[XPU][9/N] clean up existing ipex code/doc (#34111)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Kunshang Ji
2026-02-11 00:27:15 -08:00
committed by GitHub
parent 21dfb842d7
commit cb9574eb85
10 changed files with 16 additions and 50 deletions
+4 -5
View File
@@ -23,12 +23,11 @@ if current_platform.is_cuda():
elif current_platform.is_xpu():
from vllm import _custom_ops as ops
from vllm._xpu_ops import xpu_ops
reshape_and_cache_flash = ops.reshape_and_cache_flash
from vllm._ipex_ops import ipex_ops
flash_attn_varlen_func = ipex_ops.flash_attn_varlen_func # type: ignore[assignment]
get_scheduler_metadata = ipex_ops.get_scheduler_metadata # type: ignore[assignment]
flash_attn_varlen_func = xpu_ops.flash_attn_varlen_func # type: ignore[assignment]
get_scheduler_metadata = xpu_ops.get_scheduler_metadata # type: ignore[assignment]
elif current_platform.is_rocm():
try:
from flash_attn import flash_attn_varlen_func # type: ignore[no-redef]
@@ -153,7 +152,7 @@ def is_flash_attn_varlen_func_available() -> bool:
Platform-specific sources:
- CUDA: vllm.vllm_flash_attn.flash_attn_varlen_func
- XPU: ipex_ops.flash_attn_varlen_func
- XPU: xpu_ops.flash_attn_varlen_func
- ROCm: upstream flash_attn.flash_attn_varlen_func (if available)
Note: This is separate from the AITER flash attention backend (rocm_aiter_fa.py)
+1 -1
View File
@@ -9,7 +9,7 @@ from vllm.platforms import current_platform
if current_platform.is_cuda_alike():
from vllm import _custom_ops as ops
elif current_platform.is_xpu():
from vllm._ipex_ops import ipex_ops as ops # type: ignore[no-redef]
from vllm._xpu_ops import xpu_ops as ops # type: ignore[no-redef]
class PagedAttention: