[core] add nccl symmetric memory for all reduce (#24532)

Signed-off-by: Amir Samani <asamani@nvidia.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Amir Samani
2025-09-23 14:33:06 -04:00
committed by GitHub
co-authored by Michael Goin
parent a3a7828010
commit 8c1c81a3de
12 changed files with 489 additions and 6 deletions
+32
View File
@@ -1383,6 +1383,38 @@ def find_nccl_library() -> str:
return so_file
def find_nccl_include_paths() -> Optional[list[str]]:
"""
We either use the nccl.h specified by the `VLLM_NCCL_INCLUDE_PATH`
environment variable, or we find the library file brought by
nvidia-nccl-cuXX. load_inline by default uses
torch.utils.cpp_extension.include_paths
"""
paths: list[str] = []
inc = envs.VLLM_NCCL_INCLUDE_PATH
if inc and os.path.isdir(inc):
paths.append(inc)
try:
import importlib.util
spec = importlib.util.find_spec("nvidia.nccl")
if spec and getattr(spec, "submodule_search_locations", None):
for loc in spec.submodule_search_locations:
inc_dir = os.path.join(loc, "include")
if os.path.exists(os.path.join(inc_dir, "nccl.h")):
paths.append(inc_dir)
except Exception:
pass
seen = set()
out: list[str] = []
for p in paths:
if p and p not in seen:
out.append(p)
seen.add(p)
return out or None
prev_set_stream = torch.cuda.set_stream
_current_stream_tls = threading.local()