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

[fix] pp+mtp bs 1 correctness

See merge request dcutoolkit/deeplearing/vllm!214

vllm/v1/worker/gpu_model_runner.py

@@ -528,10 +528,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
             if not is_last_rank:
                 # Add new_token_ids to token_ids_cpu.
                 start_token_index = num_computed_tokens
-                end_token_index = num_computed_tokens + len(new_token_ids)
+                end_token_index = num_computed_tokens + 1
                 self.input_batch.token_ids_cpu[
                     req_index,
-                    start_token_index:end_token_index] = new_token_ids
+                    start_token_index:end_token_index] = new_token_ids[-1]
                 self.input_batch.num_tokens_no_spec[
                     req_index] = end_token_index
                 self.input_batch.num_tokens[req_index] = end_token_index