From f35d7bef702578ec645b2cbea2b1095c53674248 Mon Sep 17 00:00:00 2001 From: Klaud-Cold Date: Fri, 11 Sep 2026 00:53:05 +0000 Subject: [PATCH 1/2] config: update dsr1-fp4-b200-sglang SGLang image to v0.5.19-cu130 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Move the B200 DeepSeek-R1-0528 NVFP4 single-node SGLang recipe from lmsysorg/sglang:v0.5.16-cu130 (build commit sgl-project/sglang@fdebc938f7f4d16fe6b9f55dcd9a767cf0899ea1, CUDA 13.0.1, FlashInfer 0.6.14, sgl-kernel 0.4.5) to the v0.5.19 release image lmsysorg/sglang:v0.5.19-cu130 (Docker Hub digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9, build commit sgl-project/sglang@0bcd822377da7b5718e674eaf9c870d349424dd1 = tag v0.5.19, CUDA 13.0.3, FlashInfer 0.6.18, sgl-kernel 0.4.6.post1). Model, TP4/EP1 and TP4/EP4 DP-attention topologies, the 8k1k workload, concurrency ranges and benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh are unchanged; the dsr1-fp4-b200-sglang-mtp sibling is not touched. 将 B200 DeepSeek-R1-0528 NVFP4 单节点 SGLang 配方的镜像从 lmsysorg/sglang:v0.5.16-cu130（构建提交 sgl-project/sglang@fdebc938f7f4d16fe6b9f55dcd9a767cf0899ea1，CUDA 13.0.1， FlashInfer 0.6.14，sgl-kernel 0.4.5）切换到 v0.5.19 发布版 lmsysorg/sglang:v0.5.19-cu130（Docker Hub 摘要 sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9， 构建提交 sgl-project/sglang@0bcd822377da7b5718e674eaf9c870d349424dd1 = 标签 v0.5.19，CUDA 13.0.3，FlashInfer 0.6.18，sgl-kernel 0.4.6.post1）。 模型、TP4/EP1 与 TP4/EP4 DP-attention 拓扑、8k1k 工作负载、并发范围以及 benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh 均保持不变； 未改动 dsr1-fp4-b200-sglang-mtp 同族配方。 Co-Authored-By: Claude Fable 5.1 --- configs/nvidia-master.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 9eeeffd50..9db3aa21d 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -860,7 +860,7 @@ dsr1-fp8-b300-dynamo-trt: ep: 8 dp-attn: true dsr1-fp4-b200-sglang: - image: lmsysorg/sglang:v0.5.16-cu130 + image: lmsysorg/sglang:v0.5.19-cu130 model: nvidia/DeepSeek-R1-0528-FP4-V2 model-prefix: dsr1 runner: cluster:b200-nscale From 25f8f98fa258680daf485f6d6b60084eb961d24b Mon Sep 17 00:00:00 2001 From: Klaud-Cold Date: Fri, 11 Sep 2026 03:05:03 +0000 Subject: [PATCH 2/2] changelog: record dsr1-fp4-b200-sglang image update to v0.5.19-cu130 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Append the perf-changelog entry for moving the B200 DeepSeek-R1-0528 NVFP4 single-node SGLang recipe to lmsysorg/sglang:v0.5.19-cu130 (PR #2989). The entry selects the whole family without scenario, eval-selection or append-only modifiers; all prior bytes are preserved. 为将 B200 DeepSeek-R1-0528 NVFP4 单节点 SGLang 配方切换到 lmsysorg/sglang:v0.5.19-cu130（PR #2989）追加 perf-changelog 条目。 该条目选择整个配方族，不带 scenario、评测选择或 append-only 修饰符； 所有既有字节均保持不变。 Co-Authored-By: Claude Fable 5.1 --- perf-changelog.yaml | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index f85ee11f6..4bc0ff86c 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -7162,3 +7162,10 @@ description: - "Use the pinned SGLang nightly-dev-cu13-20260901-07c8f729 image with FlashInfer 0.6.18, which includes the BF16 TRTLLM MoE allocation fix for small-batch Blackwell execution. Model, TP8, HiCache, MTP settings and concurrency grid are unchanged." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2829 + +- config-keys: + - dsr1-fp4-b200-sglang + description: + - "Update SGLang image from lmsysorg/sglang:v0.5.16-cu130 (build commit sgl-project/sglang@fdebc938f7f4d16fe6b9f55dcd9a767cf0899ea1, CUDA 13.0.1, FlashInfer 0.6.14, sgl-kernel 0.4.5) to the v0.5.19 release image lmsysorg/sglang:v0.5.19-cu130 (digest sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9, build commit sgl-project/sglang@0bcd822377da7b5718e674eaf9c870d349424dd1, CUDA 13.0.3, FlashInfer 0.6.18, sgl-kernel 0.4.6.post1)." + - "benchmarks/single_node/fixed_seq_len/dsr1_fp4_b200.sh is unchanged: nvidia/DeepSeek-R1-0528-FP4-V2 with modelopt_fp4, trtllm_mla attention, flashinfer_trtllm MoE, fp8_e4m3 KV cache, TP4/EP1 concurrency 1-32 and TP4/EP4 DP-attention concurrency 64-256 on the 8k1k workload." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2989