From 0762e65d6101bd1551c3eee72dfb4913d1f8cd7d Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Sun, 20 Sep 2026 12:45:27 +0000 Subject: [PATCH 1/3] perf(mi325x): update dsr1-fp8-mi325x-sglang SGLang image to v0.5.20-rocm720-mi30x MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Update the dsr1-fp8-mi325x-sglang master image from lmsysorg/sglang:v0.5.19-rocm700-mi30x to lmsysorg/sglang:v0.5.20-rocm720-mi30x (digest sha256:4abecde156b531809c7f597e501930ea2023deb2c822b69f92efb66f3fdada86). SGLang v0.5.20 retired the rocm700 MI30x image lane, so the ROCm 7.2 MI30x variant already used by sibling MI325X SGLang recipes is selected. The recipe renames --cuda-graph-max-bs to --cuda-graph-max-bs-decode because v0.5.20 removed the deprecated alias (sgl-project/sglang#38375); the value 128 and all other flags, model, topology, workload and evals are unchanged. 将 dsr1-fp8-mi325x-sglang 的主镜像从 lmsysorg/sglang:v0.5.19-rocm700-mi30x 更新为 lmsysorg/sglang:v0.5.20-rocm720-mi30x。SGLang v0.5.20 已停止发布 rocm700 MI30x 镜像，因此选用同一 MI325X 集群上其他 SGLang 配方已使用的 ROCm 7.2 MI30x 变体。由于 v0.5.20 移除了已弃用的 --cuda-graph-max-bs 别名， 配方脚本改用 --cuda-graph-max-bs-decode，数值 128 及其他参数、模型、拓扑、 负载与评测均保持不变。 AI model disclosure: prepared by claude-fable-5-1 (Claude Fable 5.1) as Klaud Cold. Co-Authored-By: Claude Fable 5.1 --- benchmarks/single_node/fixed_seq_len/dsr1_fp8_mi325x.sh | 2 +- configs/amd-master.yaml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/benchmarks/single_node/fixed_seq_len/dsr1_fp8_mi325x.sh b/benchmarks/single_node/fixed_seq_len/dsr1_fp8_mi325x.sh index 0f6a801136..78a60cec69 100644 --- a/benchmarks/single_node/fixed_seq_len/dsr1_fp8_mi325x.sh +++ b/benchmarks/single_node/fixed_seq_len/dsr1_fp8_mi325x.sh @@ -37,7 +37,7 @@ python3 -m sglang.launch_server \ --model-path=$MODEL --host=0.0.0.0 --port=$PORT --trust-remote-code \ --tensor-parallel-size=$TP \ --mem-fraction-static=0.8 \ ---cuda-graph-max-bs=128 \ +--cuda-graph-max-bs-decode=128 \ --chunked-prefill-size=131072 \ --num-continuous-decode-steps=4 \ --max-prefill-tokens=131072 \ diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 9b4fcadeb7..11a38ee897 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -78,7 +78,7 @@ dsr1-fp8-mi300x-sglang: - { tp: 8, conc-start: 4, conc-end: 64 } dsr1-fp8-mi325x-sglang: - image: lmsysorg/sglang:v0.5.19-rocm700-mi30x + image: lmsysorg/sglang:v0.5.20-rocm720-mi30x model: deepseek-ai/DeepSeek-R1-0528 model-prefix: dsr1 runner: mi325x From 0f3172bf06652264470f2fcae9c9be0d27a8092f Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Sun, 20 Sep 2026 13:20:21 +0000 Subject: [PATCH 2/3] perf(mi325x): add changelog entry for dsr1-fp8-mi325x-sglang v0.5.20 image MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Append the perf-changelog.yaml entry for the dsr1-fp8-mi325x-sglang SGLang image update after the updated-image smoke run passed. 在更新镜像的 smoke 运行通过后，为 dsr1-fp8-mi325x-sglang 的 SGLang 镜像更新 追加 perf-changelog.yaml 条目。 Co-Authored-By: Claude Fable 5.1 --- perf-changelog.yaml | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index c5b4309f5f..bda8ea6615 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -8455,3 +8455,9 @@ description: - "Update B200 vLLM AgentX to DSpark6 and a new image with TP8 and DEP8 configurations." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3274 + +- config-keys: + - dsr1-fp8-mi325x-sglang + description: + - "Update SGLang image from v0.5.19-rocm700-mi30x to v0.5.20-rocm720-mi30x and rename --cuda-graph-max-bs to its decode flag." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3318 From 89cf7a165dc472b73b1b84032ebd442a44dd3cdf Mon Sep 17 00:00:00 2001 From: "claude[bot]" <41898282+claude[bot]@users.noreply.github.com> Date: Sun, 20 Sep 2026 13:20:51 +0000 Subject: [PATCH 3/3] perf(mi325x): shorten dsr1-fp8-mi325x-sglang changelog description MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Keep the perf-changelog.yaml description within 120 characters. 将 perf-changelog.yaml 的描述控制在 120 个字符以内。 Co-Authored-By: Claude Fable 5.1 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index bda8ea6615..35e5afc5c8 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -8459,5 +8459,5 @@ - config-keys: - dsr1-fp8-mi325x-sglang description: - - "Update SGLang image from v0.5.19-rocm700-mi30x to v0.5.20-rocm720-mi30x and rename --cuda-graph-max-bs to its decode flag." + - "Update SGLang image from v0.5.19-rocm700-mi30x to v0.5.20-rocm720-mi30x and rename the CUDA-graph max-bs flag." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3318