Skip to content

[Klaud Cold] Update qwen3.5-fp8-h200-sglang-agentic-mtp SGLang image to v0.5.19-cu130 / 将 qwen3.5-fp8-h200-sglang-agentic-mtp 的 SGLang 镜像更新至 v0.5.19-cu130 #3198

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
adibarra merged 4 commits into from
Sep 17, 2026
Merged

[Klaud Cold] Update qwen3.5-fp8-h200-sglang-agentic-mtp SGLang image to v0.5.19-cu130 / 将 qwen3.5-fp8-h200-sglang-agentic-mtp 的 SGLang 镜像更新至 v0.5.19-cu130 #3198

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
01bfabe
perf: update qwen3.5-fp8-h200-sglang-agentic-mtp SGLang image to v0.5…
github-actions[bot] Sep 16, 2026
3d79a08
perf: add changelog entry for the qwen3.5-fp8-h200-sglang-agentic-mtp…
github-actions[bot] Sep 16, 2026
69d978f
Merge branch 'main' into klaud/auto-4b59b883da204568-747c6b1c3639f933
github-actions[bot] Sep 16, 2026
d057231
Merge remote-tracking branch 'origin/main' into codex/pr-3198-sync
adibarra Sep 17, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5130,7 +5130,7 @@ qwen3.5-fp8-h100-sglang-agentic:
# fixed-seq-len H200 MTP entry already runs conc 4-128 at TP8/EP8 versus H100's
# 4-32, so the GPU-resident arm extends to 24 and HiCache to 48.
qwen3.5-fp8-h200-sglang-agentic-mtp:
image: lmsysorg/sglang:nightly-dev-cu13-20260914-4358a161
image: lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9
model: Qwen/Qwen3.5-397B-A17B-FP8
model-prefix: qwen3.5
runner: cluster:h200-dgxc
Expand Down
7 changes: 7 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7979,3 +7979,10 @@
- "Rename --cuda-graph-max-bs to --cuda-graph-max-bs-decode: the 2026-09-15 SGLang nightly splits the flag into decode/prefill variants and rejects the old spelling as ambiguous (run 35013444957, canary c4)"
- "将 --cuda-graph-max-bs 改为 --cuda-graph-max-bs-decode：2026-09-15 SGLang nightly 将该参数拆分为 decode/prefill 两个变体，旧写法因歧义被拒绝（运行 35013444957，canary c4）"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3144

- config-keys:
- qwen3.5-fp8-h200-sglang-agentic-mtp
description:
- "Update SGLang image from lmsysorg/sglang:nightly-dev-cu13-20260914-4358a161 (2026-09-14 cu13 dev nightly, sgl-project/sglang@4358a161, a main snapshot 633 commits past the v0.5.19 release) to the digest-pinned v0.5.19 release image lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9 (Docker Hub push 2026-09-04T22:50Z, build commit sgl-project/sglang@0bcd8223 = tag v0.5.19, sglang-kernel 0.4.6.post1, flashinfer 0.6.18). Every flag and SGLANG_SIMULATE_ACC_* setting in benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh exists in v0.5.19 (SGLANG_ENABLE_SPEC_V2 only logs a removal warning at both commits); the release predates the 09-04..09-13 upstream HiCache and Qwen3.5 GDN fixes on main (sglang PRs 36267, 33922, 34820, 38481, 38350). The same script already passed a full H200 sweep on v0.5.19-cu130 for the TP8/EP1 HiCache sibling (#2966). Model, TP8/EP8 resident and HiCache DRAM arms, EAGLE MTP settings (3 steps, top-k 1, 4 draft tokens, golden AL 3.39), the with-subagents 256k corpus and the c4/8/12/16 grid are unchanged."
- "将 SGLang 镜像从 lmsysorg/sglang:nightly-dev-cu13-20260914-4358a161（2026-09-14 cu13 dev nightly，sgl-project/sglang@4358a161，比 v0.5.19 发布版领先 633 个提交的 main 快照）更新为按摘要固定的 v0.5.19 发布版镜像 lmsysorg/sglang:v0.5.19-cu130@sha256:d6e7288627be8b02be88e4bba38e73f6d50e2826869f753c13a4c4385ab3eda9（Docker Hub 推送时间 2026-09-04T22:50Z，构建提交 sgl-project/sglang@0bcd8223 即标签 v0.5.19，sglang-kernel 0.4.6.post1，flashinfer 0.6.18）。benchmarks/single_node/agentic/qwen3.5_fp8_h200_mtp.sh 使用的全部参数与 SGLANG_SIMULATE_ACC_* 设置在 v0.5.19 中均存在（SGLANG_ENABLE_SPEC_V2 在两个提交中都只打印移除警告）；发布版不含 09-04 至 09-13 期间 main 上的 HiCache 与 Qwen3.5 GDN 修复（sglang PR 36267、33922、34820、38481、38350）。同一脚本已在 TP8/EP1 HiCache 姊妹配方上以 v0.5.19-cu130 通过完整 H200 sweep（#2966）。模型、TP8/EP8 驻留与 HiCache DRAM 分支、EAGLE MTP 设置（3 步、top-k 1、4 个草稿 token、黄金 AL 3.39）、with-subagents 256k 语料及 c4/8/12/16 网格均保持不变。"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3198
Loading