[NVIDIA][AgentX] Move GB300 DSV4.1-Flash SGLang to the 20261030 nightly with the tuned low-latency settings and TP4 c1 megakernel / [NVIDIA][AgentX] GB300 DSV4.1-Flash SGLang 升级到 20261030 nightly，采用调优后的低延迟配置并在 TP4 c1 启用 megakernel

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4-mtp/agentic.yaml

@@ -64,6 +64,8 @@ override_tp2_c16:
       args:
         tensor-parallel-size: 2
         expert-parallel-size: 2
+        # auto resolves to MegaAttention, which zero-pads per-rank heads to 64 (2x at TP2, 4x at TP4).
+        dsv4-attn-backend: flashmla
         <<: &tp2_bounded_replay
           chunked-prefill-size: 16384
           enable-decoder-swa-bounded-replay: true
@@ -90,6 +92,7 @@ override_tp2_c32:
       args:
         tensor-parallel-size: 2
         expert-parallel-size: 2
+        dsv4-attn-backend: flashmla
         <<: *tp2_bounded_replay
         prefill-decode-interval: 48
         tokenizer-worker-num: 4
@@ -187,14 +190,31 @@ override_tp4_c1:
       args:
         tensor-parallel-size: 4
         expert-parallel-size: 1
-        chunked-prefill-size: 16384
-        prefill-decode-interval: 0
-        max-running-requests: 1
+        mem-fraction-static: 0.91
+        kv-cache-dtype: fp8_e4m3
+        dsv4-attn-backend: flashmla
+        moe-a2a-backend: none
+        enable-nvidia-megakernel: true
+        startup-weight-load-mode: serial
         enable-decoder-swa-bounded-replay: true
-        cuda-graph-backend-prefill: disabled
+        context-length: 1048576
+        max-running-requests: 1
+        cuda-graph-max-bs-decode: 1
+        prefill-decode-interval: 16
+        schedule-policy: lpm
+        swa-prefix-tails: 128
+        chunked-prefill-size: 16384
+        enable-metrics-for-all-schedulers: true
       env:
-        SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: '0'
-        SGLANG_RAGGED_VERIFY_MODE: static
+        SGLANG_SPEC_TP_SYNC: rng,-dspark-accept-sample
+        SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: '1'
+        SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT: per_rank
+        SGLANG_NVIDIA_MEGAKERNEL_SLOTS: hop_a,moe,gate,attn,embed,two_shot,markov,draft_lm_head,aux_mean_stats,gate_wide,shared_ll,hop_a_batch2,engram,attn_qkv_a,attn_qkv_a_bf16,index_q_mxfp8_in,attn_indexer,hc_prenorm,hop_b,hop_b_wide,early_trigger,hop_b_start_trigger,draft,draft_moe,draft_gate,draft_tail,draft_post,draft_seam,draft_attn,draft_qfront,draft_wo_a,draft_swapab,draft_qkv_a,draft_prenorm,draft_qkv_a_ll,draft_wo_b_ll,sched_ahead,req_pos_once,kv_before_wait,target_lm_head,hop_b_quad_trigger,lm_head_pinned_tactic
+        SGLANG_ENABLE_NV_MEGA_RUNTIME_OPTIMIZATION: '1'
+        SGLANG_SIMULATE_ACC_GREEDY: '0'
+        SGLANG_OPT_DSV41_DSPARK_SIM_DRAW_AHEAD: '1'
+        NCCL_CUMEM_ENABLE: '0'
+        NCCL_MNNVL_ENABLE: '0'
   benchmark:
     env:
       CONC: '1'
@@ -207,6 +227,7 @@ override_tp4_c4:
       args:
         tensor-parallel-size: 4
         expert-parallel-size: 1
+        dsv4-attn-backend: flashmla
         chunked-prefill-size: 16384
         prefill-decode-interval: 0
         swa-prefix-tails: 4096
@@ -228,6 +249,7 @@ override_tp4_c16:
       args:
         tensor-parallel-size: 4
         expert-parallel-size: 1
+        dsv4-attn-backend: flashmla
         chunked-prefill-size: 16384
         prefill-decode-interval: 192
         tokenizer-worker-num: 4

inferencex-e2e/configs/nvidia-master.yaml

@@ -8024,9 +8024,9 @@ minimaxm31flash-fp4-b300-sglang-agentic-dspark:
       - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [48], srt-recipe: 'agentic-tp2-spf-interval.yaml:override_tp2_c48_spf_interval24' }
       - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [50], srt-recipe: 'agentic-tp2-spf-interval.yaml:override_tp2_c50_spf_interval20' }
 
-# Official SGLang nightly; TP4 C1 uses GPU Engram, other points use host DRAM.
+# Official SGLang nightly; every point uses the host-DRAM Engram table.
 dsv41flash-fp4-gb300-sglang-agentic-dspark:
-  image: lmsysorg/sglang:nightly-dev-cu13-20261021-c1c287a4@sha256:b0473834763b2753e069cb96c1acaf235ba7fd632ccd9bbeb44da809c437fdeb
+  image: lmsysorg/sglang:nightly-dev-cu13-20261030-302b78cc@sha256:d4b7701dc94d287e4fc135dfb4c1a88df0778a936255021aeb26eb54d396e579
   model: deepseek-ai/DeepSeek-V4.1-Flash
   model-prefix: dsv41flash
   runner: cluster:gb300-nv

inferencex-e2e/perf-changelog.yaml

@@ -11760,3 +11760,12 @@
     - "Run the Dynamo frontend, etcd, nats and the Mooncake master on the benchmark client node and disable the OTel collector on every disagg point."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4632
   scenario-type: [agentic-coding]
+
+- config-keys:
+    - dsv41flash-fp4-gb300-sglang-agentic-dspark
+  description:
+    - "Update the SGLang image from nightly-dev-cu13-20261021-c1c287a4 to nightly-dev-cu13-20261030-302b78cc, pinned by digest. The image includes sgl-project/sglang#46404 (DSpark simulated-acceptance bonus drawn at the simulated row), #46840 (simulated acceptance keeps the real verify accept cost), #46891 (no host sync in DSpark draft-block metadata) and sgl-nvidia-mega 54b428921749 (megakernel API 2, sgl-project/sglang#47145)."
+    - "TP4 c1 runs the NVIDIA megakernel (fp8_e4m3 KV cache, mem-fraction-static 0.91, decode graph batch 1, Engram host table) with SGLANG_ENABLE_NV_MEGA_RUNTIME_OPTIMIZATION=1 (sgl-project/sglang#47105), which enables the DSpark and Qwen3.5/EAGLE megakernel runtime optimizations; DSpark simulated draws run ahead of the draft launch (SGLANG_OPT_DSV41_DSPARK_SIM_DRAW_AHEAD)."
+    - "TP4 c1/c4/c16 and TP2/EP2 c16/c32 pin dsv4-attn-backend flashmla; auto resolves to MegaAttention, which zero-pads per-rank attention heads to 64. DP-attention points are unchanged."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4603
+  scenario-type: [agentic-coding]