build: CUTLASS DSL 4.7.1 to match B12X

[voipmonitor]

B12X master moved to nvidia-cutlass-dsl 4.7.1 (b12x 4bacd509, Luke). That blocks canonical image publishing: b12x requires nvidia-cutlass-dsl==4.7.1, but 4.6.2 is selected. FlashInfer pins the compiler set that B12X uses, so it moves with it.

The following move to 4.7.1:

• requirements.txt;
• the cu12/cu13 extras;
• ci/cuda-versions.json (dependency policy);
• ci/lil_wheels/runtime.lock.

The NGC PyTorch 26.08 builder ships 4.6.2, so ci/lil_wheels/Dockerfile first installs 4.7.1 with --require-hashes from the new ci/lil_wheels/cutlass-dsl.lock. Those hashes are the ones in blackwell-llm-docker's qwen38-runtime.lock.

Draft: this merges together with the vLLM, B12X-beta and docker changes of the same switch, after a GPU gate.

ci/lil_wheels/tests: 22 passed. ci/validate_cuda_versions.py: OK.

🤖 Generated with Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[voipmonitor]

Pre-commit on this PR fails on files it does not touch: end-of-file/whitespace fixes in scripts/b12x and docs/b12x/evidence, 410 files reformatted, mypy errors in flashinfer/b12x. None of this PR's six files are among them. The upstream Release workflow, which this fork's runners cannot serve, was queued by the ci/cuda-versions.json change and has been cancelled.

🤖 Generated with Claude Code