xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-20 05:07:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

[misc][perf] remove old code (#11425)

Browse Source 
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao 2024-12-23 00:01:08 -08:00 committed by GitHub
parent 048fc57a0f
commit f30581c518
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 0 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
vllm/_custom_ops.py
View File

@ -1,5 +1,4 @@
import contextlib import contextlib
import functools
import importlib import importlib
from typing import TYPE_CHECKING, List, Optional, Tuple, Union from typing import TYPE_CHECKING, List, Optional, Tuple, Union
@ -36,34 +35,6 @@ else:
from torch.library import impl_abstract as register_fake from torch.library import impl_abstract as register_fake
def hint_on_error(fn):
@functools.wraps(fn)
def wrapper(*args, **kwargs):
try:
return fn(*args, **kwargs)
except NotImplementedError as e:
msg = (
"Error in calling custom op %s: %s\n"
"Not implemented or built, mostly likely because the current current device "
"does not support this kernel (less likely TORCH_CUDA_ARCH_LIST was set "
"incorrectly while building)")
logger.error(msg, fn.__name__, e)
raise NotImplementedError(msg % (fn.__name__, e)) from e
except AttributeError as e:
msg = (
"Error in calling custom op %s: %s\n"
"Possibly you have built or installed an obsolete version of vllm.\n"
"Please try a clean build and install of vllm,"
"or remove old built files such as vllm/*cpython*.so and build/ ."
)
logger.error(msg, fn.__name__, e)
raise e
return wrapper
# activation ops # activation ops
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None: def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
torch.ops._C.silu_and_mul(out, x) torch.ops._C.silu_and_mul(out, x)
@ -1101,25 +1072,3 @@ def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
def register_graph_buffers(fa: int, handles: List[List[int]], def register_graph_buffers(fa: int, handles: List[List[int]],
offsets: List[List[int]]) -> None: offsets: List[List[int]]) -> None:
torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets) torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
# temporary fix for https://github.com/vllm-project/vllm/issues/5456
# TODO: remove this in v0.6.0
names_and_values = globals()
names_and_values_to_update = {}
# prepare variables to avoid dict size change during iteration
k, v, arg = None, None, None
fn_type = type(lambda x: x)
for k, v in names_and_values.items():
# find functions that are defined in this file and have torch.Tensor
# in their annotations. `arg == "torch.Tensor"` is used to handle
# the case when users use `import __annotations__` to turn type
# hints into strings.
if isinstance(v, fn_type) \
and v.__code__.co_filename == __file__ \
and any(arg is torch.Tensor or arg == "torch.Tensor"
for arg in v.__annotations__.values()):
names_and_values_to_update[k] = hint_on_error(v)
names_and_values.update(names_and_values_to_update)
del names_and_values_to_update, names_and_values, v, k, fn_type