[WideEP] Update NCCL to 2.30.7 to enable DeepEPv2 in the vllm/vllm-openai image (#45321)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com> Signed-off-by: Tyler Michael Smith <tyler@vllm.ai> Signed-off-by: Tyler Michael Smith <tyler@tylermsmith.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Ilya Markov <ilmarkov@users.noreply.github.com> Co-authored-by: OpenAI Codex <codex@openai.com> Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
co-authored by
Claude
Ilya Markov
OpenAI Codex
Codex
parent
e222c33f2f
commit
5d8e90a966
@@ -36,6 +36,10 @@ if has_deep_ep_v2():
|
||||
P = ParamSpec("P")
|
||||
|
||||
|
||||
class GINNotAvailableError(RuntimeError):
|
||||
pass
|
||||
|
||||
|
||||
@dataclasses.dataclass
|
||||
class ProcessGroupInfo:
|
||||
world_size: int
|
||||
@@ -96,19 +100,27 @@ def parallel_launch(
|
||||
**kwargs: P.kwargs,
|
||||
) -> None:
|
||||
assert not kwargs
|
||||
spawn(
|
||||
_worker_parallel_launch,
|
||||
args=(
|
||||
world_size,
|
||||
world_size,
|
||||
0,
|
||||
f"tcp://{os.getenv('LOCALHOST', 'localhost')}:{get_open_port()}",
|
||||
worker,
|
||||
try:
|
||||
spawn(
|
||||
_worker_parallel_launch,
|
||||
args=(
|
||||
world_size,
|
||||
world_size,
|
||||
0,
|
||||
f"tcp://{os.getenv('LOCALHOST', 'localhost')}:{get_open_port()}",
|
||||
worker,
|
||||
)
|
||||
+ args,
|
||||
nprocs=world_size,
|
||||
join=True,
|
||||
)
|
||||
+ args,
|
||||
nprocs=world_size,
|
||||
join=True,
|
||||
)
|
||||
except Exception as exc:
|
||||
# pytest.skip cannot propagate directly through torch.multiprocessing.
|
||||
if "GINNotAvailableError" in str(exc):
|
||||
import pytest
|
||||
|
||||
pytest.skip("NCCL GIN not available (no IBGDA-capable hardware)")
|
||||
raise
|
||||
|
||||
|
||||
## DeepEP specific utils
|
||||
@@ -225,6 +237,18 @@ def make_deepep_v2_a2a(
|
||||
):
|
||||
import deep_ep
|
||||
|
||||
from vllm.utils.nccl import query_nccl_gin_type
|
||||
|
||||
# ElasticBuffer can segfault when GIN is unavailable. Initialize the
|
||||
# lazy communicator and reject unsupported systems before entering DeepEP.
|
||||
probe = torch.zeros(1, device=pgi.device)
|
||||
torch.distributed.all_reduce(probe, group=pg)
|
||||
gin_type = query_nccl_gin_type(pg)
|
||||
if gin_type is None:
|
||||
raise RuntimeError("Failed to determine NCCL GIN support")
|
||||
if gin_type == 0:
|
||||
raise GINNotAvailableError("NCCL GIN not available")
|
||||
|
||||
buffer = deep_ep.ElasticBuffer(
|
||||
group=pg,
|
||||
num_max_tokens_per_rank=v2_args.max_tokens_per_rank,
|
||||
|
||||
Reference in New Issue
Block a user