[Klaud Cold] Update qwen3.5-fp4-b300-sglang (+mtp) SGLang image to v0.5.12-cu130

[functionstackx]

Summary

Update SGLang image from v0.5.11-cu130 (5d old) to v0.5.12-cu130

Recipes touched: `qwen3.5-fp4-b300-sglang`, qwen3.5-fp4-b300-sglang-mtp

Test plan

• full-sweep-enabled sweep passes.

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sglang/tree/main/docs_new

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26009982423
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26009982423

[claude]

LGTM, straightforward SGLang image version bump (v0.5.11-cu130 → v0.5.12-cu130) with matching changelog entry.

Extended reasoning...

Overview

This PR bumps the SGLang container image for two recipes (qwen3.5-fp4-b300-sglang and qwen3.5-fp4-b300-sglang-mtp) from v0.5.11-cu130 to v0.5.12-cu130 in .github/configs/nvidia-master.yaml, and appends a corresponding entry to perf-changelog.yaml linking back to this PR.

Security risks

None. This is a pinned upstream image version tag bump in CI config. The image source (lmsysorg/sglang) is unchanged, only the version tag advances by a patch release.

Level of scrutiny

Low. This is a mechanical, config-only change with no logic, no auth/crypto, and no production code paths. It follows the exact same pattern as numerous prior image-bump PRs in this repo (e.g. the recent #1458 entry directly above it in the changelog).

Other factors

The PR carries the full-sweep-enabled label so CI sweeps will validate the new image. No bugs were found by the bug hunting system, and there are no outstanding human reviewer comments — only the standard recipe-reminder bot messages.

[functionstackx]

Same vision-encoder workaround as PRs #1422 / #1451

Confirmed same flash-attn-4 cute sm_103 assertion in the Qwen-3.5-VL vision encoder — all 24 8k1k jobs failed with flash_fwd_sm100.py:162 AssertionError (1k1k may survive since vision encoder isn't exercised during warmup). Same family as the bf16 (#1422) and fp8 (#1451) siblings.

Added --mm-attention-backend triton_attn to both qwen3.5_fp4_b300.sh and qwen3.5_fp4_b300_mtp.sh. Pushed as 108982c2. Text decoder stays on trtllm_mha.

Upstream / root cause cross-links

• Issue: sgl-project/sglang#25564 — cutedsl Arch enum aliasing on non-cu13 path collapses sm_100..sm_110f to exclude sm_103
• Suggested durable fix: bump sglang's nvidia-cutlass-dsl dep to [cu13] extra (@mmangkad comment)
• Alternate fix in flight: Dao-AILab/flash-attention#2572 (awaiting @tridao)

Once either upstream lands, this workaround can come back out.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26009984417
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26009984417

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26047575440
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26047575440

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26047575440
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26047575440

[functionstackx]

Handing off to @Oseltamivir — tracked alongside 7 other stuck Klaud-Cold PRs in #1511. /loop will stop auto-retrying this one.

AI-generated via Claude Code /loop.

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26144009563
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26144009563

[functionstackx]

/reuse-sweep-run

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=26146822534
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=26146822534