fix(exl3): satisfy PR 228 release checks

[malaiwah]

Problem

PR #228's self-hosted pre-commit run timed out, but a local exact-head replay exposed real changed-file failures before the unrelated repository-wide baseline failures: Ruff formatting, EXL3 technical-term spelling, missing SPDX, mypy narrowing/return annotations, a forbidden stdlib re import, and newly introduced torch.cuda API calls.

Change

• Apply Ruff/SPDX output and allowlist the EXL3 suh/SUH and subtiles terms.
• Narrow optional metadata/config values instead of suppressing mypy.
• Use QuantizationMethods for the quantization-config override contract.
• Replace PR-introduced CUDA calls with torch.accelerator equivalents.
• Keep the test harness on the same accelerator API and assert its optional captured kwargs before indexing.

No EXL3 math, launch geometry, routing, capacity policy, cache key, or checkpoint layout changes.

Validation

• Exact changed-file pre-commit hook set: all passed, including Ruff, typos, mypy 3.10, SPDX, forbidden imports, and the new-CUDA-call guard.
• Focused CPU suite: 105 passed (test_exl3.py, online cache, prefill plan, warmup, and quantization config args).
• Repository-wide --all-files remains red on numerous pre-existing, unrelated Minimax/MLA/DeepSeek/parser/type errors; those are outside this repair and explain why changed-file evidence is stated explicitly.

This PR targets #228's head branch so merging it updates #228 without replaying or widening its feature commits.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• dev/*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: ab9b71e7-f639-4747-8540-6d8f05a2ac11

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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.}

[malaiwah]

Runtime requalification passed for this repair applied to the current release integration.

Exact composition

• B12X [gg-rebased] perf(dspark): port adaptive draft capacity experiments #126: c0a36cec766ce529d203f535fcaaa9c76338a551
• this repair: 27bb05185a47847e39e34dec17e3bce54c5379e5
• release integration base: r31-vllm258 / sha256:6152ba038ff84b7ae1417d23f314147c71f1721db7d66efe870e0daee1315243
• image: b12x-pr126-vllm-pr306-gate:20260814-v5
• image digest: sha256:abd3c39d481245e2fec139753da0c5c225eef3858245eea4cd77cc10ee29c785

The PR's runtime patch was applied onto the release integration's versions of exl3.py and kernel_warmup.py, rather than replacing them with the older #228 snapshots. This preserves the integration's successor changes, including the #258 prompt-logprobs environment contract. Deployed repaired hashes: config/quantization.py da0a291b…; exl3.py 3d68fffb…; exl3_online_cache.py 08f0fcbc…; kernel_warmup.py 4af35f71…. The retained integrated envs.py hash is 3eee46ff….

Gates

• changed-file pre-commit: every hook passed, including Ruff, formatting, typos, mypy 3.10, SPDX, forbidden imports, and the new-torch.cuda guard;
• focused CPU tests: 105 passed;
• fresh-cache 4x RTX PRO 6000 Blackwell Server Edition, TP4/DCP4, MTP3 startup: passed;
• full/piecewise CUDA graph capture: passed on all ranks (45 s, 0.17 GiB actual graph memory/rank);
• two API batteries: both passed arithmetic correctness, 2,050-token partial prefill, and concurrency 1/2/5/8;
• zero _pack_topk_routes* JIT events after startup;
• first battery raw seconds: 0.6345 arithmetic, 2.5777 first partial prefill, and 1.0496/1.3221/2.3083/2.7122 for concurrency 1/2/5/8. Raw evidence only; no performance claim.

Repository-wide pre-commit --all-files still fails on many unrelated files already present on #228's base branch (Minimax, MLA, DeepSeek, parsers, and global mypy/forbidden-import/new-CUDA checks). This repair intentionally does not absorb that unrelated repository cleanup. The current PR pre-run check is skipped only because my account lacks the required ready/verified label or four merged PRs; a maintainer must label or merge this repair into #228's head branch.