[XPU][CI] enhance xpu test support (#20652)

Signed-off-by: Ma, Liangliang <liangliang.ma@intel.com>
Co-authored-by: zhenwei-intel <zhenweiliu@habana.ai>

tests/conftest.py

@@ -759,7 +759,8 @@ class VllmRunner:
     - `trust_remote_code`: Set to `True` instead of `False` for convenience.
     - `seed`: Set to `0` instead of `None` for test reproducibility.
     - `max_model_len`: Set to `1024` instead of `None` to reduce memory usage.
-    - `block_size`: Set to `16` instead of `None` to reduce memory usage.
+    - `block_size`: To reduce memory usage, set default to `64` if on XPU
+        devices, otherwise default to `16`.
     - `enable_chunked_prefill`: Set to `False` instead of `None` for
       test reproducibility.
     - `enforce_eager`: Set to `False` to test CUDA graph.
@@ -777,7 +778,7 @@ class VllmRunner:
         dtype: str = "auto",
         disable_log_stats: bool = True,
         tensor_parallel_size: int = 1,
-        block_size: int = 16,
+        block_size: int = 16 if not torch.xpu.is_available() else 64,
         enable_chunked_prefill: Optional[bool] = False,
         swap_space: int = 4,
         enforce_eager: Optional[bool] = False,

vllm/distributed/device_communicators/xpu_communicator.py

@@ -53,3 +53,6 @@ class XpuCommunicator(DeviceCommunicatorBase):
         else:
             output_tensor = None
         return output_tensor
+
+    def broadcast(self, input_: torch.Tensor, src: int = 0) -> None:
+        dist.broadcast(input_, src=src, group=self.device_group)

vllm/distributed/parallel_state.py

@@ -240,6 +240,8 @@ class GroupCoordinator:
 
         if current_platform.is_cuda_alike():
             self.device = torch.device(f"cuda:{local_rank}")
+        elif current_platform.is_xpu():
+            self.device = torch.device(f"xpu:{local_rank}")
         elif current_platform.is_out_of_tree():
             self.device = torch.device(
                 f"{current_platform.device_name}:{local_rank}")

vllm/platforms/xpu.py

@@ -91,6 +91,7 @@ class XPUPlatform(Platform):
 
         # FIXME: Temporarily forcing eager mode
         # remove after t.compile support stabilizes.
+
         if (envs.VLLM_USE_V1 and vllm_config.model_config is not None
                 and not vllm_config.model_config.enforce_eager):
             from vllm.config import CompilationLevel
@@ -111,9 +112,6 @@ class XPUPlatform(Platform):
                     "mode.")
                 model_config.enforce_eager = True
 
-        if vllm_config.device_config is not None:
-            assert vllm_config.device_config.device_type == "xpu"
-
         # check and update parallel config
         parallel_config = vllm_config.parallel_config
         parallel_config.worker_cls = "vllm.v1.worker.xpu_worker.XPUWorker"
@@ -131,8 +129,10 @@ class XPUPlatform(Platform):
                 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
                 logger.warning(
                     "Please use spawn as start method if you want to use mp.")
-        elif parallel_config.distributed_executor_backend != "ray" and \
+        elif (parallel_config.distributed_executor_backend != "ray"
-                parallel_config.distributed_executor_backend != "uni":
+              and parallel_config.distributed_executor_backend != "uni"
+              and parallel_config.distributed_executor_backend
+              != "external_launcher"):
             logger.warning(
                 "%s is not supported on XPU, fallback to ray distributed"
                 " executor backend.",

vllm/v1/worker/xpu_model_runner.py

@@ -27,7 +27,7 @@ class XPUModelRunner(GPUModelRunner):
         self.cascade_attn_enabled = False
 
     def _init_device_properties(self) -> None:
-        pass
+        self.num_sms = None
 
     def _sync_device(self) -> None:
         torch.xpu.synchronize()