Merge branch 'v0.9.2-dev-fix-zero' into 'v0.9.2-dev'

fix: 解决原版0消耗chunk-prefill崩溃问题

See merge request dcutoolkit/deeplearing/vllm!393

vllm/zero_overhead/v1/gpu_model_runner.py

@@ -796,6 +796,7 @@ class V1ZeroModelRunner(GPUModelRunner):
                         req_state = self.requests[req_id]
                         token_idx = self.last_sampled_token_lens[req_idx]
                         if token_idx == -1:
+                            self.fix_sampled_token_ids[req_idx].clear()
                             continue
                         fix_len = len(self.fix_sampled_token_ids[req_idx])
                         req_state.output_token_ids[token_idx:token_idx + fix_len] = self.fix_sampled_token_ids[req_idx]