[Core] Pipeline Parallel Support (#4412)

Signed-off-by: Muralidhar Andoorveedu <muralidhar.andoorveedu@centml.ai>
This commit is contained in:
Murali Andoorveedu
2024-07-02 10:58:08 -07:00
committed by GitHub
parent 15aba081f3
commit c5832d2ae9
82 changed files with 1096 additions and 400 deletions
+6 -6
View File
@@ -69,7 +69,7 @@ class DistributedGPUExecutor(GPUExecutor):
if self.parallel_worker_tasks is None:
self.parallel_worker_tasks = self._run_workers(
"start_worker_execution_loop",
async_run_remote_workers_only=True,
async_run_tensor_parallel_workers_only=True,
**self.extra_execute_model_run_workers_kwargs)
# Only the driver worker returns the sampling results.
@@ -138,17 +138,17 @@ class DistributedGPUExecutor(GPUExecutor):
self,
method: str,
*args,
async_run_remote_workers_only: bool = False,
async_run_tensor_parallel_workers_only: bool = False,
max_concurrent_workers: Optional[int] = None,
**kwargs,
) -> Any:
"""Runs the given method on all workers.
Args:
async_run_remote_workers_only: If True the method will be run only
in the remote workers, not the driver worker. It will also be
run asynchronously and return a list of futures rather than
blocking on the results.
async_run_tensor_parallel_workers_only: If True the method will be
run only in the remote TP workers, not the driver worker.
It will also be run asynchronously and return a list of futures
rather than blocking on the results.
"""
raise NotImplementedError