[ fix ] fix code and config about vllm because of new torch version

autotest/config-npu.yaml

@@ -217,7 +217,8 @@ case:
                 accelerator: NPU
                 output_path: /mnt/hwfile/llmrazor/qa-llm-cic/qa-llm-cicd/test_output
             resource:
-                image: ccr-hw/910c_cann910_pta291_rl_s1_vllm:v0.1.0-20260826T114552Z
+                image: ccr-hw/910c_cann910_pta291_rl_s1_vllm:v0.5.0-20260917T114217Z
+                pip_package: unset ASCEND_WORK_PATH;pip install -e . --no-deps; pip install more-itertools triton
                 envs:
                     - MODEL_PATH=/mnt/hwfile/llmrazor/qa-llm-cicd/qa_test_models/Qwen3-30B-A3B
                     - DATA_PATH=/mnt/hwfile/llmrazor/qa-llm-cicd/xtuner_resource/datasets/gsm8k/train-mini.jsonl
@@ -230,6 +231,9 @@ case:
                     - XTUNER_ACTIVATION_OFFLOAD=1
                     - VLLM_VERSION=0.11.0
                     - VLLM_USE_V1=1
+                    - TORCH_NPU_DEVICE_CAPABILITY=9.0
+                    - PIP_INDEX_URL=http://pkg.pjlab.org.cn/repository/pypi-tsinghua/simple
+                    - PIP_TRUSTED_HOST=pkg.pjlab.org.cn
             assert_info:
                 base_metric: npu-qwen3-rl-vllm/tracker.jsonl
                 check_metrics:

xtuner/v1/rl/rollout/vllm.py

@@ -28,9 +28,8 @@ def stateless_init_process_group(master_address, master_port, rank, world_size,
     """VLLM provides `StatelessProcessGroup` to create a process group without
     considering the global process group in torch.distributed.
 
-    It is recommended to create `StatelessProcessGroup`, and then initialize
-    the data-plane communication (NCCL) between external (train processes)
-    and vLLM workers.
+    It is recommended to create `StatelessProcessGroup`, and then initialize the data-plane communication (NCCL)
+    between external (train processes) and vLLM workers.
     """
     from vllm.distributed.utils import StatelessProcessGroup
 
@@ -210,6 +209,55 @@ def __init__(
         )
         self.tp_size = self.config.tensor_parallel_size // self.dp_size
 
+    def _get_request_payload(self, rollout_state: RolloutState) -> dict:
+        """Build the chat-completions request body for one generation
+        request."""
+        sample_params = rollout_state.sample_params
+        payload: dict[str, Any] = {
+            "model": self.config.model_path,
+            "messages": rollout_state.message,
+            "stream": sample_params.stream,
+        }
+
+        if rollout_state.tools is not None:
+            payload["tools"] = rollout_state.tools
+        if rollout_state.tool_choice is not None:
+            payload["tool_choice"] = rollout_state.tool_choice
+
+        # vLLM-Ascend (USE_TOKEN_IN=1) accepts explicit input_ids; partial-rollout
+        # replay passes the request prefix via rollout_state.tokens.
+        if rollout_state.tokens is not None:
+            payload["input_ids"] = rollout_state.tokens
+        elif "train_prompt_ids" in rollout_state.extra_fields:
+            payload["input_ids"] = rollout_state.extra_fields["train_prompt_ids"]
+
+        if "image_data" in rollout_state.extra_fields:
+            image_data = rollout_state.extra_fields["image_data"]
+            assert isinstance(payload["messages"], list), "image_data requires messages to be a list"
+            image_index = 0
+            for message in payload["messages"]:
+                if not isinstance(message, dict) or message.get("role") != "user":
+                    continue
+                new_content = []
+                for content_part in message.get("content", []):
+                    if not isinstance(content_part, dict):
+                        new_content.append(content_part)
+                        continue
+                    if content_part.get("type") == "image_url":
+                        content_part["image_url"]["url"] = f"file://{image_data[image_index]}"
+                        content_part["image_url"].pop("image_wh", None)
+                        image_index += 1
+                    new_content.append(content_part)
+                message["content"] = new_content
+            assert image_index == len(image_data), f"Expected {len(image_data)} images, but processed {image_index}."
+
+        vllm_sample_params = self._transform_sample_params(sample_params.model_dump())
+        vllm_sample_params["return_routed_experts"] = (
+            self.enable_return_routed_experts and sample_params.return_routed_experts
+        )
+        payload.update(vllm_sample_params)
+        return payload
+
     async def _create_request(
         self,
         url: str,
@@ -285,9 +333,6 @@ def _transform_sample_params(self, sample_params: Dict, extra_params: Dict = {})
     def get_logprobs(self, input_ids, sampling_params):
         pass
 
-    def generate(self, input_ids, sampling_params):
-        pass
-
     def sleep(self, level=1):
         url = f"{self.server_url}/{self.endpoints['sleep']}"
         headers = {"Content-Type": "application/json"}
@@ -368,7 +413,7 @@ def _transform_rollout_config_to_server_configs(self) -> Namespace:
             "cudagraph_capture_sizes": [16, 12, 8, 4, 2, 1],
             "cudagraph_mode": "FULL_DECODE_ONLY",
         }
-        args["additional_config"] = {"enable_cpu_binding": True}
+        args["additional_config"] = {"enable_cpu_binding": True, "weight_nz_mode": 0}
         args["limit_mm_per_prompt"] = {"image": 10, "video": 0}
         args["enable_log_requests"] = False
         args["uvicorn_log_level"] = "error"