[Perf][ROCm] Add AITER custom AG/RS (DP only) (#48247)

Signed-off-by: simondanielsson <simon.danielsson99@hotmail.com>
This commit is contained in:
Simon Danielsson
2026-09-11 13:19:57 +08:00
committed by GitHub
parent 980c16c8e4
commit ae48466cf3
8 changed files with 549 additions and 74 deletions
+12 -10
View File
@@ -673,11 +673,11 @@ class GroupCoordinator:
else:
stream = graph_capture_context.stream
# only cuda uses this function,
# only cuda/rocm uses this function,
# so we don't abstract it into the base class
maybe_ca_context = nullcontext()
maybe_fi_pcie_ipc_context: AbstractContextManager[Any] = nullcontext()
maybe_aiter_context = nullcontext()
maybe_aiter_ar_context = nullcontext()
from vllm.distributed.device_communicators.cuda_communicator import (
CudaCommunicator,
)
@@ -698,12 +698,10 @@ class GroupCoordinator:
if fi_pcie_ipc_ar_comm is not None:
maybe_fi_pcie_ipc_context = fi_pcie_ipc_ar_comm.capture()
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
# Capture each group's own comm. A global lookup would double-capture
aiter_ar_comm = getattr(self.device_communicator, "aiter_ar_comm", None)
if aiter_ar_comm is not None:
maybe_aiter_ar_context = aiter_ar_comm.capture() # type: ignore
# ensure all initialization operations complete before attempting to
# capture the graph on another stream
@@ -715,7 +713,7 @@ class GroupCoordinator:
torch.cuda.stream(stream),
maybe_ca_context,
maybe_fi_pcie_ipc_context,
maybe_aiter_context,
maybe_aiter_ar_context,
):
yield graph_capture_context
@@ -1640,7 +1638,11 @@ def graph_capture(
context = graph_capture_context or GraphCaptureContext(
torch.cuda.Stream(device=device)
)
with get_tp_group().graph_capture(context), get_pp_group().graph_capture(context):
with (
get_tp_group().graph_capture(context),
get_pp_group().graph_capture(context),
get_dp_group().graph_capture(context),
):
yield context