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

vllm/v1/request.py

@@ -79,6 +79,7 @@ class Request:
         self._all_token_ids: list[int] = self.prompt_token_ids.copy()
         self.spec_token_ids: list[int] = []
         self.num_computed_tokens = 0
+        self.num_generated_token_ids = 0
         self.cache_salt: Optional[str] = cache_salt
 
         # Multi-modal related

vllm/v1/worker/gpu_model_runner.py

@@ -496,8 +496,8 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 elif num_new_tokens > 0:
                     req_state.output_token_ids.extend(
                         new_token_ids[-num_new_tokens:])
-                if len(spec_token_ids) > 0:
-                    req_state.spec_token_ids = spec_token_ids
+            if len(spec_token_ids) > 0:
+                req_state.spec_token_ids = spec_token_ids
 
             # Update the block IDs.
             if not resumed_from_preemption:
@@ -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