vllm-cpu 0.11.0.post2__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/v1/worker/xpu_worker.py

@@ -0,0 +1,179 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import os
+
+import torch
+import torch.distributed
+
+import vllm.envs as envs
+from vllm.config import VllmConfig
+from vllm.distributed import get_world_group
+from vllm.logger import init_logger
+from vllm.model_executor import set_random_seed
+from vllm.platforms import current_platform
+from vllm.v1.worker.gpu_worker import (Worker,
+                                       init_worker_distributed_environment)
+from vllm.v1.worker.xpu_model_runner import XPUModelRunner
+
+logger = init_logger(__name__)
+
+
+class XPUWorker(Worker):
+    """A XPU worker class."""
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        local_rank: int,
+        rank: int,
+        distributed_init_method: str,
+        is_driver_worker: bool = False,
+    ):
+        super().__init__(vllm_config, local_rank, rank,
+                         distributed_init_method, is_driver_worker)
+        device_config = self.device_config
+        assert device_config.device_type == "xpu"
+        assert current_platform.is_xpu()
+
+        # Torch profiler. Enabled and configured through env vars:
+        # VLLM_TORCH_PROFILER_DIR=/path/to/save/trace
+        if envs.VLLM_TORCH_PROFILER_DIR:
+            torch_profiler_trace_dir = envs.VLLM_TORCH_PROFILER_DIR
+            logger.info("Profiling enabled. Traces will be saved to: %s",
+                        torch_profiler_trace_dir)
+            logger.debug(
+                "Profiler config: record_shapes=%s,"
+                "profile_memory=%s,with_stack=%s,with_flops=%s",
+                envs.VLLM_TORCH_PROFILER_RECORD_SHAPES,
+                envs.VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
+                envs.VLLM_TORCH_PROFILER_WITH_STACK,
+                envs.VLLM_TORCH_PROFILER_WITH_FLOPS,
+            )
+            self.profiler = torch.profiler.profile(
+                activities=[
+                    torch.profiler.ProfilerActivity.CPU,
+                    torch.profiler.ProfilerActivity.XPU,
+                ],
+                record_shapes=envs.VLLM_TORCH_PROFILER_RECORD_SHAPES,
+                profile_memory=envs.VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY,
+                with_stack=envs.VLLM_TORCH_PROFILER_WITH_STACK,
+                with_flops=envs.VLLM_TORCH_PROFILER_WITH_FLOPS,
+                on_trace_ready=torch.profiler.tensorboard_trace_handler(
+                    torch_profiler_trace_dir, use_gzip=True))
+        else:
+            self.profiler = None
+
+    # we provide this function due to `torch.xpu.mem_get_info()` doesn't
+    # return correct free_gpu_memory on intel client GPU. We need to
+    # calculate/estiamte it.
+    def xpu_get_mem_info(self):
+        if current_platform.is_data_center_gpu():
+            return torch.xpu.mem_get_info()
+        else:
+            _, total_gpu_memory = torch.xpu.mem_get_info()
+            # FIXME: memory_allocated() doesn't count non-torch allocations,
+            # and we don't have any API to get it. so we mark it as 128MB.
+            used_memory = torch.xpu.memory_allocated()
+            non_torch_allocations = 128 * 1024 * 1024
+            free_gpu_memory = total_gpu_memory - (used_memory +
+                                                  non_torch_allocations)
+            return free_gpu_memory, total_gpu_memory
+
+    @torch.inference_mode()
+    def determine_available_memory(self) -> int:
+        """Profiles the peak memory usage of the model to determine how many
+        KV blocks may be allocated without OOMs.
+        The engine will first conduct a profiling of the existing memory usage.
+        Then, it calculates the maximum possible number of GPU and CPU blocks
+        that can be allocated with the remaining free memory.
+        .. tip::
+            You may limit the usage of GPU memory
+            by adjusting the `gpu_memory_utilization` parameter.
+        """
+        # Profile the memory usage of the model and get the maximum number of
+        # cache blocks that can be allocated with the remaining free memory.
+        torch.xpu.empty_cache()
+        torch.xpu.reset_peak_memory_stats()
+
+        free_gpu_memory, total_gpu_memory = torch.xpu.mem_get_info()
+        current_allocated_bytes = torch.xpu.memory_allocated()
+        msg = ("Before memory profiling run, "
+               f"total GPU memory: {total_gpu_memory / 1024**2:.2f} MB, "
+               f"model load takes {current_allocated_bytes / 1024**2:.2f} MB, "
+               f"free gpu memory is {free_gpu_memory / 1024**2:.2f} MB.")
+        logger.info(msg)
+        # Execute a forward pass with dummy inputs to profile the memory usage
+        # of the model.
+        self.model_runner.profile_run()
+
+        free_gpu_memory, _ = self.xpu_get_mem_info()
+        # NOTE(woosuk): Here we assume that the other processes using the same
+        # GPU did not change their memory usage during the profiling.
+        assert self.init_gpu_memory > free_gpu_memory, (
+            "Error in memory profiling. "
+            f"Initial free memory {self.init_gpu_memory}, current free memory"
+            f" {free_gpu_memory}. This happens when the GPU memory was "
+            "not properly cleaned up before initializing the vLLM instance.")
+
+        # Get the peak memory allocation recorded by torch
+        peak_memory = torch.xpu.memory_stats()["allocated_bytes.all.peak"]
+
+        torch.xpu.empty_cache()
+        torch_allocated_bytes = torch.xpu.memory_stats(
+        )["allocated_bytes.all.current"]
+        total_allocated_bytes = self.xpu_get_mem_info(
+        )[1] - self.xpu_get_mem_info()[0]
+
+        non_torch_allocations = total_allocated_bytes - torch_allocated_bytes
+        if non_torch_allocations > 0:
+            peak_memory += non_torch_allocations
+        available_kv_cache_memory = (
+            total_gpu_memory * self.cache_config.gpu_memory_utilization -
+            peak_memory)
+
+        msg = ("After memory profiling run, "
+               f"peak memory usage is {peak_memory / 1024**2:.2f} MB,"
+               f"torch mem is {torch_allocated_bytes / 1024**2:.2f} MB, "
+               f"non-torch mem is {non_torch_allocations / 1024**2:.2f} MB, "
+               f"free gpu memory is {free_gpu_memory / 1024**2:.2f} MB.")
+        logger.info(msg)
+
+        return int(available_kv_cache_memory)
+
+    def init_device(self):
+        if self.device_config.device.type == "xpu" and current_platform.is_xpu(
+        ):
+            self.device = torch.device(f"xpu:{self.local_rank}")
+            current_platform.set_device(self.device)
+            current_platform.check_if_supports_dtype(self.model_config.dtype)
+            torch.xpu.empty_cache()
+            self.init_gpu_memory = torch.xpu.get_device_properties(
+                self.local_rank).total_memory
+        else:
+            raise RuntimeError(
+                f"Not support device type: {self.device_config.device}")
+
+        ENV_CCL_ZE_IPC_EXCHANGE = os.getenv("CCL_ZE_IPC_EXCHANGE", "pidfd")
+        ENV_CCL_ATL_TRANSPORT = os.getenv("CCL_ATL_TRANSPORT", "ofi")
+        ENV_LOCAL_WORLD_SIZE = os.getenv("LOCAL_WORLD_SIZE",
+                                         str(self.parallel_config.world_size))
+        os.environ["CCL_ZE_IPC_EXCHANGE"] = ENV_CCL_ZE_IPC_EXCHANGE
+        os.environ["CCL_ATL_TRANSPORT"] = ENV_CCL_ATL_TRANSPORT
+        os.environ["LOCAL_WORLD_SIZE"] = ENV_LOCAL_WORLD_SIZE
+        os.environ["LOCAL_RANK"] = str(self.local_rank)
+
+        init_worker_distributed_environment(self.vllm_config, self.rank,
+                                            self.distributed_init_method,
+                                            self.local_rank,
+                                            current_platform.dist_backend)
+
+        # global all_reduce needed for overall oneccl warm up
+        torch.distributed.all_reduce(torch.zeros(1).xpu(),
+                                     group=get_world_group().device_group)
+
+        # Set random seed.
+        set_random_seed(self.model_config.seed)
+
+        # Construct the model runner
+        self.model_runner = XPUModelRunner(  # type: ignore
+            self.vllm_config, self.device)

vllm/version.py

@@ -0,0 +1,41 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+try:
+    from ._version import __version__, __version_tuple__
+except Exception as e:
+    import warnings
+
+    warnings.warn(f"Failed to read commit hash:\n{e}",
+                  RuntimeWarning,
+                  stacklevel=2)
+
+    __version__ = "dev"
+    __version_tuple__ = (0, 0, __version__)
+
+
+def _prev_minor_version_was(version_str):
+    """Check whether a given version matches the previous minor version.
+
+    Return True if version_str matches the previous minor version.
+
+    For example - return True if the current version if 0.7.4 and the
+    supplied version_str is '0.6'.
+
+    Used for --show-hidden-metrics-for-version.
+    """
+    # Match anything if this is a dev tree
+    if __version_tuple__[0:2] == (0, 0):
+        return True
+
+    # Note - this won't do the right thing when we release 1.0!
+    assert __version_tuple__[0] == 0
+    assert isinstance(__version_tuple__[1], int)
+    return version_str == f"{__version_tuple__[0]}.{__version_tuple__[1] - 1}"
+
+
+def _prev_minor_version():
+    """For the purpose of testing, return a previous minor version number."""
+    # In dev tree, this will return "0.-1", but that will work fine"
+    assert isinstance(__version_tuple__[1], int)
+    return f"{__version_tuple__[0]}.{__version_tuple__[1] - 1}"

vllm/worker/worker_base.py

@@ -0,0 +1,279 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+from typing import (Any, Callable, Dict, List, Optional, Set, Tuple, TypeVar,
+                    Union)
+
+import cloudpickle
+import torch.nn as nn
+
+from vllm.config import VllmConfig, set_current_vllm_config
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.sequence import ExecuteModelRequest
+from vllm.utils import (enable_trace_function_call_for_thread,
+                        resolve_obj_by_qualname, run_method,
+                        update_environment_variables,
+                        warn_for_unimplemented_methods)
+from vllm.v1.outputs import SamplerOutput
+
+logger = init_logger(__name__)
+
+_R = TypeVar("_R")
+
+
+@warn_for_unimplemented_methods
+class WorkerBase:
+    """Worker interface that allows vLLM to cleanly separate implementations for
+    different hardware. Also abstracts control plane communication, e.g., to
+    communicate request metadata to other workers.
+    """
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+    ) -> None:
+        self.vllm_config = vllm_config
+        self.model_config = vllm_config.model_config
+        self.cache_config = vllm_config.cache_config
+        self.lora_config = vllm_config.lora_config
+        self.load_config = vllm_config.load_config
+        self.parallel_config = vllm_config.parallel_config
+        self.scheduler_config = vllm_config.scheduler_config
+        self.device_config = vllm_config.device_config
+        self.speculative_config = vllm_config.speculative_config
+        self.observability_config = vllm_config.observability_config
+        self.kv_transfer_config = vllm_config.kv_transfer_config
+        self.compilation_config = vllm_config.compilation_config
+        from vllm.platforms import current_platform
+        self.current_platform = current_platform
+
+    def init_device(self) -> None:
+        """Initialize device state, such as loading the model or other on-device
+        memory allocations.
+        """
+        raise NotImplementedError
+
+    def initialize_cache(self, num_gpu_blocks: int,
+                         num_cpu_blocks: int) -> None:
+        """Initialize the KV cache with the given size in blocks.
+        """
+        raise NotImplementedError
+
+    def get_model(self) -> nn.Module:
+        raise NotImplementedError
+
+    def apply_model(self, fn: Callable[[nn.Module], _R]) -> _R:
+        """Apply a function on the model inside this worker."""
+        return fn(self.get_model())
+
+    def load_model(self) -> None:
+        """Load model onto target device."""
+        raise NotImplementedError
+
+    def execute_model(
+        self,
+        execute_model_req: Optional[ExecuteModelRequest] = None
+    ) -> Optional[List[SamplerOutput]]:
+        raise NotImplementedError
+
+    def start_worker_execution_loop(self) -> None:
+        """Execute model loop in parallel worker.
+
+        You can stop the loop by executing a driver worker with an empty output.
+        See `stop_remote_worker_execution_loop` for more details.
+        """
+        with self.current_platform.inference_mode():
+            while True:
+                output = self.execute_model(execute_model_req=None)
+                if output is None:
+                    return None
+
+    def determine_num_available_blocks(self) -> Tuple[int, int]:
+        """Determine the number of available blocks for the GPU KV cache and
+        swappable CPU KV cache.
+
+        The implementation may run profiling or other heuristics to determine
+        the size of caches.
+
+        Returns a Tuple[num_gpu_blocks, num_cpu_blocks], where num_gpu_blocks
+        are blocks that are "active" on the device and can be appended to.
+        num_cpu_blocks refers to "swapped" blocks in CPU memory and cannot be
+        appended to.
+        """
+        raise NotImplementedError
+
+    def get_cache_block_size_bytes(self) -> int:
+        """Return the size of a single cache block, in bytes. Used in
+        speculative decoding.
+        """
+        raise NotImplementedError
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        raise NotImplementedError
+
+    def remove_lora(self, lora_id: int) -> bool:
+        raise NotImplementedError
+
+    def pin_lora(self, lora_id: int) -> bool:
+        raise NotImplementedError
+
+    def list_loras(self) -> Set[int]:
+        raise NotImplementedError
+
+    @property
+    def vocab_size(self) -> int:
+        """Get vocabulary size from model configuration."""
+        return self.model_config.get_vocab_size()
+
+    def shutdown(self) -> None:
+        """Clean up resources held by the worker."""
+        return
+
+
+class WorkerWrapperBase:
+    """
+    This class represents one process in an executor/engine. It is responsible
+    for lazily initializing the worker and handling the worker's lifecycle.
+    We first instantiate the WorkerWrapper, which remembers the worker module
+    and class name. Then, when we call `update_environment_variables`, and the
+    real initialization happens in `init_worker`.
+    """
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        rpc_rank: int = 0,
+    ) -> None:
+        """
+        Initialize the worker wrapper with the given vllm_config and rpc_rank.
+        Note: rpc_rank is the rank of the worker in the executor. In most cases,
+        it is also the rank of the worker in the distributed group. However,
+        when multiple executors work together, they can be different.
+        e.g. in the case of SPMD-style offline inference with TP=2,
+        users can launch 2 engines/executors, each with only 1 worker.
+        All workers have rpc_rank=0, but they have different ranks in the TP
+        group.
+        """
+        self.rpc_rank = rpc_rank
+        self.worker: Optional[WorkerBase] = None
+        self.vllm_config: Optional[VllmConfig] = None
+        # do not store this `vllm_config`, `init_worker` will set the final
+        # one. TODO: investigate if we can remove this field in
+        # `WorkerWrapperBase`, `init_cached_hf_modules` should be
+        # unnecessary now.
+        if vllm_config.model_config is not None:
+            # it can be None in tests
+            trust_remote_code = vllm_config.model_config.trust_remote_code
+            if trust_remote_code:
+                # note: lazy import to avoid importing torch before initializing
+                from vllm.utils import init_cached_hf_modules
+                init_cached_hf_modules()
+
+    def shutdown(self) -> None:
+        if self.worker is not None:
+            self.worker.shutdown()
+
+    def adjust_rank(self, rank_mapping: Dict[int, int]) -> None:
+        """
+        Adjust the rpc_rank based on the given mapping.
+        It is only used during the initialization of the executor,
+        to adjust the rpc_rank of workers after we create all workers.
+        """
+        if self.rpc_rank in rank_mapping:
+            self.rpc_rank = rank_mapping[self.rpc_rank]
+
+    def update_environment_variables(self, envs_list: List[Dict[str,
+                                                                str]]) -> None:
+        envs = envs_list[self.rpc_rank]
+        key = 'CUDA_VISIBLE_DEVICES'
+        if key in envs and key in os.environ:
+            # overwriting CUDA_VISIBLE_DEVICES is desired behavior
+            # suppress the warning in `update_environment_variables`
+            del os.environ[key]
+        update_environment_variables(envs)
+
+    def init_worker(self, all_kwargs: List[Dict[str, Any]]) -> None:
+        """
+        Here we inject some common logic before initializing the worker.
+        Arguments are passed to the worker class constructor.
+        """
+        kwargs = all_kwargs[self.rpc_rank]
+        self.vllm_config = kwargs.get("vllm_config")
+        assert self.vllm_config is not None, (
+            "vllm_config is required to initialize the worker")
+        enable_trace_function_call_for_thread(self.vllm_config)
+
+        from vllm.plugins import load_general_plugins
+        load_general_plugins()
+
+        if isinstance(self.vllm_config.parallel_config.worker_cls, str):
+            worker_class = resolve_obj_by_qualname(
+                self.vllm_config.parallel_config.worker_cls)
+        else:
+            logger.warning(
+                "passing worker_cls as a class object is strongly deprecated,"
+                " as the serialization of class objects can be tricky and"
+                " error-prone. To be safe, please keep the class in a separate"
+                " module and pass the qualified name of the class as a string."
+            )
+            assert isinstance(self.vllm_config.parallel_config.worker_cls,
+                              bytes)
+            worker_class = cloudpickle.loads(
+                self.vllm_config.parallel_config.worker_cls)
+        if self.vllm_config.parallel_config.worker_extension_cls:
+            worker_extension_cls = resolve_obj_by_qualname(
+                self.vllm_config.parallel_config.worker_extension_cls)
+            extended_calls = []
+            if worker_extension_cls not in worker_class.__bases__:
+                # check any conflicts between worker and worker_extension_cls
+                for attr in dir(worker_extension_cls):
+                    if attr.startswith("__"):
+                        continue
+                    assert not hasattr(worker_class, attr), (
+                        f"Worker class {worker_class} already has an attribute"
+                        f" {attr}, which conflicts with the worker"
+                        f" extension class {worker_extension_cls}.")
+                    if callable(getattr(worker_extension_cls, attr)):
+                        extended_calls.append(attr)
+                # dynamically inherit the worker extension class
+                worker_class.__bases__ = worker_class.__bases__ + (
+                    worker_extension_cls, )
+                logger.info(
+                    "Injected %s into %s for extended collective_rpc calls %s",
+                    worker_extension_cls, worker_class, extended_calls)
+        with set_current_vllm_config(self.vllm_config):
+            # To make vLLM config available during worker initialization
+            self.worker = worker_class(**kwargs)
+            assert self.worker is not None
+
+    def initialize_from_config(self, kv_cache_configs: List[Any]) -> None:
+        kv_cache_config = kv_cache_configs[self.rpc_rank]
+        with set_current_vllm_config(self.vllm_config):
+            self.worker.initialize_from_config(kv_cache_config)  # type: ignore
+
+    def init_device(self):
+        with set_current_vllm_config(self.vllm_config):
+            # To make vLLM config available during device initialization
+            self.worker.init_device()  # type: ignore
+
+    def execute_method(self, method: Union[str, bytes], *args, **kwargs):
+        try:
+            # method resolution order:
+            # if a method is defined in this class, it will be called directly.
+            # otherwise, since we define `__getattr__` and redirect attribute
+            # query to `self.worker`, the method will be called on the worker.
+            return run_method(self, method, args, kwargs)
+        except Exception as e:
+            # if the driver worker also execute methods,
+            # exceptions in the rest worker may cause deadlock in rpc like ray
+            # see https://github.com/vllm-project/vllm/issues/3455
+            # print the error and inform the user to solve the error
+            msg = (f"Error executing method {method!r}. "
+                   "This might cause deadlock in distributed execution.")
+            logger.exception(msg)
+            raise e
+
+    def __getattr__(self, attr):
+        return getattr(self.worker, attr)