[2/N][Sync sglang-miles] Support INT4 QAT for RL

[zyzshishui]

Motivation

Synced from https://github.com/sgl-project/sglang/pull/18639 and add tests with super tiny fix

Modifications

• Add post-process weights API path for restoring weights and running quantization post-processing after weight updates.
• Support asymmetric WNA16 compressed-tensors QAT wiring, including zero-point handling for Marlin MoE.
• Explicitly reject unsupported asymmetric WNA16 Triton MoE path.

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 026995f3fa

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Add admin auth to post-process weights endpoint

This route mutates model weights but is missing @auth_level(AuthLevel.ADMIN_OPTIONAL), so it falls back to AuthLevel.NORMAL. In the current auth logic, NORMAL endpoints are allowed without credentials when only admin_api_key is configured, which means an unauthenticated caller can trigger weight post-processing/restore operations and disrupt serving.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Avoid reprocessing all quantized modules indiscriminately

This applies process_weights_after_loading to every module that exposes it, but many quantization schemes are not idempotent (for example, compressed-tensors linear paths convert parameter types/layouts and expect load-time structures on first pass). Calling this endpoint after normal startup/update can therefore assert or repack already-converted weights, leading to failures or corrupted outputs unless processing is limited to modules explicitly restored to load format.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Post-process the same worker targeted by tensor updates

The new post-process path always runs on tp_worker, but tensor updates can target draft_worker when draft mode is enabled and disable_draft_model is unset/false. In that common flow, updated draft weights never receive the required post-processing, leaving draft execution with stale/incompatible weight layout.

Useful? React with 👍 / 👎.