vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/v1/worker/gpu_worker.py

@@ -0,0 +1,393 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""A GPU worker class."""
+import gc
+import os
+from typing import TYPE_CHECKING, Optional
+
+import torch
+import torch.distributed
+import torch.nn as nn
+
+import vllm.envs as envs
+from vllm.config import VllmConfig
+from vllm.device_allocator.cumem import CuMemAllocator
+from vllm.distributed import (ensure_model_parallel_initialized,
+                              init_distributed_environment,
+                              set_custom_all_reduce)
+from vllm.distributed.kv_transfer import ensure_kv_transfer_initialized
+from vllm.distributed.parallel_state import get_pp_group, get_tp_group
+from vllm.logger import init_logger
+from vllm.lora.request import LoRARequest
+from vllm.model_executor import set_random_seed
+from vllm.platforms import current_platform
+from vllm.sequence import IntermediateTensors
+from vllm.utils import GiB_bytes, MemorySnapshot, memory_profiling
+from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
+from vllm.v1.outputs import ModelRunnerOutput
+from vllm.v1.utils import report_usage_stats
+from vllm.v1.worker.gpu_model_runner import GPUModelRunner
+from vllm.v1.worker.worker_base import WorkerBase
+
+logger = init_logger(__name__)
+
+if TYPE_CHECKING:
+    from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
+    from vllm.v1.core.sched.output import SchedulerOutput
+
+
+class Worker(WorkerBase):
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        local_rank: int,
+        rank: int,
+        distributed_init_method: str,
+        is_driver_worker: bool = False,
+    ):
+
+        super().__init__(vllm_config=vllm_config,
+                         local_rank=local_rank,
+                         rank=rank,
+                         distributed_init_method=distributed_init_method,
+                         is_driver_worker=is_driver_worker)
+
+        if self.model_config.trust_remote_code:
+            # note: lazy import to avoid importing torch before initializing
+            from vllm.utils import init_cached_hf_modules
+            init_cached_hf_modules()
+
+        # Buffers saved before sleep
+        self._sleep_saved_buffers: dict[str, torch.Tensor] = {}
+
+        # 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)
+            self.profiler = torch.profiler.profile(
+                activities=[
+                    torch.profiler.ProfilerActivity.CPU,
+                    torch.profiler.ProfilerActivity.CUDA,
+                ],
+                with_stack=True,
+                on_trace_ready=torch.profiler.tensorboard_trace_handler(
+                    torch_profiler_trace_dir, use_gzip=True))
+        else:
+            self.profiler = None
+
+    def sleep(self, level: int = 1) -> None:
+        free_bytes_before_sleep = torch.cuda.mem_get_info()[0]
+
+        # Save the buffers before level 2 sleep
+        if level == 2:
+            model = self.model_runner.model
+            self._sleep_saved_buffers = {
+                name: buffer.cpu().clone()
+                for name, buffer in model.named_buffers()
+            }
+
+        allocator = CuMemAllocator.get_instance()
+        allocator.sleep(offload_tags=("weights", ) if level == 1 else tuple())
+        free_bytes_after_sleep, total = torch.cuda.mem_get_info()
+        freed_bytes = free_bytes_after_sleep - free_bytes_before_sleep
+        used_bytes = total - free_bytes_after_sleep
+        assert freed_bytes >= 0, "Memory usage increased after sleeping."
+        logger.info(
+            "Sleep mode freed %.2f GiB memory, "
+            "%.2f GiB memory is still in use.", freed_bytes / GiB_bytes,
+            used_bytes / GiB_bytes)
+
+    def wake_up(self, tags: Optional[list[str]] = None) -> None:
+        allocator = CuMemAllocator.get_instance()
+        allocator.wake_up(tags)
+
+        # Restore the buffers after level 2 sleep
+        if len(self._sleep_saved_buffers):
+            model = self.model_runner.model
+            for name, buffer in model.named_buffers():
+                if name in self._sleep_saved_buffers:
+                    buffer.data.copy_(self._sleep_saved_buffers[name].data)
+            self._sleep_saved_buffers = {}
+
+    def init_device(self):
+        if self.device_config.device.type == "cuda":
+            # torch.distributed.all_reduce does not free the input tensor until
+            # the synchronization point. This causes the memory usage to grow
+            # as the number of all_reduce calls increases. This env var disables
+            # this behavior.
+            # Related issue:
+            # https://discuss.pytorch.org/t/cuda-allocation-lifetime-for-inputs-to-distributed-all-reduce/191573
+            os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
+
+            # This env var set by Ray causes exceptions with graph building.
+            os.environ.pop("NCCL_ASYNC_ERROR_HANDLING", None)
+            self.device = torch.device(f"cuda:{self.local_rank}")
+            torch.cuda.set_device(self.device)
+
+            _check_if_gpu_supports_dtype(self.model_config.dtype)
+            gc.collect()
+            torch.cuda.empty_cache()
+
+            # take current memory snapshot
+            self.init_snapshot = MemorySnapshot()
+            self.requested_memory = (self.init_snapshot.total_memory *
+                                     self.cache_config.gpu_memory_utilization)
+            if self.init_snapshot.free_memory < self.requested_memory:
+                GiB = lambda b: round(b / GiB_bytes, 2)
+                raise ValueError(
+                    f"Free memory on device "
+                    f"({GiB(self.init_snapshot.free_memory)}/"
+                    f"{GiB(self.init_snapshot.total_memory)} GiB) on startup "
+                    f"is less than desired GPU memory utilization "
+                    f"({self.cache_config.gpu_memory_utilization}, "
+                    f"{GiB(self.requested_memory)} GiB). Decrease GPU memory "
+                    f"utilization or reduce GPU memory used by other processes."
+                )
+        else:
+            raise RuntimeError(
+                f"Not support device type: {self.device_config.device}")
+        # Initialize the distributed environment.
+        init_worker_distributed_environment(self.vllm_config, self.rank,
+                                            self.distributed_init_method,
+                                            self.local_rank)
+        # Set random seed.
+        set_random_seed(self.model_config.seed)
+
+        # Construct the model runner
+        self.model_runner: GPUModelRunner = GPUModelRunner(
+            self.vllm_config, self.device)
+
+        if self.rank == 0:
+            # If usage stat is enabled, collect relevant info.
+            report_usage_stats(self.vllm_config)
+
+    # FIXME(youkaichao & ywang96): Use TorchDispatchMode instead of memory pool
+    # to hijack tensor allocation.
+    def load_model(self) -> None:
+        if self.vllm_config.model_config.enable_sleep_mode:
+            allocator = CuMemAllocator.get_instance()
+            assert allocator.get_current_usage() == 0, (
+                "Sleep mode can only be "
+                "used for one instance per process.")
+            context = allocator.use_memory_pool(tag="weights")
+        else:
+            from contextlib import nullcontext
+            context = nullcontext()
+        with context:
+            self.model_runner.load_model()
+
+    @torch.inference_mode()
+    def determine_available_memory(self) -> int:
+        """Profiles the peak memory usage of the model to determine how much 
+        memory can be used for KV cache without OOMs.
+
+        The engine will first conduct a profiling of the existing memory usage.
+        Then, it calculate the free memory that can be used for KV cache in
+        bytes.
+
+        Tip:
+            You may limit the usage of GPU memory
+            by adjusting the `gpu_memory_utilization` parameter.
+        """
+        torch.cuda.empty_cache()
+        torch.cuda.reset_peak_memory_stats()
+        GiB = lambda b: b / GiB_bytes
+
+        # Execute a forward pass with dummy inputs to profile the memory usage
+        # of the model.
+        with memory_profiling(
+                self.init_snapshot,
+                weights_memory=int(
+                    self.model_runner.model_memory_usage)) as profile_result:
+            self.model_runner.profile_run()
+
+        free_gpu_memory = profile_result.after_profile.free_memory
+        # 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_snapshot.free_memory > free_gpu_memory, (
+            "Error in memory profiling. "
+            f"Initial free memory {GiB(self.init_snapshot.free_memory)} GiB, "
+            f"current free memory {GiB(free_gpu_memory)} GiB. "
+            "This happens when other processes sharing the same container "
+            "release GPU memory while vLLM is profiling during initialization. "
+            "To fix this, ensure consistent GPU memory allocation or "
+            "isolate vLLM in its own container.")
+        available_kv_cache_memory = self.requested_memory \
+            - profile_result.non_kv_cache_memory
+
+        logger.debug(
+            "Initial free memory: %.2f GiB, free memory: %.2f GiB, "
+            "requested GPU memory: %.2f GiB",
+            GiB(self.init_snapshot.free_memory), GiB(free_gpu_memory),
+            GiB(self.requested_memory))
+        logger.debug(profile_result)
+        logger.info("Available KV cache memory: %.2f GiB",
+                    GiB(available_kv_cache_memory))
+        gc.collect()
+
+        return int(available_kv_cache_memory)
+
+    def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
+        return self.model_runner.get_kv_cache_spec()
+
+    def initialize_from_config(self, kv_cache_config: KVCacheConfig) -> None:
+        """Allocate GPU KV cache with the specified kv_cache_config."""
+        if self.vllm_config.model_config.enable_sleep_mode:
+            allocator = CuMemAllocator.get_instance()
+            context = allocator.use_memory_pool(tag="kv_cache")
+        else:
+            from contextlib import nullcontext
+            context = nullcontext()
+        with context:
+            self.model_runner.initialize_kv_cache(kv_cache_config)
+
+    def compile_or_warm_up_model(self) -> None:
+        # warm up sizes that are not in cudagraph capture sizes,
+        # but users still want to compile for better performance,
+        # e.g. for the max-num-batched token size in chunked prefill.
+        warmup_sizes = self.vllm_config.compilation_config.compile_sizes.copy()
+        if not self.model_config.enforce_eager:
+            warmup_sizes = [
+                x for x in warmup_sizes if x not in
+                self.vllm_config.compilation_config.cudagraph_capture_sizes
+            ]
+        for size in sorted(warmup_sizes, reverse=True):
+            logger.info("Compile and warming up model for size %d", size)
+            self.model_runner._dummy_run(size)
+        if not self.model_config.enforce_eager:
+            self.model_runner.capture_model()
+
+        # Warm up sampler and preallocate memory buffer for logits and other
+        # sampling related tensors of max possible shape to avoid memory
+        # fragmentation issue.
+        # NOTE: This is called after `capture_model` on purpose to prevent
+        # memory buffers from being cleared by `torch.cuda.empty_cache`.
+        if get_pp_group().is_last_rank:
+            max_num_reqs = min(self.scheduler_config.max_num_seqs,
+                               self.scheduler_config.max_num_batched_tokens)
+            self.model_runner._dummy_sampler_run(
+                hidden_states=self.model_runner._dummy_run(
+                    num_tokens=max_num_reqs))
+
+        # Reset the seed to ensure that the random state is not affected by
+        # the model initialization and profiling.
+        set_random_seed(self.model_config.seed)
+
+    def get_model(self) -> nn.Module:
+        return self.model_runner.get_model()
+
+    @torch.inference_mode()
+    def execute_model(
+        self,
+        scheduler_output: "SchedulerOutput",
+    ) -> Optional[ModelRunnerOutput]:
+        intermediate_tensors = None
+        if not get_pp_group().is_first_rank:
+            intermediate_tensors = IntermediateTensors(
+                get_pp_group().recv_tensor_dict(
+                    all_gather_group=get_tp_group()))
+
+        output = self.model_runner.execute_model(scheduler_output,
+                                                 intermediate_tensors)
+        parallel_config = self.vllm_config.parallel_config
+        if parallel_config.distributed_executor_backend != "external_launcher" \
+            and not get_pp_group().is_last_rank:
+            assert isinstance(output, IntermediateTensors)
+            get_pp_group().send_tensor_dict(output.tensors,
+                                            all_gather_group=get_tp_group())
+            return None
+        assert isinstance(output, ModelRunnerOutput)
+        return output if self.is_driver_worker else None
+
+    def profile(self, is_start: bool = True):
+        if self.profiler is None:
+            raise RuntimeError("Profiler is not enabled.")
+        if is_start:
+            self.profiler.start()
+        else:
+            self.profiler.stop()
+            print(self.profiler.key_averages().table(
+                sort_by="self_cuda_time_total"))
+
+    def execute_dummy_batch(self) -> None:
+        self.model_runner._dummy_run(1)
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        return self.model_runner.add_lora(lora_request)
+
+    def remove_lora(self, lora_id: int) -> bool:
+        return self.model_runner.remove_lora(lora_id)
+
+    def list_loras(self) -> set[int]:
+        return self.model_runner.list_loras()
+
+    def pin_lora(self, lora_id: int) -> bool:
+        return self.model_runner.pin_lora(lora_id)
+
+    def check_health(self) -> None:
+        # worker will always be healthy as long as it's running.
+        return
+
+    def save_sharded_state(
+        self,
+        path: str,
+        pattern: Optional[str] = None,
+        max_size: Optional[int] = None,
+    ) -> None:
+        from vllm.model_executor.model_loader import ShardedStateLoader
+        ShardedStateLoader.save_model(
+            self.model_runner.model,
+            path,
+            pattern=pattern,
+            max_size=max_size,
+        )
+
+    def save_tensorized_model(
+        self,
+        tensorizer_config: "TensorizerConfig",
+    ) -> None:
+        self.model_runner.save_tensorized_model(
+            tensorizer_config=tensorizer_config, )
+
+
+def init_worker_distributed_environment(
+    vllm_config: VllmConfig,
+    rank: int,
+    distributed_init_method: Optional[str] = None,
+    local_rank: int = -1,
+    backend: str = "nccl",
+) -> None:
+    """Initialize the distributed environment."""
+    parallel_config = vllm_config.parallel_config
+    set_custom_all_reduce(not parallel_config.disable_custom_all_reduce)
+
+    init_distributed_environment(parallel_config.world_size, rank,
+                                 distributed_init_method, local_rank, backend)
+
+    ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
+                                      parallel_config.pipeline_parallel_size)
+
+    ensure_kv_transfer_initialized(vllm_config)
+
+
+def _check_if_gpu_supports_dtype(torch_dtype: torch.dtype):
+    # Check if the GPU supports the dtype.
+    if torch_dtype == torch.bfloat16:  # noqa: SIM102
+        if not current_platform.has_device_capability(80):
+            capability = current_platform.get_device_capability()
+            gpu_name = current_platform.get_device_name()
+
+            if capability is None:
+                compute_str = "does not have a compute capability"
+            else:
+                version_str = capability.as_version_str()
+                compute_str = f"has compute capability {version_str}"
+
+            raise ValueError(
+                "Bfloat16 is only supported on GPUs with compute capability "
+                f"of at least 8.0. Your {gpu_name} GPU {compute_str}. "
+                "You can use float16 instead by explicitly setting the "
+                "`dtype` flag in CLI, for example: --dtype=half.")

vllm/v1/worker/lora_model_runner_mixin.py

@@ -0,0 +1,173 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Define LoRA functionality mixin for model runners.
+"""
+
+from contextlib import contextmanager
+
+import numpy as np
+import torch.nn as nn
+
+from vllm.config import LoRAConfig, ModelConfig, SchedulerConfig
+from vllm.logger import init_logger
+from vllm.lora.layers import LoRAMapping
+from vllm.lora.request import LoRARequest
+from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
+from vllm.model_executor.models import supports_lora, supports_multimodal
+from vllm.v1.worker.gpu_input_batch import InputBatch
+
+logger = init_logger(__name__)
+
+
+# Defined as a mixin for GPUModelRunner
+class LoRAModelRunnerMixin:
+
+    LORA_WARMUP_RANK = 8
+
+    def load_lora_model(self, model: nn.Module, model_config: ModelConfig,
+                        scheduler_config: SchedulerConfig,
+                        lora_config: LoRAConfig, device: str) -> nn.Module:
+
+        if not supports_lora(model):
+            raise ValueError(
+                f"{model.__class__.__name__} does not support LoRA yet.")
+
+        if supports_multimodal(model):
+            logger.warning("Regarding multimodal models, vLLM currently "
+                           "only supports adding LoRA to language model.")
+
+        # Use get_text_config() in case of multimodal models
+        text_config = model_config.hf_config.get_text_config()
+
+        # Add LoRA Manager to the Model Runner
+        self.lora_manager = LRUCacheWorkerLoRAManager(
+            scheduler_config.max_num_seqs,
+            scheduler_config.max_num_batched_tokens,
+            model_config.get_vocab_size(),
+            lora_config,
+            device,
+            model.embedding_modules,
+            model.embedding_padding_modules,
+            max_position_embeddings=text_config.max_position_embeddings,
+        )
+        return self.lora_manager.create_lora_manager(model)
+
+    def _set_active_loras(self, prompt_lora_mapping: tuple[int, ...],
+                          token_lora_mapping: tuple[int, ...],
+                          lora_requests: set[LoRARequest]) -> None:
+        if not self.lora_manager:
+            raise RuntimeError("LoRA is not enabled.")
+
+        # Set is_prefill to True, so we always use the SGMV kernels on
+        # non-cuda platforms.
+        # On cuda platforms we use the same kernels for prefill and
+        # decode and this flag is generally ignored.
+        lora_mapping = LoRAMapping(token_lora_mapping,
+                                   prompt_lora_mapping,
+                                   is_prefill=True)
+        self.lora_manager.set_active_adapters(lora_requests, lora_mapping)
+
+    def set_active_loras(self, input_batch: InputBatch,
+                         num_scheduled_tokens: np.ndarray) -> None:
+
+        prompt_lora_mapping: tuple[int, ...]  # of size input_batch.num_reqs
+        token_lora_mapping: tuple[int,
+                                  ...]  # of size np.sum(num_scheduled_tokens)
+        lora_requests: set[LoRARequest]
+        prompt_lora_mapping, token_lora_mapping, lora_requests = \
+                            input_batch.make_lora_inputs(num_scheduled_tokens)
+        return self._set_active_loras(prompt_lora_mapping, token_lora_mapping,
+                                      lora_requests)
+
+    @contextmanager
+    def maybe_setup_dummy_loras(self, lora_config):
+        if lora_config is None:
+            yield
+        else:
+            # __enter__ code
+            assert self.lora_manager is not None, "LoRA is not enabled"
+
+            num_loras = lora_config.max_loras
+
+            # Make dummy lora requests
+            lora_requests: set[LoRARequest] = {
+                LoRARequest(lora_name=f"warmup_{lora_id}",
+                            lora_int_id=lora_id,
+                            lora_path="/not/a/real/path")
+                for lora_id in range(1, num_loras + 1)
+            }
+
+            with self.lora_manager.dummy_lora_cache():
+                # Add the dummy LoRAs here so _set_active_loras doesn't try to
+                # load from disk.
+                for lr in lora_requests:
+                    self.lora_manager.add_dummy_lora(
+                        lr, rank=self.LORA_WARMUP_RANK)
+
+                yield
+
+            # __exit__ code
+            self.lora_manager.remove_all_adapters()
+
+    @contextmanager
+    def maybe_select_dummy_loras(self, lora_config: LoRAConfig,
+                                 num_scheduled_tokens: np.ndarray):
+        if lora_config is None:
+            yield
+        else:
+            # __enter__ code
+            assert self.lora_manager is not None, "LoRA is not enabled"
+
+            num_reqs = len(num_scheduled_tokens)
+            num_loras = lora_config.max_loras
+
+            # Make prompt lora mapping
+            # Assign LoRA IDs cyclically to simulate a worst-case scenario.
+            prompt_lora_mapping = (np.arange(num_reqs, dtype=np.int32) %
+                                   num_loras) + 1
+
+            # Make token lora mapping
+            token_lora_mapping = np.repeat(prompt_lora_mapping,
+                                           num_scheduled_tokens)
+
+            # Make dummy lora requests
+            lora_requests: set[LoRARequest] = {
+                LoRARequest(lora_name=f"warmup_{lora_id}",
+                            lora_int_id=lora_id,
+                            lora_path="/not/a/real/path")
+                for lora_id in range(1, num_loras + 1)
+            }
+
+            self._set_active_loras(tuple(prompt_lora_mapping),
+                                   tuple(token_lora_mapping), lora_requests)
+
+            yield
+
+    @contextmanager
+    def maybe_dummy_run_with_lora(self, lora_config: LoRAConfig,
+                                  num_scheduled_tokens: np.ndarray):
+        with self.maybe_setup_dummy_loras(
+                lora_config), self.maybe_select_dummy_loras(
+                    lora_config, num_scheduled_tokens):
+            yield
+
+    def add_lora(self, lora_request: LoRARequest) -> bool:
+        if not self.lora_manager:
+            raise RuntimeError("LoRA is not enabled.")
+        return self.lora_manager.add_adapter(lora_request)
+
+    def remove_lora(self, lora_id: int) -> bool:
+        if not self.lora_manager:
+            raise RuntimeError("LoRA is not enabled.")
+        return self.lora_manager.remove_adapter(lora_id)
+
+    def pin_lora(self, lora_id: int) -> bool:
+        if not self.lora_manager:
+            raise RuntimeError("LoRA is not enabled.")
+        return self.lora_manager.pin_adapter(lora_id)
+
+    def list_loras(self) -> set[int]:
+        if not self.lora_manager:
+            raise RuntimeError("LoRA is not enabled.")
+        return self.lora_manager.list_adapters()