Merge branch 'v0.9.2-dev_pp_bug' into 'v0.9.2-dev'

fix PP 场景 decode 阶段 token 被误丢弃导致卡住

See merge request dcutoolkit/deeplearing/vllm!363

vllm/v1/worker/gpu_model_runner.py

@@ -1600,6 +1600,11 @@ class GPUModelRunnerBase(LoRAModelRunnerMixin):
             seq_len = (req_state.num_computed_tokens +
                        scheduler_output.num_scheduled_tokens[req_id])
             if seq_len < req_state.num_tokens:
+                # If we have already started decoding, seeing a "partial prefill"
+                # condition is suspicious and can lead to discarding the sampled
+                # token forever (PP stall).
+                if req_state.output_token_ids:
+                    continue
                 # Ignore the sampled token for partial prefills.
                 # Rewind the generator state as if the token was not sampled.
                 # This relies on cuda-specific torch-internal impl details
@@ -3461,6 +3466,11 @@ class GPUModelRunnerMTP(GPUModelRunnerBase):
             seq_len = (req_state.num_computed_tokens +
                        scheduler_output.num_scheduled_tokens[req_id])
             if seq_len < req_state.num_tokens:
+                # If we have already started decoding, seeing a "partial prefill"
+                # condition is suspicious and can lead to discarding the sampled
+                # token forever (PP stall).
+                if req_state.output_token_ids:
+                    continue
                 # Ignore the sampled token for partial prefills.
                 # Rewind the generator state as if the token was not sampled.
                 # This relies on cuda-specific torch-internal impl details