[ROCm][FEAT] AITER Fused Allreduce + RMSNorm (#37646)

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: Rita Brugarolas Brufau <rita.brugarolasbrufau@amd.com>
Signed-off-by: junkang1991 <junkangchow@gmail.com>
Co-authored-by: Rita Brugarolas <Rita.BrugarolasBrufau@amd.com>
Co-authored-by: junkang1991 <junkangchow@gmail.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
vllmellm
2026-05-01 23:07:18 +08:00
committed by GitHub
co-authored by Rita Brugarolas junkang1991 Luka Govedič TJian
parent bc635fad23
commit 529c671e80
9 changed files with 454 additions and 20 deletions
+9 -1
View File
@@ -472,6 +472,7 @@ class GroupCoordinator:
# only cuda uses this function,
# so we don't abstract it into the base class
maybe_ca_context = nullcontext()
maybe_aiter_context = nullcontext()
from vllm.distributed.device_communicators.cuda_communicator import (
CudaCommunicator,
)
@@ -482,13 +483,20 @@ class GroupCoordinator:
if ca_comm is not None:
maybe_ca_context = ca_comm.capture() # type: ignore
from vllm._aiter_ops import rocm_aiter_ops
if rocm_aiter_ops.is_enabled():
aiter_ar = rocm_aiter_ops.get_aiter_allreduce()
if aiter_ar is not None:
maybe_aiter_context = aiter_ar.capture() # type: ignore
# ensure all initialization operations complete before attempting to
# capture the graph on another stream
curr_stream = torch.cuda.current_stream()
if curr_stream != stream:
stream.wait_stream(curr_stream)
with torch.cuda.stream(stream), maybe_ca_context:
with torch.cuda.stream(stream), maybe_ca_context, maybe_aiter_context:
yield graph_capture_context
def all_reduce(self, input_: torch.Tensor) -> torch.Tensor: