[llm] ray.llm support custom accelerators (ray-project#51359)

Signed-off-by: liuxsh9 <liuxiaoshuang4@huawei.com>
Signed-off-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
Co-authored-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
Signed-off-by: Dhakshin Suriakannu <d_suriakannu@apple.com>

Author: 2 people

python/ray/data/llm.py

@@ -19,6 +19,9 @@ class ProcessorConfig(_ProcessorConfig):
             On the other hand, small batch sizes are more fault-tolerant and could
             reduce bubbles in the data pipeline. You can tune the batch size to balance
             the throughput and fault-tolerance based on your use case.
+        resources_per_bundle: The resource bundles for placement groups.
+            You can specify a custom device label e.g. {'NPU': 1}.
+            The default resource bundle for LLM Stage is always a GPU resource i.e. {'GPU': 1}.
         accelerator_type: The accelerator type used by the LLM stage in a processor.
             Default to None, meaning that only the CPU will be used.
         concurrency: The number of workers for data parallelism. Default to 1.

python/ray/llm/_internal/batch/processor/base.py

@@ -30,6 +30,12 @@ class ProcessorConfig(BaseModelExtended):
         "You can tune the batch size to balance the throughput and fault-tolerance "
         "based on your use case. Defaults to 64.",
     )
+    resources_per_bundle: Optional[Dict[str, float]] = Field(
+        default=None,
+        description="This will override the default resource bundles for placement groups. "
+        "You can specify a custom device label e.g. {'NPU': 1}. "
+        "The default resource bundle for LLM Stage is always a GPU resource i.e. {'GPU': 1}.",
+    )
     accelerator_type: Optional[str] = Field(
         default=None,
         description="The accelerator type used by the LLM stage in a processor. "

python/ray/llm/_internal/batch/processor/vllm_engine_proc.py

@@ -193,6 +193,7 @@ def build_vllm_engine_processor(
                 # This is used to make sure we overlap batches to avoid the tail
                 # latency of each batch.
                 max_concurrency=config.max_concurrent_batches,
+                resources=config.resources_per_bundle,
                 accelerator_type=config.accelerator_type,
                 runtime_env=config.runtime_env,
             ),

python/ray/llm/_internal/batch/stages/vllm_engine_stage.py

@@ -570,26 +570,42 @@ def __del__(self):
             self.llm.shutdown()
 
 
-def _ray_scheduling_strategy_fn(num_gpus_per_instance: int, accelerator_type: str):
-    """
-    Create a Ray scheduling strategy for vLLM engine.
+def _ray_scheduling_strategy_fn(
+    num_bundles_per_replica: int,
+    accelerator_type: Optional[str] = None,
+    resources_per_bundle: Optional[Dict[str, float]] = None,
+):
+    """Create a Ray scheduling strategy for the engine.
 
     Args:
-        num_gpus_per_instance: The number of GPUs per instance.
-        accelerator_type: The accelerator type.
+        num_bundles_per_replica: The number of device bundles per
+            engine replica.
+        accelerator_type: The accelerator type. If None, the
+            accelerator_type label will not be set.
+        resources_per_bundle: The custom resources per bundle.
+            If None, we default to 1xGPU + 1xCPU bundle.
 
     Returns:
         The Ray scheduling strategy.
     """
 
     def _get_bundle() -> Dict[str, float]:
-        bundle: Dict[str, float] = {"GPU": 1, "CPU": 1}
+
+        bundle = {}
+        # Custom resources
+        if resources_per_bundle:
+            bundle = resources_per_bundle
+        else:
+            # GPU bundles
+            bundle = {"GPU": 1, "CPU": 1}
+
+        # Accelerator type
         if accelerator_type:
             bundle[f"accelerator_type:{accelerator_type}"] = 0.001
         return bundle
 
     pg = ray.util.placement_group(
-        [_get_bundle()] * num_gpus_per_instance,
+        [_get_bundle()] * num_bundles_per_replica,
         strategy="STRICT_PACK",
     )
     return dict(
@@ -626,30 +642,40 @@ def post_init(cls, values):
         if accelerator_type:
             ray_remote_args["accelerator_type"] = accelerator_type
 
-        # Setup num_gpus required per vLLM engine.
+        # Setup num_workers required per vLLM engine.
         tp_size = engine_kwargs.get("tensor_parallel_size", 1)
         pp_size = engine_kwargs.get("pipeline_parallel_size", 1)
-        num_gpus = tp_size * pp_size
+        num_bundles_per_replica = tp_size * pp_size
 
         # Use the MP backend by default.
         engine_kwargs.setdefault("distributed_executor_backend", "mp")
         executor_backend = engine_kwargs.get("distributed_executor_backend")
 
-        # When Ray is used in the vLLM engine, we set num_gpus to 0 so that
+        # When Ray is used in the vLLM engine, we set num_devices to 0 so that
         # Ray Data won't reserve GPUs in advance. Instead, we specify scheduling
         # strategy in .map_batches() arguments and let vLLM Ray executor to
         # create placement groups for each TP/PP worker.
-        if executor_backend == "ray" and num_gpus > 1:
+        resources_per_bundle = map_batches_kwargs.pop("resources", None)
+        if executor_backend == "ray" and num_bundles_per_replica > 1:
             # Note that we have to use partial() to pass a function
             # instead of an object.
             map_batches_kwargs["ray_remote_args_fn"] = partial(
                 _ray_scheduling_strategy_fn,
-                num_gpus,
+                num_bundles_per_replica,
                 accelerator_type,
+                resources_per_bundle,
             )
-            num_gpus = 0
+            ray_remote_args["num_gpus"] = 0
+        else:
+            if not resources_per_bundle:
+                # Default to GPUs per bundle if custom resources are not specified.
+                ray_remote_args["num_gpus"] = num_bundles_per_replica
+            else:
+                ray_remote_args["resources"] = {
+                    resource_key: resource_count * num_bundles_per_replica
+                    for resource_key, resource_count in resources_per_bundle.items()
+                }
 
-        map_batches_kwargs["num_gpus"] = num_gpus
         map_batches_kwargs.update(ray_remote_args)
         return values
 

python/ray/llm/_internal/serve/configs/server_models.py

@@ -190,6 +190,13 @@ class LLMConfig(BaseModelExtended):
         ),
     )
 
+    resources_per_bundle: Optional[Dict[str, float]] = Field(
+        default=None,
+        description="This will override the default resource bundles for placement groups. "
+        "You can specify a custom device label e.g. {'NPU': 1}. "
+        "The default resource bundle for LLM Stage is always a GPU resource i.e. {'GPU': 1}.",
+    )
+
     accelerator_type: Optional[str] = Field(
         default=None,
         description=f"The type of accelerator runs the model on. Only the following values are supported: {str([t.value for t in GPUType])}",

python/ray/llm/_internal/serve/deployments/llm/vllm/vllm_models.py

@@ -51,6 +51,11 @@ class VLLMEngineConfig(BaseModelExtended):
         None,
         description="Configuration for cloud storage mirror. This is for where the weights are downloaded from.",
     )
+    resources_per_bundle: Optional[Dict[str, float]] = Field(
+        default=None,
+        description="This overrides the vLLM engine worker's default resource configuration, "
+        "the number of resources returned by `placement_bundles`.",
+    )
     accelerator_type: Optional[GPUType] = Field(
         None,
         description="The type of accelerator to use. This is used to determine the placement group strategy.",
@@ -104,6 +109,7 @@ def from_llm_config(cls, llm_config: LLMConfig) -> "VLLMEngineConfig":
             model_id=llm_config.model_id,
             hf_model_id=hf_model_id,
             mirror_config=mirror_config,
+            resources_per_bundle=llm_config.resources_per_bundle,
             accelerator_type=llm_config.accelerator_type,
             engine_kwargs=llm_config.engine_kwargs,
             runtime_env=llm_config.runtime_env,
@@ -122,7 +128,7 @@ def pipeline_parallel_degree(self) -> int:
         return self.engine_kwargs.get("pipeline_parallel_size", 1)
 
     @property
-    def num_gpu_workers(self) -> int:
+    def num_devices(self) -> int:
         return self.tensor_parallel_degree * self.pipeline_parallel_degree
 
     @property
@@ -134,10 +140,13 @@ def placement_strategy(self) -> str:
 
     @property
     def placement_bundles(self) -> List[Dict[str, float]]:
-        bundle = {"GPU": 1}
+        if self.resources_per_bundle:
+            bundle = self.resources_per_bundle
+        else:
+            bundle = {"GPU": 1}
         if self.accelerator_type:
             bundle[self.ray_accelerator_type()] = 0.001
-        bundles = [bundle for _ in range(self.num_gpu_workers)]
+        bundles = [bundle for _ in range(self.num_devices)]
 
         return bundles
 

python/ray/llm/_internal/serve/observability/usage_telemetry/usage.py

@@ -266,6 +266,6 @@ def push_telemetry_report_for_all_models(
             max_replicas=max_replicas,
             tensor_parallel_degree=engine_config.tensor_parallel_degree,
             gpu_type=model.accelerator_type or DEFAULT_GPU_TYPE,
-            num_gpus=engine_config.num_gpu_workers,
+            num_gpus=engine_config.num_devices,
         )
         _push_telemetry_report(telemetry_model)

python/ray/llm/tests/serve/configs/test_models.py

@@ -212,6 +212,28 @@ def test_get_serve_options_without_accelerator_type(self):
         }
         assert serve_options == expected_options
 
+    def test_resources_per_bundle(self):
+        """Test that resources_per_bundle is correctly parsed."""
+
+        # Test the default resource bundle
+        serve_options = LLMConfig(
+            model_loading_config=dict(model_id="test_model"),
+            engine_kwargs=dict(tensor_parallel_size=3, pipeline_parallel_size=2),
+        ).get_serve_options(name_prefix="Test:")
+        assert serve_options["placement_group_bundles"] == [{"CPU": 1, "GPU": 0}] + [
+            {"GPU": 1} for _ in range(6)
+        ]
+
+        # Test the custom resource bundle
+        serve_options = LLMConfig(
+            model_loading_config=dict(model_id="test_model"),
+            engine_kwargs=dict(tensor_parallel_size=3, pipeline_parallel_size=2),
+            resources_per_bundle={"XPU": 1},
+        ).get_serve_options(name_prefix="Test:")
+        assert serve_options["placement_group_bundles"] == [{"CPU": 1, "GPU": 0}] + [
+            {"XPU": 1} for _ in range(6)
+        ]
+
 
 if __name__ == "__main__":
     sys.exit(pytest.main(["-v", __file__]))

python/ray/util/accelerators/accelerators.py

@@ -26,6 +26,8 @@
 GOOGLE_TPU_V5P = "TPU-V5P"
 GOOGLE_TPU_V5LITEPOD = "TPU-V5LITEPOD"
 GOOGLE_TPU_V6E = "TPU-V6E"
+HUAWEI_NPU_910B = "Ascend910B"
+HUAWEI_NPU_910B4 = "Ascend910B4"
 
 # Use these instead of NVIDIA_A100 if you need a specific accelerator size. Note that
 # these labels are not auto-added to nodes, you'll have to add them manually in