[platform] add ray_device_key (#11948)

Signed-off-by: youkaichao <youkaichao@gmail.com>

vllm/executor/ray_utils.py

@@ -8,6 +8,7 @@ import msgspec
 from vllm.config import ParallelConfig
 from vllm.executor.msgspec_utils import decode_hook, encode_hook
 from vllm.logger import init_logger
+from vllm.platforms import current_platform
 from vllm.sequence import ExecuteModelRequest, IntermediateTensors
 from vllm.utils import get_ip
 from vllm.worker.worker_base import WorkerWrapperBase
@@ -47,7 +48,12 @@ try:
 
         def get_node_and_gpu_ids(self) -> Tuple[str, List[int]]:
             node_id = ray.get_runtime_context().get_node_id()
-            gpu_ids = ray.get_gpu_ids()
+            device_key = current_platform.ray_device_key
+            if not device_key:
+                raise RuntimeError("current platform %s does not support ray.",
+                                   current_platform.device_name)
+            gpu_ids = ray.get_runtime_context().get_accelerator_ids(
+            )[device_key]
             return node_id, gpu_ids
 
         def execute_model_spmd(
@@ -249,11 +255,12 @@ def initialize_ray_cluster(
         # Placement group is already set.
         return
 
-    device_str = "GPU"
-    if current_platform.is_tpu():
-        device_str = "TPU"
-    elif current_platform.is_hpu():
-        device_str = 'HPU'
+    device_str = current_platform.ray_device_key
+    if not device_str:
+        raise ValueError(
+            f"current platform {current_platform.device_name} does not "
+            "support ray.")
+
     # Create placement group for worker processes
     current_placement_group = ray.util.get_current_placement_group()
     if current_placement_group:

vllm/platforms/cuda.py

@@ -77,6 +77,7 @@ class CudaPlatformBase(Platform):
     device_name: str = "cuda"
     device_type: str = "cuda"
     dispatch_key: str = "CUDA"
+    ray_device_key: str = "GPU"
 
     @classmethod
     def get_device_capability(cls,

vllm/platforms/hpu.py

@@ -19,6 +19,7 @@ class HpuPlatform(Platform):
     device_name: str = "hpu"
     device_type: str = "hpu"
     dispatch_key: str = "HPU"
+    ray_device_key: str = "HPU"
 
     @classmethod
     def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,

vllm/platforms/interface.py

@@ -82,6 +82,10 @@ class Platform:
     # check https://github.com/pytorch/pytorch/blob/313dac6c1ca0fa0cde32477509cce32089f8532a/torchgen/model.py#L134 # noqa
     # use "CPU" as a fallback for platforms not registered in PyTorch
     dispatch_key: str = "CPU"
+    # available ray device keys:
+    # https://github.com/ray-project/ray/blob/10ba5adadcc49c60af2c358a33bb943fb491a171/python/ray/_private/ray_constants.py#L438 # noqa
+    # empty string means the device does not support ray
+    ray_device_key: str = ""
     # The torch.compile backend for compiling simple and
     # standalone functions. The default value is "inductor" to keep
     # the same behavior as PyTorch.

vllm/platforms/neuron.py

@@ -16,6 +16,7 @@ class NeuronPlatform(Platform):
     _enum = PlatformEnum.NEURON
     device_name: str = "neuron"
     device_type: str = "neuron"
+    ray_device_key: str = "neuron_cores"
     supported_quantization: list[str] = ["neuron_quant"]
 
     @classmethod

vllm/platforms/rocm.py

@@ -64,6 +64,8 @@ class RocmPlatform(Platform):
     device_name: str = "rocm"
     device_type: str = "cuda"
     dispatch_key: str = "CUDA"
+    ray_device_key: str = "GPU"
+
     supported_quantization: list[str] = [
         "awq", "gptq", "fp8", "compressed_tensors", "compressed-tensors",
         "fbgemm_fp8", "gguf"

vllm/platforms/tpu.py

@@ -19,6 +19,8 @@ class TpuPlatform(Platform):
     device_name: str = "tpu"
     device_type: str = "tpu"
     dispatch_key: str = "XLA"
+    ray_device_key: str = "TPU"
+
     supported_quantization: list[str] = [
         "tpu_int8", "compressed-tensors", "compressed_tensors"
     ]

vllm/platforms/xpu.py

@@ -19,6 +19,9 @@ class XPUPlatform(Platform):
     device_name: str = "xpu"
     device_type: str = "xpu"
     dispatch_key: str = "XPU"
+    # Intel XPU's device key is "GPU" for Ray.
+    # see https://github.com/ray-project/ray/blob/6a5eb5865eeb9ccf058a79b44f107e327e360673/python/ray/_private/accelerators/intel_gpu.py#L20 # noqa: E501
+    ray_device_key: str = "GPU"
 
     @classmethod
     def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,

vllm/v1/executor/ray_utils.py

@@ -41,7 +41,12 @@ try:
 
         def get_node_and_gpu_ids(self) -> Tuple[str, List[int]]:
             node_id = ray.get_runtime_context().get_node_id()
-            gpu_ids = ray.get_gpu_ids()
+            device_key = current_platform.ray_device_key
+            if not device_key:
+                raise RuntimeError("current platform %s does not support ray.",
+                                   current_platform.device_name)
+            gpu_ids = ray.get_runtime_context().get_accelerator_ids(
+            )[device_key]
             return node_id, gpu_ids
 
         def setup_device_if_necessary(self):
@@ -211,7 +216,11 @@ def initialize_ray_cluster(
         # Placement group is already set.
         return
 
-    device_str = "GPU" if not current_platform.is_tpu() else "TPU"
+    device_str = current_platform.ray_device_key
+    if not device_str:
+        raise ValueError(
+            f"current platform {current_platform.device_name} does not "
+            "support ray.")
     # Create placement group for worker processes
     current_placement_group = ray.util.get_current_placement_group()
     if current_placement_group: