[yushengsu-thu]

Motivation and Modification

Scenario: lora_a, lora_b, and lora_c are separate LoRA adapters sharing one base model. A has finished rollout and training; its updated weights will become lora_a@2. lora_b, and lora_c are still generating.

1. SGLang PR — allow publication alongside generation
   Most important file: tokenizer_control_mixin.py.
   Previously, updates acquired a writer lock while the engine was running, making them mutually exclusive with generation.
   For validated fresh-version publication, the PR selects a reader lock:

# Inside _weight_update_session_call()
if not is_paused:
    lock = (
        self.model_update_lock.reader_lock
        if self._weight_update_new_loras
        else self.model_update_lock.writer_lock
    )
    async with lock:
        results = await communicator(obj)

lora_b, and lora_c generation already uses reader locks, so publication can coexist with it.
Supporting changes make this safe:

• Register lora_a’s new version with ready=False, blocking premature sampling.
• Validate session targets, preserve received IPC tensors, and check completeness and TP worker results.
• Mark the adapter ready only after successful loading.
  _weight_update_new_loras contains validated pending adapters; a session ID alone does not enable this path. Base updates and same-name overwrites retain exclusive locking.

---

CI States

Latest PR Test (Base): ❌ Run #34296559933
Latest PR Test (Extra): ❌ Run #34296559792
Latest PR Test (AMD ROCm 7.2): ❌ Run #34296560027