[Refactor] Share workspace and model runner init between GPU and XPU workers (#59200)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
eef8b15d7c
commit
643bbef55c
@@ -522,6 +522,10 @@ class Worker(WorkerBase):
|
|||||||
else:
|
else:
|
||||||
raise RuntimeError(f"Unsupported device type: {self.device_config.device}")
|
raise RuntimeError(f"Unsupported device type: {self.device_config.device}")
|
||||||
|
|
||||||
|
self._init_workspace_and_model_runner()
|
||||||
|
|
||||||
|
def _init_workspace_and_model_runner(self) -> None:
|
||||||
|
"""Set up the workspace manager, build the model runner, report usage."""
|
||||||
# DSpark target and draft CUDA graphs retain workspace views concurrently.
|
# DSpark target and draft CUDA graphs retain workspace views concurrently.
|
||||||
num_ubatches = 2 if self.vllm_config.parallel_config.enable_dbo else 1
|
num_ubatches = 2 if self.vllm_config.parallel_config.enable_dbo else 1
|
||||||
init_workspace_manager(
|
init_workspace_manager(
|
||||||
@@ -529,8 +533,13 @@ class Worker(WorkerBase):
|
|||||||
num_ubatches,
|
num_ubatches,
|
||||||
_num_workspace_lanes(self.vllm_config, self.use_v2_model_runner),
|
_num_workspace_lanes(self.vllm_config, self.use_v2_model_runner),
|
||||||
)
|
)
|
||||||
|
self.model_runner: GPUModelRunner = self._make_model_runner()
|
||||||
|
if self.rank == 0:
|
||||||
|
# If usage stat is enabled, collect relevant info.
|
||||||
|
report_usage_stats(self.vllm_config)
|
||||||
|
|
||||||
# Construct the model runner
|
def _make_model_runner(self) -> "GPUModelRunner":
|
||||||
|
"""Construct the platform's model runner; subclasses override this."""
|
||||||
if self.use_v2_model_runner:
|
if self.use_v2_model_runner:
|
||||||
if self.vllm_config.is_mm_encoder_only:
|
if self.vllm_config.is_mm_encoder_only:
|
||||||
from vllm.v1.worker.mm_encoder_model_runner import (
|
from vllm.v1.worker.mm_encoder_model_runner import (
|
||||||
@@ -542,19 +551,10 @@ class Worker(WorkerBase):
|
|||||||
)
|
)
|
||||||
|
|
||||||
# HACK(woosuk): This is a temporary fix to avoid type errors.
|
# HACK(woosuk): This is a temporary fix to avoid type errors.
|
||||||
self.model_runner: GPUModelRunner = GPUModelRunnerV2( # type: ignore
|
return GPUModelRunnerV2(self.vllm_config, self.device) # type: ignore
|
||||||
self.vllm_config, self.device
|
from vllm.v1.worker.gpu_model_runner import GPUModelRunner as GPUModelRunnerV1
|
||||||
)
|
|
||||||
else:
|
|
||||||
from vllm.v1.worker.gpu_model_runner import (
|
|
||||||
GPUModelRunner as GPUModelRunnerV1,
|
|
||||||
)
|
|
||||||
|
|
||||||
self.model_runner = GPUModelRunnerV1(self.vllm_config, self.device)
|
return GPUModelRunnerV1(self.vllm_config, self.device)
|
||||||
|
|
||||||
if self.rank == 0:
|
|
||||||
# If usage stat is enabled, collect relevant info.
|
|
||||||
report_usage_stats(self.vllm_config)
|
|
||||||
|
|
||||||
def handle_ft_command(self, ft_request):
|
def handle_ft_command(self, ft_request):
|
||||||
assert self.worker_sentinel is not None
|
assert self.worker_sentinel is not None
|
||||||
|
|||||||
@@ -10,9 +10,7 @@ from vllm.logger import init_logger
|
|||||||
from vllm.platforms import current_platform
|
from vllm.platforms import current_platform
|
||||||
from vllm.utils.mem_utils import MemorySnapshot, format_gib
|
from vllm.utils.mem_utils import MemorySnapshot, format_gib
|
||||||
from vllm.utils.torch_utils import set_random_seed
|
from vllm.utils.torch_utils import set_random_seed
|
||||||
from vllm.v1.utils import report_usage_stats
|
|
||||||
from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
|
from vllm.v1.worker.gpu_worker import Worker, init_worker_distributed_environment
|
||||||
from vllm.v1.worker.workspace import init_workspace_manager
|
|
||||||
from vllm.v1.worker.xpu_model_runner import (
|
from vllm.v1.worker.xpu_model_runner import (
|
||||||
XPUMMEncoderModelRunner,
|
XPUMMEncoderModelRunner,
|
||||||
XPUModelRunner,
|
XPUModelRunner,
|
||||||
@@ -214,11 +212,9 @@ class XPUWorker(Worker):
|
|||||||
"worker requested memory: %sGiB", format_gib(self.requested_memory)
|
"worker requested memory: %sGiB", format_gib(self.requested_memory)
|
||||||
)
|
)
|
||||||
|
|
||||||
# Initialize workspace manager
|
self._init_workspace_and_model_runner()
|
||||||
num_ubatches = 2 if self.vllm_config.parallel_config.enable_dbo else 1
|
|
||||||
init_workspace_manager(self.device, num_ubatches)
|
|
||||||
|
|
||||||
# Construct the model runner
|
def _make_model_runner(self):
|
||||||
if self.use_v2_model_runner:
|
if self.use_v2_model_runner:
|
||||||
model_runner = (
|
model_runner = (
|
||||||
XPUMMEncoderModelRunner
|
XPUMMEncoderModelRunner
|
||||||
@@ -227,13 +223,7 @@ class XPUWorker(Worker):
|
|||||||
)
|
)
|
||||||
else:
|
else:
|
||||||
model_runner = XPUModelRunner # type: ignore[assignment]
|
model_runner = XPUModelRunner # type: ignore[assignment]
|
||||||
self.model_runner = model_runner( # type: ignore
|
return model_runner(self.vllm_config, self.device) # type: ignore
|
||||||
self.vllm_config, self.device
|
|
||||||
)
|
|
||||||
|
|
||||||
if self.rank == 0:
|
|
||||||
# If usage stat is enabled, collect relevant info.
|
|
||||||
report_usage_stats(self.vllm_config)
|
|
||||||
|
|
||||||
def shutdown(self) -> None:
|
def shutdown(self) -> None:
|
||||||
logger.info(
|
logger.info(
|
||||||
|
|||||||
Reference in New Issue
Block a user