA][AgentX] 更新 B300 DeepSeek-V4.1-Flash SGLang 聚合曲线 (#3744)

* feat(agentx): refresh B300 DSV4.1 Flash SGLang curve

更新 B300 DeepSeek-V4.1-Flash SGLang AgentX 镜像、聚合式九点配置及配套文档。

* chore(changelog): link B300 SGLang refresh PR

将 B300 DeepSeek-V4.1-Flash SGLang 更新条目链接至 upstream PR。

* config(agentx): use default DeepSeek V4 indexer path on B300

Remove the explicit FP4 indexer option from the B300 DeepSeek-V4.1-Flash SGLang recipe and align the performance changelog with the final configuration.

---------

Co-authored-by: functionstackx <47992694+functionstackx@users.noreply.github.com>

Author: shyeh25

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b300-fp4-mtp/agentic.yaml

@@ -1,12 +1,19 @@
-# DeepSeek-V4.1-Flash AgentX on B300 with SGLang native DSpark. The KV cache
-# is GPU-resident.
 base:
   schema: 2
   name: dsv41flash-fp4-b300-sglang-agentic
   model:
     path: hf:deepseek-ai/DeepSeek-V4.1-Flash
-    container: lmsysorg/sglang:nightly-dev-cu13-20260922-582389ce@sha256:35ea4d321b0735051dcce2599fc495853ad1ef30f6a1908f0a75e74204362d14
+    container: lmsysorg/sglang:nightly-dev-cu13-20261004-295a53c9@sha256:7cfd4dd8e0e902d2a3c0560565963c4e172091e1586aaed15ff5a0a131394d34
     precision: fp4
+  identity:
+    model:
+      repo: deepseek-ai/DeepSeek-V4.1-Flash
+      revision: dba1be0a40aa45a94ad051997016db3960a90277
+    container:
+      image: lmsysorg/sglang:nightly-dev-cu13-20261004-295a53c9@sha256:7cfd4dd8e0e902d2a3c0560565963c4e172091e1586aaed15ff5a0a131394d34
+    frameworks:
+      sglang: 0.0.0.dev1+g295a53c90
+      sglang-router: 0.3.2
   resources:
     gpu_type: b300
     gpus_per_node: 8
@@ -27,13 +34,13 @@ base:
         trust-remote-code: true
         # Overlap checkpoint reads with weight loading from local NVMe.
         weight-loader-prefetch-checkpoints: true
-        chunked-prefill-size: 4096
+        chunked-prefill-size: 16384
         # Keep active decode requests progressing while long prefixes are queued.
         prefill-decode-interval: 16
         # DSpark is the checkpoint's bundled draft; the block size is its only knob.
         speculative-algorithm: DSPARK
         speculative-dspark-block-size: 5
-        # Admission never exceeds the captured decode graph batch.
+        # Capture decode graphs up to batch 64.
         cuda-graph-max-bs-decode: 64
         reasoning-parser: auto
         tool-call-parser: auto
@@ -60,17 +67,15 @@ base:
       AIPERF_HTTP_TCP_USER_TIMEOUT: '900000'
       AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
 
-# One variant per point. Admission is 2x CONC, capped at the graph batch. SWA
-# prefix tails are 64x CONC within [128, 4096]. TP2 raises the static fraction
-# (0.80 from c16, 0.85 from c32) and interleaves decode more often from c32.
-# Saturation points get a longer warmup drain.
+# TP arms admit up to 2x CONC, capped at 64; DP4 C192 admits 384 globally.
+# SWA prefix tails are 64x per-rank CONC, with a minimum of 128.
 override_tp4_c1:
   roles:
     agg:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        expert-parallel-size: 4
+        expert-parallel-size: 1
         mem-fraction-static: 0.7
         swa-prefix-tails: 128
         max-running-requests: 2
@@ -79,118 +84,29 @@ override_tp4_c1:
       CONC: '1'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
 
-override_tp4_c2:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        expert-parallel-size: 4
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 128
-        max-running-requests: 4
-  benchmark:
-    env:
-      CONC: '2'
-      AGENTIC_WARMUP_GRACE_PERIOD: '1800'
-
 override_tp4_c4:
   roles:
     agg:
       gpus: 4
       args:
         tensor-parallel-size: 4
-        expert-parallel-size: 4
+        expert-parallel-size: 1
         mem-fraction-static: 0.7
         swa-prefix-tails: 256
         max-running-requests: 8
+        prefill-decode-interval: 192
   benchmark:
     env:
       CONC: '4'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
 
-override_tp4_c8:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        expert-parallel-size: 4
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 512
-        max-running-requests: 16
-  benchmark:
-    env:
-      CONC: '8'
-      AGENTIC_WARMUP_GRACE_PERIOD: '1800'
-
-override_tp4_c16:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        expert-parallel-size: 4
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 1024
-        max-running-requests: 32
-  benchmark:
-    env:
-      CONC: '16'
-      AGENTIC_WARMUP_GRACE_PERIOD: '1800'
-
-override_tp4_c32:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        expert-parallel-size: 4
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 2048
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '32'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
-override_tp4_c64:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        expert-parallel-size: 4
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 4096
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '64'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
-override_tp4_c128:
-  roles:
-    agg:
-      gpus: 4
-      args:
-        tensor-parallel-size: 4
-        expert-parallel-size: 4
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 4096
-        max-running-requests: 64
-  benchmark:
-    env:
-      CONC: '128'
-      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
-
 override_tp2_c1:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        expert-parallel-size: 2
+        expert-parallel-size: 1
         mem-fraction-static: 0.7
         swa-prefix-tails: 128
         max-running-requests: 2
@@ -199,110 +115,116 @@ override_tp2_c1:
       CONC: '1'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
 
-override_tp2_c2:
-  roles:
-    agg:
-      gpus: 2
-      args:
-        tensor-parallel-size: 2
-        expert-parallel-size: 2
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 128
-        max-running-requests: 4
-  benchmark:
-    env:
-      CONC: '2'
-      AGENTIC_WARMUP_GRACE_PERIOD: '1800'
-
 override_tp2_c4:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        expert-parallel-size: 2
+        expert-parallel-size: 1
         mem-fraction-static: 0.7
         swa-prefix-tails: 256
         max-running-requests: 8
+        prefill-decode-interval: 96
   benchmark:
     env:
       CONC: '4'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
 
-override_tp2_c8:
+override_tp2_c16:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        expert-parallel-size: 2
-        mem-fraction-static: 0.7
-        swa-prefix-tails: 512
-        max-running-requests: 16
+        expert-parallel-size: 1
+        mem-fraction-static: 0.8
+        swa-prefix-tails: 1024
+        max-running-requests: 32
+        prefill-decode-interval: 192
   benchmark:
     env:
-      CONC: '8'
+      CONC: '16'
       AGENTIC_WARMUP_GRACE_PERIOD: '1800'
 
-override_tp2_c16:
+override_tp2_c24:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        expert-parallel-size: 2
+        expert-parallel-size: 1
         mem-fraction-static: 0.8
-        swa-prefix-tails: 1024
-        max-running-requests: 32
+        swa-prefix-tails: 1536
+        max-running-requests: 48
+        prefill-decode-interval: 96
   benchmark:
     env:
-      CONC: '16'
-      AGENTIC_WARMUP_GRACE_PERIOD: '1800'
+      CONC: '24'
+      AGENTIC_WARMUP_GRACE_PERIOD: '3600'
 
 override_tp2_c32:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        expert-parallel-size: 2
+        expert-parallel-size: 1
         mem-fraction-static: 0.85
         swa-prefix-tails: 2048
         max-running-requests: 64
-        prefill-decode-interval: 4
+        prefill-decode-interval: 64
   benchmark:
     env:
       CONC: '32'
       AGENTIC_WARMUP_GRACE_PERIOD: '3600'
 
-override_tp2_c64:
+override_tp2_c48:
   roles:
     agg:
       gpus: 2
       args:
         tensor-parallel-size: 2
-        expert-parallel-size: 2
-        mem-fraction-static: 0.85
-        swa-prefix-tails: 4096
+        expert-parallel-size: 1
+        mem-fraction-static: 0.8
+        swa-prefix-tails: 3072
         max-running-requests: 64
-        prefill-decode-interval: 4
+        prefill-decode-interval: 64
   benchmark:
     env:
-      CONC: '64'
+      CONC: '48'
       AGENTIC_WARMUP_GRACE_PERIOD: '3600'
 
-override_tp2_c128:
+override_dp4_c192:
+  frontend:
+    type: sglang-router
+    enable_multiple_frontends: false
+    args:
+      policy: cache_aware
+      request-id-headers: x-correlation-id
+      dp-aware: true
+      connect-timeout-secs: 900
+      request-timeout-secs: 14400
+      disable-health-check: true
+      disable-retries: true
   roles:
     agg:
-      gpus: 2
+      gpus: 4
       args:
-        tensor-parallel-size: 2
-        expert-parallel-size: 2
-        mem-fraction-static: 0.85
-        swa-prefix-tails: 4096
-        max-running-requests: 64
-        prefill-decode-interval: 4
-  benchmark:
-    env:
-      CONC: '128'
+        tensor-parallel-size: 4
+        expert-parallel-size: 4
+        data-parallel-size: 4
+        enable-dp-attention: true
+        enable-dp-lm-head: true
+        enable-dp-attention-local-control-broadcast: true
+        load-balance-method: total_requests
+        tokenizer-worker-num: 8
+        chunked-prefill-size: 32768
+        mem-fraction-static: 0.76
+        swa-prefix-tails: 3072
+        max-running-requests: 384
+  benchmark:
+    env:
+      CONC: '192'
       AGENTIC_WARMUP_GRACE_PERIOD: '3600'
+      AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID: 'true'

inferencex-e2e/configs/nvidia-master.yaml

@@ -9207,11 +9207,9 @@ qwen3.5-fp8-b300-dynamo-sglang-agentic-disagg:
           ep: 2
           dp-attn: false
 
-# Official 2026-09-21 CUDA 13 nightly, amd64 manifest pinned for B300.
-# Latest multiarch index: sha256:987c7e4bd26918647211a5dcad72a2bdf2a5f394ac1469eff730e7517fc139be.
-# Uses the cookbook's automatic backends and the existing B200 AgentX memory bounds.
+# Official 2026-10-04 CUDA 13 nightly, multiarch manifest pinned for B300.
 dsv41flash-fp4-b300-sglang-agentic-dspark:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260922-582389ce@sha256:35ea4d321b0735051dcce2599fc495853ad1ef30f6a1908f0a75e74204362d14
+  image: lmsysorg/sglang:nightly-dev-cu13-20261004-295a53c9@sha256:7cfd4dd8e0e902d2a3c0560565963c4e172091e1586aaed15ff5a0a131394d34
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:b300-dsxe
@@ -9222,8 +9220,9 @@ dsv41flash-fp4-b300-sglang-agentic-dspark:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b300-fp4-mtp/agentic.yaml }
-      - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b300-fp4-mtp/agentic.yaml }
+      - { tp: 4, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 4], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b300-fp4-mtp/agentic.yaml }
+      - { tp: 2, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 4, 16, 24, 32, 48], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b300-fp4-mtp/agentic.yaml }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [192], router: { name: sglang-router, version: "0.3.2" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/b300-fp4-mtp/agentic.yaml }
 
 # Official SGLang nightly; TP4 C1/C2 use GPU Engram, other points use host DRAM.
 dsv41flash-fp4-gb300-sglang-agentic-dspark:

inferencex-e2e/perf-changelog.yaml

@@ -9507,3 +9507,13 @@
     - "Add Qwen3.5 FP8 GB300 AgentX PP4 prefill and DEP16 decode variants with MTP6, HiCache, and decode sequence-length splits 1."
     - "Select five disaggregated configurations and three existing aggregate configurations using the SGLang nightly image and pinned Dynamo revision."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3840
+
+- config-keys:
+    - dsv41flash-fp4-b300-sglang-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Repin DeepSeek-V4.1-Flash FP4 B300 SGLang AgentX from nightly-dev-cu13-20260922-582389ce to nightly-dev-cu13-20261004-295a53c9, pinned by multiarch digest, and publish nine points: TP4/EP1 at concurrency 1 and 4, TP2/EP1 at concurrency 1, 4, 16, 24, 32 and 48, and aggregated TP4/EP4/DP4 at concurrency 192."
+    - "For the TP arms, switch to EP1, use 16K prefill chunks, and set point-specific prefill-decode scheduling."
+    - "Add an aggregated DP4 C192 point with four internal TP4/EP4 data-parallel ranks behind sglang-router 0.3.2, using cache-aware routing and correlation-ID affinity."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3744