From 34299f7ab35a715c143f04ac9e6dc35e1084b641 Mon Sep 17 00:00:00 2001 From: Jiejing Zhang Date: Tue, 15 Sep 2026 07:02:36 +0000 Subject: [PATCH 1/2] perf: use HIP Top-K for GLM-5.2 on MI355X MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 关闭 SGLANG_OPT_USE_TOPK_V2，使 GLM-5.2 在 MI355X 上使用预编译 HIP Top-K 路径。 --- .../single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh | 5 +++-- perf-changelog.yaml | 9 +++++++++ 2 files changed, 12 insertions(+), 2 deletions(-) diff --git a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh index f46ad956e..86328e283 100644 --- a/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh +++ b/benchmarks/single_node/agentic/glm5.2_fp4_mi355x_sglang_mtp.sh @@ -61,8 +61,9 @@ export AIPERF_HTTP_TCP_USER_TIMEOUT=900000 # inter-turn idle gaps can reuse a socket exactly as the server closes it. # Outlast the client pool so the race cannot occur. export SGLANG_TIMEOUT_KEEP_ALIVE=900 -# PR #36684 & PR #36851 turned the v2 fused top-k on for GLM-5.x on ROCm. -export SGLANG_OPT_USE_TOPK_V2=true +# Use the precompiled HIP Top-K path; at c8 it improves P90 interactivity with +# essentially unchanged throughput per GPU compared with fused Top-K v2. +export SGLANG_OPT_USE_TOPK_V2=false # HiCache L2 (host DRAM), optionally extended with Mooncake L3. # KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=hicache or mooncake. diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 381227752..a26a0cb5d 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -7753,3 +7753,12 @@ - "Set a 64-token CUDA graph capture floor for B300 DSpark at c1/c2/c4; preserve c8+ sizing, full AgentX traces, AL 3.51, and Engram UVA" - "将 B300 DSpark 的 c1/c2/c4 CUDA graph 最小捕获范围设为 64 tokens；保持 c8+ 大小、完整 AgentX 轨迹、AL 3.51 和 Engram UVA" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3116 + +- config-keys: + - glm5.2-fp4-mi355x-sglang-agentic-mtp + scenario-type: + - agentic-coding + description: + - "Disable SGLANG_OPT_USE_TOPK_V2 so GLM-5.2 uses the precompiled HIP Top-K path on MI355X; at concurrency 8, P90 interactivity improves from 96.97 to 106.69 tok/s/user (+10.0%) while throughput per GPU changes from 9,553.2 to 9,539.0 tok/s (-0.15%)." + - "关闭 SGLANG_OPT_USE_TOPK_V2，使 GLM-5.2 在 MI355X 上使用预编译 HIP Top-K 路径；并发 8 时，P90 interactivity 从 96.97 提升到 106.69 tok/s/user（+10.0%），单卡吞吐从 9,553.2 变为 9,539.0 tok/s（-0.15%）。" + pr-link: TBD From aac8318502c4407a7c0722b9a1acf8a1b968e502 Mon Sep 17 00:00:00 2001 From: Jiejing Zhang Date: Tue, 15 Sep 2026 07:07:31 +0000 Subject: [PATCH 2/2] chore: link GLM-5.2 HIP Top-K changelog MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 将 GLM-5.2 HIP Top-K 性能记录关联到 PR #3117。 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index a26a0cb5d..172188f04 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -7761,4 +7761,4 @@ description: - "Disable SGLANG_OPT_USE_TOPK_V2 so GLM-5.2 uses the precompiled HIP Top-K path on MI355X; at concurrency 8, P90 interactivity improves from 96.97 to 106.69 tok/s/user (+10.0%) while throughput per GPU changes from 9,553.2 to 9,539.0 tok/s (-0.15%)." - "关闭 SGLANG_OPT_USE_TOPK_V2，使 GLM-5.2 在 MI355X 上使用预编译 HIP Top-K 路径；并发 8 时，P90 interactivity 从 96.97 提升到 106.69 tok/s/user（+10.0%），单卡吞吐从 9,553.2 变为 9,539.0 tok/s（-0.15%）。" - pr-link: TBD + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3117