fix: resolve audited training and inference bugs

- reject prompts that encode to zero tokens in add_task instead of admitting a task whose prefill can never run, and surface empty-id run_batch calls as prompt_empty errors
- deliver the STOP stream callback when cancelling a live task so clients observe termination instead of hanging until socket timeout
- strip the torch.compile _orig_mod. prefix at every unwrap_model site and when loading checkpoints so FSDP state dicts and saved weights no longer leak the wrapper name into downstream keys
- reject online_* train strategies with nprocs > 1 at config validation time, explaining the NCCL all-gather deadlock they would otherwise hit mid-run
- apply the frequency penalty before temperature scaling (OpenAI semantics) so the penalty survives temperature=0 instead of being annihilated by the 1e8 logit blowup, and exclude penalty pipelines from the greedy fast path
- return logprobs from the raw pre-strategy distribution so they match training-side policy logprobs for PPO/GRPO importance ratios

astrai/config/train_config.py

@@ -213,8 +213,17 @@ class TrainConfig(BaseConfig):
 
     @model_validator(mode="after")
     def _validate_online_strategy(self) -> "TrainConfig":
-        if self.strategy.startswith("online_") and self.reward_model_fn is None:
-            raise ValueError(
-                f"reward_model_fn is required for online RL strategy {self.strategy!r}"
-            )
+        if self.strategy.startswith("online_"):
+            if self.reward_model_fn is None:
+                raise ValueError(
+                    f"reward_model_fn is required for online RL strategy "
+                    f"{self.strategy!r}"
+                )
+            if self.nprocs > 1:
+                raise ValueError(
+                    f"online RL strategy {self.strategy!r} requires single-process "
+                    f"training (nprocs=1): per-rank rollouts issue different "
+                    f"numbers of forward passes and desynchronize the "
+                    f"ddp/fsdp collectives, deadlocking NCCL"
+                )
         return self