From e9bee3539eb3a4f1e7bcbb62394a234bbc770a5a Mon Sep 17 00:00:00 2001 From: Yifan Qiao Date: Wed, 19 Aug 2026 10:23:09 +0000 Subject: [PATCH] perf(gb300): add AgentX aggregate TP4 points MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 为 GB300 AgentX 聚合式 TP4 扫描新增并发 1、2、4 和 6，同时保留现有并发 8 数据点。 --- configs/nvidia-master.yaml | 2 +- perf-changelog.yaml | 9 +++++++++ 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 238aaf320f..731bb63b79 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7787,7 +7787,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg: ep: 1 dp-attn: false - spec-decoding: mtp - conc-list: [8] + conc-list: [1, 2, 4, 6, 8] prefill: num-worker: 1 tp: 4 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 52c74bf168..b0a67a950f 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6195,3 +6195,12 @@ - "Use AITER INT4 quick-reduce, ptpc_fp8 online quantization excluding embeddings, lm_head, gates, and experts, an FP8 KV cache, and three-token MTP with golden synthetic acceptance length 2.99 for benchmark runs." - "Set max-num-seqs to twice the concurrency, select CUDA graph capture sizes by concurrency, and cap max-num-batched-tokens at 16384." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2576 + +- config-keys: + - dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg + scenario-type: + - agentic-coding + description: + - "Expand the GB300 AgentX aggregate TP4 sweep to concurrency 1, 2, 4, 6, and 8." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2676 + append-only: true