biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Hardware][Intel GPU] Add intel GPU pipeline parallel support. (#7810)

Browse Source
This commit is contained in:
Kunshang Ji
2024-08-28 01:07:02 +08:00
committed by GitHub
parent 9db642138b
commit 076169f603
6 changed files with 82 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
vllm/executor/multiproc_gpu_executor.py
View File

@@ -30,16 +30,12 @@ class MultiprocessingGPUExecutor(DistributedGPUExecutor):
uses_ray: bool = False
def _init_executor(self) -> None:
self._check_executor_parameters()
# Create the parallel GPU workers.
world_size = self.parallel_config.world_size
tensor_parallel_size = self.parallel_config.tensor_parallel_size
# Set CUDA_VISIBLE_DEVICES for the driver, inherited by workers
if "CUDA_VISIBLE_DEVICES" not in os.environ:
update_environment_variables({
"CUDA_VISIBLE_DEVICES": (",".join(map(str, range(world_size))))
})
# Ensure that VLLM_INSTANCE_ID is set, to be inherited by workers
os.environ["VLLM_INSTANCE_ID"] = get_vllm_instance_id()
@@ -68,16 +64,6 @@ class MultiprocessingGPUExecutor(DistributedGPUExecutor):
if world_size > 1:
maybe_set_triton_cache_manager()
cuda_device_count = cuda_device_count_stateless()
# Use confusing message for more common TP-only case.
assert tensor_parallel_size <= cuda_device_count, (
f"please set tensor_parallel_size ({tensor_parallel_size}) "
f"to less than max local gpu count ({cuda_device_count})")
assert world_size <= cuda_device_count, (
f"please ensure that world_size ({world_size}) "
f"is less than than max local gpu count ({cuda_device_count})")
# Multiprocessing-based executor does not support multi-node setting.
# Since it only works for single node, we can use the loopback address
# 127.0.0.1 for communication.
@@ -139,6 +125,26 @@ class MultiprocessingGPUExecutor(DistributedGPUExecutor):
max_concurrent_workers=self.parallel_config.
max_parallel_loading_workers)
def _check_executor_parameters(self):
world_size = self.parallel_config.tensor_parallel_size
tensor_parallel_size = self.parallel_config.tensor_parallel_size
# Set CUDA_VISIBLE_DEVICES for the driver, inherited by workers
if "CUDA_VISIBLE_DEVICES" not in os.environ:
update_environment_variables({
"CUDA_VISIBLE_DEVICES": (",".join(map(str, range(world_size))))
})
cuda_device_count = cuda_device_count_stateless()
# Use confusing message for more common TP-only case.
assert tensor_parallel_size <= cuda_device_count, (
f"please set tensor_parallel_size ({tensor_parallel_size}) "
f"to less than max local gpu count ({cuda_device_count})")
assert world_size <= cuda_device_count, (
f"please ensure that world_size ({world_size}) "
f"is less than than max local gpu count ({cuda_device_count})")
def shutdown(self):
if (worker_monitor := getattr(self, "worker_monitor",
None)) is not None: