A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 09:00 UTC
[2/N][Attention] Fix pre-commit errors (#32052)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Jan 9, 2026 at 19:27 UTC
03089019759b4e9af69c160ad0c516e3dcdb1420
3 files changed
+6
-16
tools/pre_commit/mypy.py
-2
@@ -74,8 +74,6 @@ EXCLUDE = [
74
"vllm/model_executor/layers/fla/ops",
75
# Ignore triton kernels in ops.
76
"vllm/v1/attention/ops",
77
- # TODO(matt): remove.
78
- "vllm/v1/attention/backends/fa_utils.py",
77
]
78
79
vllm/v1/attention/backends/fa_utils.py
+4
-8
@@ -7,10 +7,7 @@ from vllm.platforms import current_platform
7
logger = init_logger(__name__)
8
9
if current_platform.is_cuda():
10
- from vllm import _custom_ops
11
-
12
- ops = _custom_ops
13
- reshape_and_cache_flash = ops.reshape_and_cache_flash
10
+ from vllm._custom_ops import reshape_and_cache_flash
11
from vllm.vllm_flash_attn import ( # type: ignore[attr-defined]
12
flash_attn_varlen_func,
13
get_scheduler_metadata,
@@ -19,10 +16,9 @@ if current_platform.is_cuda():
16
elif current_platform.is_xpu():
17
from vllm._ipex_ops import ipex_ops
18
22
- ops = ipex_ops
23
- reshape_and_cache_flash = ops.reshape_and_cache_flash
24
- flash_attn_varlen_func = ops.flash_attn_varlen_func
25
- get_scheduler_metadata = ops.get_scheduler_metadata
19
+ reshape_and_cache_flash = ipex_ops.reshape_and_cache_flash
20
+ flash_attn_varlen_func = ipex_ops.flash_attn_varlen_func
21
+ get_scheduler_metadata = ipex_ops.get_scheduler_metadata
22
23
elif current_platform.is_rocm():
24
try:
vllm/v1/attention/ops/paged_attn.py
+2
-6
@@ -7,13 +7,9 @@ import torch
7
from vllm.platforms import current_platform
8
9
if current_platform.is_cuda_alike():
10
- from vllm import _custom_ops
11
-
12
- ops = _custom_ops
10
+ from vllm import _custom_ops as ops
11
elif current_platform.is_xpu():
14
- from vllm._ipex_ops import ipex_ops
15
-
16
- ops = ipex_ops
12
+ from vllm._ipex_ops import ipex_ops as ops # type: ignore[no-redef]
13
14
15
class PagedAttention: