From 18d87ef4edfbe1ceda53e3dd67c998af0c3172ac Mon Sep 17 00:00:00 2001 From: hshrivastava-droid Date: Mon, 10 Aug 2026 14:06:04 -0700 Subject: [PATCH 1/3] perf(qwen3.5-fp4-b200-sglang-mtp): add TEP2 arm Add a TP2/EP2 arm to the Qwen3.5-397B-A17B NVFP4 B200 SGLang MTP 8k/1k sweep at concurrency 16/32/64, alongside the existing TP2/EP1 rows. --- configs/nvidia-master.yaml | 1 + perf-changelog.yaml | 11 +++++++++++ 2 files changed, 12 insertions(+) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 15b610cb4..c938a6f6f 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -1246,6 +1246,7 @@ qwen3.5-fp4-b200-sglang-mtp: search-space: - { tp: 4, ep: 1, conc-start: 4, conc-end: 4, spec-decoding: mtp } - { tp: 2, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } + - { tp: 2, ep: 2, conc-list: [16, 32, 64], spec-decoding: mtp } qwen3.5-fp8-b200-sglang-mtp: image: lmsysorg/sglang:v0.5.14-cu130 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 448540716..7c3fd9575 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5736,3 +5736,14 @@ - "Upgrade SGLang from nightly-dev-cu13-20260711-7de33ce8 to nightly-dev-cu13-20260719-99f5a6f4, the latest nightly before the scheduler WAR-barrier regression in SGLang #31687" - "Restore the pre-watchdog AgentX concurrency grid and use AIPerf's default 300-second per-trajectory idle-gap cap" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2319 + +- config-keys: + - qwen3.5-fp4-b200-sglang-mtp + scenario-type: + - fixed-seq-len + description: + - "Add a TEP2 arm (tp 2, ep 2) to the Qwen3.5-397B-A17B NVFP4 B200 SGLang MTP 8k/1k sweep at concurrency 16, 32, and 64" + - "Expert parallelism 2 measured against a same-node ep 1 reference at concurrency 32: +1.11% tok/s/GPU and +2.28% tok/s/user, non-overlapping on both axes across two samples" + - "Combined against the current published tp 2 ep 1 rows: 7534.5 tok/s/GPU at 120.16 tok/s/user (+5.85% / +5.17%) at concurrency 16, 10319.8 at 80.41 (+8.94% / +7.26%) at 32, and 13290.8 at 51.97 (+9.46% / +9.06%) at 64" + - "Correctness held on the promoted configuration: GSM8K strict-match 0.967-0.975 and flexible-extract 0.960-0.968 against a 0.94 threshold, coherence passed, decode CUDA graph active and MTP acceptance baseline-like with no backend fallback" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX From e36439994615e02aaaf042482234dc7060a6126d Mon Sep 17 00:00:00 2001 From: hshrivastava-droid Date: Mon, 10 Aug 2026 14:06:32 -0700 Subject: [PATCH 2/3] Update perf-changelog pr-link for #2550 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 7c3fd9575..73b805acb 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5746,4 +5746,4 @@ - "Expert parallelism 2 measured against a same-node ep 1 reference at concurrency 32: +1.11% tok/s/GPU and +2.28% tok/s/user, non-overlapping on both axes across two samples" - "Combined against the current published tp 2 ep 1 rows: 7534.5 tok/s/GPU at 120.16 tok/s/user (+5.85% / +5.17%) at concurrency 16, 10319.8 at 80.41 (+8.94% / +7.26%) at 32, and 13290.8 at 51.97 (+9.46% / +9.06%) at 64" - "Correctness held on the promoted configuration: GSM8K strict-match 0.967-0.975 and flexible-extract 0.960-0.968 against a 0.94 threshold, coherence passed, decode CUDA graph active and MTP acceptance baseline-like with no backend fallback" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2550 From f6c282cbf06a5c43dc4625e51d218a3664939ab8 Mon Sep 17 00:00:00 2001 From: hshrivastava-droid Date: Wed, 12 Aug 2026 11:28:35 -0700 Subject: [PATCH 3/3] chore(changelog): trim TEP2 entry description to two bullets MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 精简 perf-changelog 条目：仅保留 TEP2 组合与 #2205 NVFP4-V2 依赖两条说明。 --- perf-changelog.yaml | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 73b805acb..1d2980399 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5742,8 +5742,6 @@ scenario-type: - fixed-seq-len description: - - "Add a TEP2 arm (tp 2, ep 2) to the Qwen3.5-397B-A17B NVFP4 B200 SGLang MTP 8k/1k sweep at concurrency 16, 32, and 64" - - "Expert parallelism 2 measured against a same-node ep 1 reference at concurrency 32: +1.11% tok/s/GPU and +2.28% tok/s/user, non-overlapping on both axes across two samples" - - "Combined against the current published tp 2 ep 1 rows: 7534.5 tok/s/GPU at 120.16 tok/s/user (+5.85% / +5.17%) at concurrency 16, 10319.8 at 80.41 (+8.94% / +7.26%) at 32, and 13290.8 at 51.97 (+9.46% / +9.06%) at 64" - - "Correctness held on the promoted configuration: GSM8K strict-match 0.967-0.975 and flexible-extract 0.960-0.968 against a 0.94 threshold, coherence passed, decode CUDA graph active and MTP acceptance baseline-like with no backend fallback" + - "Add a TEP2 arm (tp 2, ep 2) to the qwen3.5-fp4-b200-sglang-mtp 8k/1k sweep at concurrency 16, 32, and 64" + - "Rides on the NVFP4-V2 checkpoint switch from #2205" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2550