[v1] torchrun compatibility (#13642)
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
@@ -541,7 +541,7 @@ class RayDistributedExecutor(DistributedExecutorBase):
|
||||
# and the TP group executes in SPMD fashion.
|
||||
if self.use_v1:
|
||||
outputs = [
|
||||
worker.execute_model.
|
||||
worker.execute_model_ray.
|
||||
bind( # type: ignore[attr-defined]
|
||||
outputs[i]) for i, worker in enumerate(tp_group)
|
||||
]
|
||||
|
||||
@@ -112,10 +112,12 @@ try:
|
||||
torch.cuda.set_device(self.worker.device)
|
||||
self.compiled_dag_cuda_device_set = True
|
||||
|
||||
def execute_model(
|
||||
def execute_model_ray(
|
||||
self,
|
||||
scheduler_output: "SchedulerOutput",
|
||||
) -> "ModelRunnerOutput":
|
||||
# this method is used to compile ray CG,
|
||||
# and it needs a special logic of self.setup_device_if_necessary()
|
||||
self.setup_device_if_necessary()
|
||||
assert self.worker is not None, "Worker is not initialized"
|
||||
if isinstance(scheduler_output, tuple):
|
||||
|
||||
@@ -93,9 +93,10 @@ class ExecutorWithExternalLauncher(UniProcExecutor):
|
||||
("ExecutorWithExternalLauncher needs deterministic "
|
||||
"execution, so it"
|
||||
"does not support delay_factor in scheduling")
|
||||
assert not envs.VLLM_USE_V1, \
|
||||
("V1 architecture cannot guarantee deterministic execution, "
|
||||
"so it is not supported in ExecutorWithExternalLauncher.")
|
||||
if envs.VLLM_USE_V1:
|
||||
assert not envs.VLLM_ENABLE_V1_MULTIPROCESSING, \
|
||||
("To get deterministic execution in V1, "
|
||||
"please set VLLM_ENABLE_V1_MULTIPROCESSING=0")
|
||||
self.driver_worker = WorkerWrapperBase(vllm_config=self.vllm_config,
|
||||
rpc_rank=0)
|
||||
# engines are launched in torchrun-compatible launchers
|
||||
|
||||
Reference in New Issue
Block a user