@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / f30581c518
Fork 0

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, 04:30 UTC

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

Signed-off-by: youkaichao <youkaichao@gmail.com>

youkaichao committed Dec 23, 2024 at 00:01 UTC f30581c51831b74795cb55419b2fffc928cfd7d2
1 file changed -51
vllm/_custom_ops.py
-51
@@ -1,5 +1,4 @@
1 import contextlib
2 -import functools
2 import importlib
3 from typing import TYPE_CHECKING, List, Optional, Tuple, Union
4
@@ -36,34 +35,6 @@ else:
35 from torch.library import impl_abstract as register_fake
36
37
39 -def hint_on_error(fn):
40 -
41 - @functools.wraps(fn)
42 - def wrapper(*args, **kwargs):
43 - try:
44 - return fn(*args, **kwargs)
45 -
46 - except NotImplementedError as e:
47 - msg = (
48 - "Error in calling custom op %s: %s\n"
49 - "Not implemented or built, mostly likely because the current current device "
50 - "does not support this kernel (less likely TORCH_CUDA_ARCH_LIST was set "
51 - "incorrectly while building)")
52 - logger.error(msg, fn.__name__, e)
53 - raise NotImplementedError(msg % (fn.__name__, e)) from e
54 - except AttributeError as e:
55 - msg = (
56 - "Error in calling custom op %s: %s\n"
57 - "Possibly you have built or installed an obsolete version of vllm.\n"
58 - "Please try a clean build and install of vllm,"
59 - "or remove old built files such as vllm/*cpython*.so and build/ ."
60 - )
61 - logger.error(msg, fn.__name__, e)
62 - raise e
63 -
64 - return wrapper
65 -
66 -
38 # activation ops
39 def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
40 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]]:
1072 def register_graph_buffers(fa: int, handles: List[List[int]],
1073 offsets: List[List[int]]) -> None:
1074 torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
1104 -
1105 -
1106 -# temporary fix for https://github.com/vllm-project/vllm/issues/5456
1107 -# TODO: remove this in v0.6.0
1108 -names_and_values = globals()
1109 -names_and_values_to_update = {}
1110 -# prepare variables to avoid dict size change during iteration
1111 -k, v, arg = None, None, None
1112 -fn_type = type(lambda x: x)
1113 -for k, v in names_and_values.items():
1114 - # find functions that are defined in this file and have torch.Tensor
1115 - # in their annotations. `arg == "torch.Tensor"` is used to handle
1116 - # the case when users use `import __annotations__` to turn type
1117 - # hints into strings.
1118 - if isinstance(v, fn_type) \
1119 - and v.__code__.co_filename == __file__ \
1120 - and any(arg is torch.Tensor or arg == "torch.Tensor"
1121 - for arg in v.__annotations__.values()):
1122 - names_and_values_to_update[k] = hint_on_error(v)
1123 -
1124 -names_and_values.update(names_and_values_to_update)
1125 -del names_and_values_to_update, names_and_values, v, k, fn_type