[Feature] Refactor Refit Unit Tests with SHA256 Verification

[zhaochenyang20]

Checklist

• If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Motivation

Recently, the SHA256 checksum has been a powerful tool to verify whether the weights have been correctly updated and remain bit-for-bit identical to the target state. @dreamyang-liu

Currently, SGLang supports three major weight refit strategies for RL systems:

1. update_weights_from_disk
2. update_weights_from_tensor
3. update_weights_from_distributed

However, the existing unit tests for these interfaces could be more robust. We believe using SHA256 to verify the weight buffers after refitting will provide a much stricter guarantee of correctness, preventing silent mismatches in RL rollout and training loops.

In this sense, we can try to:

1. Refactor LLM Refit Tests: Rewrite the unit tests for the three LLM refit interfaces using SHA256 verification.
2. Correctness Guarantee: Ensure that after any refit operation, the in-memory weights match the source perfectly.

Related resources

https://docs.sglang.io/advanced_features/sglang_for_rl.html#open-to-use-refit-functionality

#18306

#18891

[aeft]

Hi, I'd like to take this on. Expect a PR in a day.

Similar to #18306, I plan to add a get_weights_checksum Engine method and HTTP API for LLM. The core SHA256 computation will live in model_runner.py, following the same calling chain as get_weights_by_name (engine -> tokenizer_manager -> scheduler -> tp_worker -> model_runner).

I noticed #18306 implemented compute_weights_checksum in multimodal_gen/runtime/loader/weight_utils.py. I'll need the same function for LLM. Since #18306 hasn't been merged yet, I plan to implement it independently first. Once #18306 is merged, we can consolidate them into a shared location. Does this sound reasonable, or would you prefer a different approach?

I'll start with update_weights_from_tensor as the first PR to demonstrate the pattern. Then extend to from_disk and from_distributed in follow-up PRs. Or would you prefer them in a single PR instead?