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 16, 19:30 UTC
[TPU] Suppress import custom_ops warning (#7458)
Woosuk Kwon committed
Aug 13, 2024 at 00:30 UTC
d6e634f3d78649002460758f3964e4df0d39a546
2 files changed
+7
-5
vllm/_custom_ops.py
+6
-4
@@ -6,13 +6,15 @@ import torch
6
7
from vllm._core_ext import ScalarType
8
from vllm.logger import init_logger
9
+from vllm.platforms import current_platform
10
11
logger = init_logger(__name__)
12
12
-try:
13
- import vllm._C
14
-except ImportError as e:
15
- logger.warning("Failed to import from vllm._C with %r", e)
13
+if not current_platform.is_tpu():
14
+ try:
15
+ import vllm._C
16
+ except ImportError as e:
17
+ logger.warning("Failed to import from vllm._C with %r", e)
18
19
with contextlib.suppress(ImportError):
20
# ruff: noqa: F401
vllm/utils.py
+1
-1
@@ -29,7 +29,6 @@ import torch.types
29
from typing_extensions import ParamSpec, TypeIs, assert_never
30
31
import vllm.envs as envs
32
-from vllm import _custom_ops as ops
32
from vllm.logger import enable_trace_function_call, init_logger
33
34
logger = init_logger(__name__)
@@ -359,6 +358,7 @@ def is_xpu() -> bool:
358
@lru_cache(maxsize=None)
359
def get_max_shared_memory_bytes(gpu: int = 0) -> int:
360
"""Returns the maximum shared memory per thread block in bytes."""
361
+ from vllm import _custom_ops as ops
362
max_shared_mem = (
363
ops.get_max_shared_memory_per_block_device_attribute(gpu))
364
# value 0 will cause MAX_SEQ_LEN become negative and test_attention.py