[CPU] V1 support for the CPU backend (#16441)
This commit is contained in:
@@ -0,0 +1,86 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
from contextlib import contextmanager
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.model_loader import get_model
|
||||
from vllm.v1.worker.gpu_model_runner import GPUModelRunner
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class CPUModelRunner(GPUModelRunner):
|
||||
|
||||
def __init__(self, vllm_config: VllmConfig, device: torch.device):
|
||||
super().__init__(vllm_config, device)
|
||||
|
||||
assert device == torch.device("cpu")
|
||||
assert self.speculative_config is None, "spec decode is not supported."
|
||||
|
||||
self.use_cuda_graph = False
|
||||
self.cascade_attn_enabled = False
|
||||
|
||||
self._postprocess_tenosrs()
|
||||
|
||||
def _postprocess_tenosrs(self) -> None:
|
||||
# Note: replace device tensors with cpu tensors
|
||||
def replace_tensor(obj: Any, cpu_attr_name: str,
|
||||
device_attr_name) -> None:
|
||||
cpu_tensor = getattr(obj, cpu_attr_name, None)
|
||||
device_tensor = getattr(obj, device_attr_name, None)
|
||||
if cpu_tensor is not None and device_tensor is not None:
|
||||
assert isinstance(cpu_tensor, torch.Tensor)
|
||||
assert isinstance(device_tensor, torch.Tensor)
|
||||
setattr(obj, device_attr_name, cpu_tensor)
|
||||
|
||||
for k, v in vars(self).items():
|
||||
if k.endswith("_cpu") and isinstance(v, torch.Tensor):
|
||||
replace_tensor(self, k, k[:-4])
|
||||
|
||||
for k, v in vars(self.input_batch).items():
|
||||
if k.endswith("_cpu_tensor") and isinstance(v, torch.Tensor):
|
||||
replace_tensor(self.input_batch, k, k[:-11])
|
||||
|
||||
for k, v in vars(self.input_batch.block_table).items():
|
||||
if k.endswith("_cpu") and isinstance(v, torch.Tensor):
|
||||
replace_tensor(self.input_batch.block_table, k, k[:-4])
|
||||
|
||||
def load_model(self) -> None:
|
||||
logger.info("Starting to load model %s...", self.model_config.model)
|
||||
self.model = get_model(vllm_config=self.vllm_config)
|
||||
|
||||
if self.lora_config:
|
||||
self.model = self.load_lora_model(self.model, self.model_config,
|
||||
self.scheduler_config,
|
||||
self.lora_config, self.device)
|
||||
|
||||
def warming_up_model(self) -> None:
|
||||
logger.info("Warming up model for the compilation...")
|
||||
# Only generate graph for the generic shape
|
||||
self._dummy_run(max(16, self.max_num_reqs))
|
||||
logger.info("Warming up done.")
|
||||
|
||||
def _init_device_properties(self) -> None:
|
||||
pass
|
||||
|
||||
def _sync_device(self) -> None:
|
||||
pass
|
||||
|
||||
|
||||
@contextmanager
|
||||
def _set_global_compilation_settings():
|
||||
import torch._inductor.config
|
||||
|
||||
# Note: The CPPGEMM backend requires freezing parameters.
|
||||
freezing_value = torch._inductor.config.freezing
|
||||
torch._inductor.config.freezing = True
|
||||
# Note: workaround for "ValueError: fast mode: can't pickle cyclic objects
|
||||
# including object type dict"
|
||||
force_disable_caches = torch._inductor.config.force_disable_caches
|
||||
torch._inductor.config.force_disable_caches = True
|
||||
yield
|
||||
torch._inductor.config.freezing = freezing_value
|
||||
torch._inductor.config.force_disable_caches = force_disable_caches
|
||||
@@ -0,0 +1,101 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
import os
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
|
||||
from vllm import envs
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.distributed.parallel_state import get_pp_group, get_tp_group
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.utils import set_random_seed
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.v1.core.sched.output import SchedulerOutput
|
||||
from vllm.v1.outputs import ModelRunnerOutput
|
||||
from vllm.v1.worker.cpu_model_runner import CPUModelRunner
|
||||
from vllm.v1.worker.gpu_worker import (Worker,
|
||||
init_worker_distributed_environment)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class CPUWorker(Worker):
|
||||
|
||||
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=is_driver_worker)
|
||||
|
||||
self.parallel_config.disable_custom_all_reduce = True
|
||||
|
||||
def init_device(self):
|
||||
# Setup OpenMP threads affinity.
|
||||
omp_cpuids = envs.VLLM_CPU_OMP_THREADS_BIND
|
||||
if omp_cpuids == "all":
|
||||
self.local_omp_cpuid = "all"
|
||||
else:
|
||||
self.local_omp_cpuid = omp_cpuids.split("|")[self.rank]
|
||||
ret = torch.ops._C_utils.init_cpu_threads_env(self.local_omp_cpuid)
|
||||
if ret:
|
||||
logger.info(ret)
|
||||
|
||||
# Note: unique identifier for creating allreduce shared memory
|
||||
os.environ["VLLM_DIST_IDENT"] = self.distributed_init_method.split(
|
||||
":")[-1]
|
||||
# Initialize the distributed environment.
|
||||
init_worker_distributed_environment(self.vllm_config, self.rank,
|
||||
self.distributed_init_method,
|
||||
self.local_rank, "gloo")
|
||||
# Set random seed.
|
||||
set_random_seed(self.model_config.seed)
|
||||
|
||||
# Construct the model runner
|
||||
self.model_runner: CPUModelRunner = CPUModelRunner(
|
||||
self.vllm_config, torch.device("cpu"))
|
||||
|
||||
def sleep(self, level: int = 1) -> None:
|
||||
logger.warning("sleep mode is not supported on CPU, ignore it.")
|
||||
pass
|
||||
|
||||
def wake_up(self, tags: Optional[list[str]] = None) -> None:
|
||||
logger.warning("sleep mode is not supported on CPU, ignore it.")
|
||||
pass
|
||||
|
||||
def determine_available_memory(self) -> int:
|
||||
return self.cache_config.cpu_kvcache_space_bytes # type: ignore
|
||||
|
||||
def compile_or_warm_up_model(self) -> None:
|
||||
# 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)
|
||||
self.model_runner.warming_up_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)
|
||||
|
||||
if 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
|
||||
@@ -5,7 +5,7 @@ import copy
|
||||
import gc
|
||||
import time
|
||||
import weakref
|
||||
from typing import TYPE_CHECKING, Optional, Union
|
||||
from typing import TYPE_CHECKING, Any, Optional, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
@@ -38,7 +38,6 @@ from vllm.sequence import IntermediateTensors
|
||||
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
|
||||
GiB_bytes, LazyLoader, async_tensor_h2d, cdiv,
|
||||
check_use_alibi, is_pin_memory_available)
|
||||
from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
|
||||
from vllm.v1.attention.backends.utils import CommonAttentionMetadata
|
||||
from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
|
||||
from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
|
||||
@@ -203,8 +202,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
self.vllm_config.compilation_config.cudagraph_capture_sizes))
|
||||
|
||||
# Cache the device properties.
|
||||
self.device_properties = torch.cuda.get_device_properties(self.device)
|
||||
self.num_sms = self.device_properties.multi_processor_count
|
||||
self._init_device_properties()
|
||||
|
||||
# Persistent buffers for CUDA graphs.
|
||||
self.input_ids = torch.zeros(self.max_num_tokens,
|
||||
@@ -315,6 +313,17 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
self.input_batch, scheduler_output)
|
||||
return batch_reordered
|
||||
|
||||
# Note: used for model runner override.
|
||||
def _init_device_properties(self) -> None:
|
||||
"""Initialize attributes from torch.cuda.get_device_properties
|
||||
"""
|
||||
self.device_properties = torch.cuda.get_device_properties(self.device)
|
||||
self.num_sms = self.device_properties.multi_processor_count
|
||||
|
||||
# Note: used for model runner override.
|
||||
def _sync_device(self) -> None:
|
||||
torch.cuda.synchronize()
|
||||
|
||||
def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
|
||||
"""Update the cached states and the persistent batch with the scheduler
|
||||
output.
|
||||
@@ -538,8 +547,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
def _prepare_inputs(
|
||||
self,
|
||||
scheduler_output: "SchedulerOutput",
|
||||
) -> tuple[dict[str, FlashAttentionMetadata], torch.Tensor,
|
||||
Optional[SpecDecodeMetadata]]:
|
||||
) -> tuple[dict[str, Any], torch.Tensor, Optional[SpecDecodeMetadata]]:
|
||||
total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
|
||||
assert total_num_scheduled_tokens > 0
|
||||
num_reqs = self.input_batch.num_reqs
|
||||
@@ -652,7 +660,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
common_attn_metadata = CommonAttentionMetadata(
|
||||
query_start_loc=query_start_loc, seq_lens=seq_lens)
|
||||
|
||||
attn_metadata: dict[str, FlashAttentionMetadata] = {}
|
||||
attn_metadata: dict[str, Any] = {}
|
||||
# Prepare the attention metadata for each KV cache group and make layers
|
||||
# in the same group share the same metadata.
|
||||
for kv_cache_group_id, kv_cache_group_spec in enumerate(
|
||||
@@ -1710,7 +1718,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
|
||||
# Must synchronize the non-blocking GPU->CPU transfers.
|
||||
if prompt_logprobs_dict:
|
||||
torch.cuda.synchronize()
|
||||
self._sync_device()
|
||||
|
||||
return prompt_logprobs_dict
|
||||
|
||||
@@ -1740,7 +1748,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
dtype=np.int32)
|
||||
|
||||
if skip_attn:
|
||||
attn_metadata: Optional[dict[str, FlashAttentionMetadata]] = None
|
||||
attn_metadata: Optional[dict[str, Any]] = None
|
||||
else:
|
||||
query_start_loc = self.query_start_loc[:num_reqs + 1]
|
||||
seq_lens = self.seq_lens[:num_reqs]
|
||||
@@ -1964,7 +1972,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
|
||||
sampler_output = self._dummy_sampler_run(hidden_states)
|
||||
else:
|
||||
sampler_output = None
|
||||
torch.cuda.synchronize()
|
||||
self._sync_device()
|
||||
del hidden_states, sampler_output
|
||||
self.encoder_cache.clear()
|
||||
gc.collect()
|
||||
|
||||
@@ -342,13 +342,14 @@ def init_worker_distributed_environment(
|
||||
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)
|
||||
distributed_init_method, local_rank, backend)
|
||||
|
||||
ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
|
||||
parallel_config.pipeline_parallel_size)
|
||||
|
||||
Reference in New Issue
Block a user