[sglang-miles] Fix post_process_weights deadlock in RL in-place mode

[ByronHsu]

Summary

Restore the pause-aware locking behavior for post_process_weights on the sglang-miles-v0.5.13 branch.

This is the post_process_weights hunk from the earlier pause-aware deadlock fix in #22754, which appears to have been lost during the v0.5.13 tokenizer_control_mixin.py refactor.

sglang-miles-v0.5.12 skipped model_update_lock.writer_lock when the engine was already paused. During the v0.5.13 refactor to tokenizer_control_mixin.py, post_process_weights became an unconditional writer-lock acquisition. That can deadlock with pause_generation(mode="in_place"):

active /generate holds model_update_lock.reader_lock
pause_generation(in_place) preserves the request and pauses progress
post_process_weights waits for writer_lock
continue_generation is only called after post_process_weights

The result is that health generation stops making progress and post-load processing can time out.

Fix

• Check self.is_pause under is_pause_cond.
• Use model_update_lock.writer_lock only when the engine is not paused.
• Use nullcontext() while paused, matching the pause-aware behavior from the earlier miles branch.

Validation

python3 -m py_compile python/sglang/srt/managers/tokenizer_control_mixin.py

---

CI States

Latest PR Test (Base): ❌ Run #28393955205
Latest PR Test (Extra): ❌ Run #28393954957

[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!