Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Feature] add disable-custom-all-reduce #1148

Merged
merged 4 commits into from
Aug 20, 2024
Merged

[Feature] add disable-custom-all-reduce #1148

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
eddb973
add disable-custom-all-reduce
Aug 16, 2024
5e065b3
fix: isort
Aug 19, 2024
c1c4504
Merge branch 'main' into add-disable_custom_all_reduce
zhyncs Aug 19, 2024
3e92e72
Merge branch 'sgl-project:main' into add-disable_custom_all_reduce
Xu-Chen Aug 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions python/sglang/srt/model_executor/model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,7 @@
get_tp_group,
init_distributed_environment,
initialize_model_parallel,
set_custom_all_reduce,
)
from vllm.distributed.parallel_state import in_the_same_node_as
from vllm.model_executor.model_loader import get_model
Expand Down Expand Up @@ -105,6 +106,7 @@ def __init__(
nccl_init_method = f"tcp://{server_args.nccl_init_addr}"
else:
nccl_init_method = f"tcp://127.0.0.1:{self.nccl_port}"
set_custom_all_reduce(not server_args.disable_custom_all_reduce)
init_distributed_environment(
backend="nccl",
world_size=self.tp_size,
Expand Down
7 changes: 7 additions & 0 deletions python/sglang/srt/server_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,6 +86,7 @@ class ServerArgs:
enable_mla: bool = False
attention_reduce_in_fp32: bool = False
efficient_weight_load: bool = False
disable_custom_all_reduce: bool = False

# Distributed args
nccl_init_addr: Optional[str] = None
Expand Down Expand Up @@ -428,6 +429,12 @@ def add_cli_args(parser: argparse.ArgumentParser):
action="store_true",
help="Turn on memory efficient weight loading with quantization (quantize per layer during loading).",
)
parser.add_argument(
"--disable-custom-all-reduce",
action="store_true",
default=False,
help="Disable the custom all-reduce kernel and fall back to NCCL.",
)

@classmethod
def from_cli_args(cls, args: argparse.Namespace):
Expand Down
Loading