[XPU] update xpu graph usage (#43043)

Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>
This commit is contained in:
Xinyu Chen
2026-05-19 23:09:07 +08:00
committed by GitHub
parent b82e908b4c
commit d740e2c029
4 changed files with 9 additions and 19 deletions
@@ -22,6 +22,7 @@ class XpuCommunicator(DeviceCommunicatorBase):
unique_name: str = "",
):
super().__init__(cpu_group, device, device_group, unique_name)
self.ca_comm: None = None
if self.use_all2all:
if self.all2all_backend in ("naive", "allgather_reducescatter"):
from .all2all import AgRsAll2AllManager
+7 -1
View File
@@ -476,9 +476,15 @@ class GroupCoordinator:
from vllm.distributed.device_communicators.cuda_communicator import (
CudaCommunicator,
)
from vllm.distributed.device_communicators.xpu_communicator import (
XpuCommunicator,
)
if self.device_communicator is not None:
assert isinstance(self.device_communicator, CudaCommunicator)
assert isinstance(
self.device_communicator,
(CudaCommunicator, XpuCommunicator),
)
ca_comm = self.device_communicator.ca_comm
if ca_comm is not None:
maybe_ca_context = ca_comm.capture() # type: ignore
-17
View File
@@ -195,23 +195,6 @@ class XPUPlatform(Platform):
"XPU Graph is disabled by environment variable, "
"please set VLLM_XPU_ENABLE_XPU_GRAPH=1 to enable it."
)
elif parallel_config.world_size_across_dp > 1:
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
logger.warning(
"XPU Graph doesn't support capture communication ops, "
"disabling cudagraph_mode."
)
else:
if (
attention_config.backend == AttentionBackendEnum.FLASH_ATTN
and compilation_config.cudagraph_mode
not in {CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE}
):
compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
logger.warning(
"FMHA sycl-tla kernels cannot be captured with XPU graphs, "
"falling back to PIECEWISE graph mode on XPU platform."
)
# Disable fusion passes not yet supported on XPU.
pass_config = compilation_config.pass_config
+1 -1
View File
@@ -294,7 +294,7 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
# https://github.com/vllm-project/vllm/issues/22945
_cudagraph_support = (
AttentionCGSupport.ALWAYS
if get_flash_attn_version() == 3
if get_flash_attn_version() == 3 or current_platform.is_xpu()
else AttentionCGSupport.UNIFORM_BATCH
)
supports_update_block_table: bool = True