diff --git a/vllm/v1/worker/xpu_model_runner.py b/vllm/v1/worker/xpu_model_runner.py index f6747cb349..15fa8e1e76 100644 --- a/vllm/v1/worker/xpu_model_runner.py +++ b/vllm/v1/worker/xpu_model_runner.py @@ -11,6 +11,7 @@ from vllm.v1.worker.gpu.model_runner import ( GPUModelRunner as GPUModelRunnerV2, ) from vllm.v1.worker.gpu_model_runner import GPUModelRunner +from vllm.v1.worker.mm_encoder_model_runner import MMEncoderModelRunner class XPUModelRunner(GPUModelRunner): @@ -39,6 +40,18 @@ class XPUModelRunnerV2(GPUModelRunnerV2): super().__init__(vllm_config, device) +class XPUMMEncoderModelRunner(MMEncoderModelRunner): + """An encoder-only model runner for XPU devices.""" + + def __init__( + self, + vllm_config: VllmConfig, + device: torch.device, + ): + with _torch_cuda_wrapper(): + super().__init__(vllm_config, device) + + @contextmanager def _torch_cuda_wrapper(): # Replace cuda APIs with xpu APIs. Each callable gets its own functools.partial diff --git a/vllm/v1/worker/xpu_worker.py b/vllm/v1/worker/xpu_worker.py index 02cc2c9e87..8a8166c435 100644 --- a/vllm/v1/worker/xpu_worker.py +++ b/vllm/v1/worker/xpu_worker.py @@ -13,7 +13,11 @@ from vllm.utils.torch_utils import set_random_seed from vllm.v1.utils import report_usage_stats from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment from vllm.v1.worker.workspace import init_workspace_manager -from vllm.v1.worker.xpu_model_runner import XPUModelRunner, XPUModelRunnerV2 +from vllm.v1.worker.xpu_model_runner import ( + XPUMMEncoderModelRunner, + XPUModelRunner, + XPUModelRunnerV2, +) from .utils import request_memory @@ -215,7 +219,14 @@ class XPUWorker(Worker): init_workspace_manager(self.device, num_ubatches) # Construct the model runner - model_runner = XPUModelRunnerV2 if self.use_v2_model_runner else XPUModelRunner + if self.use_v2_model_runner: + model_runner = ( + XPUMMEncoderModelRunner + if self.vllm_config.is_mm_encoder_only + else XPUModelRunnerV2 + ) + else: + model_runner = XPUModelRunner # type: ignore[assignment] self.model_runner = model_runner( # type: ignore self.vllm_config, self.device )