[Kernel] Refactor FlashInfer allreduce for mnnvl backend (#34109)

Signed-off-by: hjjq <50634613+hjjq@users.noreply.github.com>
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
This commit is contained in:
Hanjie Qiu
2026-02-26 03:17:20 +00:00
committed by GitHub
co-authored by wzhao18 Wei Zhao
parent 2aa4140402
commit 71dfce6aa6
7 changed files with 593 additions and 180 deletions
@@ -34,19 +34,25 @@ class CudaCommunicator(DeviceCommunicatorBase):
# custom allreduce or torch symm mem can be used only by tp
use_custom_allreduce = False
use_torch_symm_mem = False
use_flashinfer_allreduce = False
else:
from vllm.distributed.parallel_state import _ENABLE_CUSTOM_ALL_REDUCE
use_custom_allreduce = _ENABLE_CUSTOM_ALL_REDUCE
use_torch_symm_mem = envs.VLLM_ALLREDUCE_USE_SYMM_MEM
use_flashinfer_allreduce = envs.VLLM_ALLREDUCE_USE_FLASHINFER
self.use_custom_allreduce = use_custom_allreduce
self.use_torch_symm_mem = use_torch_symm_mem
self.use_flashinfer_allreduce = use_flashinfer_allreduce
# lazy import to avoid documentation build error
from vllm.distributed.device_communicators.custom_all_reduce import (
CustomAllreduce,
)
from vllm.distributed.device_communicators.flashinfer_all_reduce import (
FlashInferAllReduce,
)
from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
from vllm.distributed.device_communicators.quick_all_reduce import (
QuickAllReduce,
@@ -65,12 +71,20 @@ class CudaCommunicator(DeviceCommunicatorBase):
self.ca_comm: CustomAllreduce | None = None
self.qr_comm: QuickAllReduce | None = None
self.symm_mem_comm: SymmMemCommunicator | None = None
self.fi_ar_comm: FlashInferAllReduce | None = None
if use_torch_symm_mem and current_platform.is_cuda():
self.symm_mem_comm = SymmMemCommunicator(
group=self.cpu_group,
device=self.device,
)
if self.use_flashinfer_allreduce and self.world_size > 1:
self.fi_ar_comm = FlashInferAllReduce(
group=self.cpu_group,
device=self.device,
)
if use_custom_allreduce and self.world_size > 1:
# Initialize a custom fast all-reduce implementation.
self.ca_comm = CustomAllreduce(
@@ -136,7 +150,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
out = torch.ops.vllm.all_reduce_symmetric_with_copy(input_)
if out is not None:
return out
# always try quick reduce first, then custom allreduce,
# always try quick reduce first, then flashinfer, then custom allreduce,
# and then pynccl. (quick reduce just for ROCM MI3*)
qr_comm = self.qr_comm
if (
@@ -147,6 +161,15 @@ class CudaCommunicator(DeviceCommunicatorBase):
out = qr_comm.quick_all_reduce(input_)
assert out is not None
return out
fi_ar_comm = self.fi_ar_comm
if (
fi_ar_comm is not None
and not fi_ar_comm.disabled
and fi_ar_comm.should_use_fi_ar(input_)
):
out = fi_ar_comm.all_reduce(input_)
assert out is not None
return out
ca_comm = self.ca_comm
if (
ca_comm is not None
@@ -270,6 +293,9 @@ class CudaCommunicator(DeviceCommunicatorBase):
self.pynccl_comm = None
if self.ca_comm is not None:
self.ca_comm = None
if self.fi_ar_comm is not None:
self.fi_ar_comm.destroy()
self.fi_ar_comm = None
if self.all2all_manager is not None:
self.all2all_manager.destroy()
self.all2all_manager = None