From 3195f3d94701dec7a6f76d9a6a648bdfeb1f7fa5 Mon Sep 17 00:00:00 2001
From: Oseltamivir <58582368+Oseltamivir@users.noreply.github.com>
Date: Tue, 14 Jul 2026 14:59:21 +0800
Subject: [PATCH] feat(dsv4): densify B300 sglang 1k1k sweep, re-run B300 trt
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- dsv4-fp4-b300-sglang: the 1k1k search space sampled only conc {1, 32,
  512, 8192}, leaving the ~50 tok/s/user frontier (conc 64-256) unsampled
  (933 tok/s/GPU vs B200's 1359). Add the DP-attn flashinfer band conc
  64-512 and extend megamoe to 2048-8192; script selects launch profiles
  by CONC range instead of exact match.
- dsv4-fp4-b300-trt(+mtp): re-run (no config change) so the density check
  compares same-pass data across frameworks.

中文：加密 B300 SGLang 1k1k 扫描点（DP-attn 64-512、megamoe 2048-8192，
脚本按 CONC 区间选择启动配置），并重跑 B300 TRT 以便同批次对比密度。
---
 .../fixed_seq_len/dsv4_fp4_b300_sglang.sh         |  4 ++--
 configs/nvidia-master.yaml                        | 12 ++++++++----
 perf-changelog.yaml                               | 15 +++++++++++++++
 3 files changed, 25 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_sglang.sh b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_sglang.sh
index b451dee0d5..553ad7279e 100755
--- a/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_sglang.sh
+++ b/benchmarks/single_node/fixed_seq_len/dsv4_fp4_b300_sglang.sh
@@ -53,7 +53,7 @@ start_gpu_monitor --output "$PWD/gpu_metrics.csv"
 # SWA ratio: 1k inputs need more SWA cache headroom than 8k inputs; 0.5 was
 # tuned empirically for the 1k1k recipe, while 0.1 is the cookbook default.
 
-if [ "$CONC" = "1" ] || [ "$CONC" = "32" ]; then
+if [ "$CONC" -le 32 ]; then
     # TP-only, no DP attention
     MEM_FRACTION_STATIC=0.90
     SWA_FULL_TOKENS_RATIO=$([[ "$ISL" == "1024" ]] && echo 0.5 || echo 0.1)
@@ -63,7 +63,7 @@ if [ "$CONC" = "1" ] || [ "$CONC" = "32" ]; then
         --disable-flashinfer-autotune
     )
 
-elif [ "$CONC" = "512" ]; then
+elif [ "$CONC" -ge 64 ] && [ "$CONC" -le 512 ]; then
     # DP attention, flashinfer_mxfp4
     export SGLANG_OPT_SWA_EVICT_DROP_PAGE_MARGIN=1
     MEM_FRACTION_STATIC=0.94
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f90c510186..ade57ec489 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -2002,8 +2002,8 @@ dsv4-fp4-b300-sglang:
   multinode: false
   # Recipes are selected inside benchmarks/single_node/dsv4_fp4_b300_sglang.sh
   # by CONC:
-  #   CONC 1|32:         TP-only, flashinfer_mxfp4
-  #   CONC 512:          DP-attn, flashinfer_mxfp4
+  #   CONC <=32:         TP-only, flashinfer_mxfp4
+  #   CONC 64-512:       DP-attn, flashinfer_mxfp4
   #   CONC 2048-8192:    DP-attn, megamoe
   # ep is implicit in sglang: --moe-a2a-backend megamoe forces ep_size=tp_size,
   # while low-latency leaves ep_size at the default of 1.
@@ -2014,8 +2014,12 @@ dsv4-fp4-b300-sglang:
       search-space:
       - { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
       - { tp: 4, ep: 1, conc-start: 32, conc-end: 32 }
-      - { tp: 4, ep: 1, dp-attn: true, conc-start: 512, conc-end: 512 }
-      - { tp: 8, ep: 8, dp-attn: true, conc-start: 8192, conc-end: 8192 }
+      # Densified mid band: the previous single conc-512 point left the
+      # ~50 tok/s/user frontier region (conc 64-256) unsampled at 1k1k.
+      - { tp: 4, ep: 1, dp-attn: true, conc-start: 64, conc-end: 512 }
+      # Megamoe band extended down to the script's existing 2048/4096 profiles
+      # (previously only 8192 at 1k1k).
+      - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 8192 }
     - isl: 8192
       osl: 1024
       search-space:
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 57d510dd15..613ee7f933 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4750,3 +4750,18 @@
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928"
     - "6 topologies across 1k/1k and 8k/1k: 1P1D TP4 STP + wide-EP (DEP4 prefill / DEP16 decode) from 1P1D up to 8P1D, recipes under benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2137
+
+- config-keys:
+    - dsv4-fp4-b300-sglang
+  description:
+    - "Densify the 1k1k sweep: the previous search space sampled only conc {1, 32, 512, 8192}, leaving the ~50 tok/s/user frontier region unsampled and noisy (933 tok/s/GPU vs B200's 1359 at 50 tok/s/user)"
+    - "Add the DP-attn flashinfer_mxfp4 band conc 64-512 (was a single 512 point) and extend the megamoe band down to conc 2048-8192 (was a single 8192 point), reusing the existing CONC 2048/4096 launch profiles in dsv4_fp4_b300_sglang.sh"
+    - "dsv4_fp4_b300_sglang.sh: select launch profiles by CONC range (<=32 TP-only, 64-512 DP-attn flashinfer) instead of exact match; megamoe profiles unchanged"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2192
+
+- config-keys:
+    - dsv4-fp4-b300-trt
+    - dsv4-fp4-b300-trt-mtp
+  description:
+    - "Re-run B300 dsv4 TRT-LLM configs (no config change, last run 2026-06-11) so the 1k1k frontier density check around 50 tok/s/user compares same-pass data across frameworks"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2192