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 9, 10:00 UTC
[ez] Delete torch25_custom_graph_pass (#33287)
Signed-off-by: angelayi <yiangela7@gmail.com>
Angela Yi committed
Jan 29, 2026 at 08:47 UTC
5a66c9cc760958036ee3f71a78f548b19e89b984
1 file changed
-44
vllm/compilation/torch25_custom_graph_pass.py
deleted
-44
@@ -1,44 +0,0 @@
1
-# SPDX-License-Identifier: Apache-2.0
2
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
-from abc import ABC, abstractmethod
4
-from typing import Any, NoReturn
5
-
6
-import torch
7
-
8
-
9
-class Torch25CustomGraphPass(ABC): # noqa (redefinition)
10
- """
11
- This class replaces CustomGraphPass from torch==2.6 when using torch<2.6.
12
- It conforms to the 2.6 interface but also supports pickling, as that's what
13
- the inductor code cache uses to determine the cache key before 2.6.
14
- (in 2.6 and above, uuid() is used.)
15
-
16
- Subclasses can just "pretend" that uuid is used.
17
- """
18
-
19
- @abstractmethod
20
- def __call__(self, graph: torch.fx.graph.Graph) -> None:
21
- """
22
- Implementation of the custom pass.
23
- """
24
-
25
- @abstractmethod
26
- def uuid(self) -> Any | None:
27
- """
28
- Return an ID to uniquely identify your custom pass implementation.
29
- Return None to skip inductor code caching entirely.
30
- """
31
-
32
- def __getstate__(self) -> Any | None:
33
- """
34
- Pickling is used instead of uuid() in torch<2.6. Just return uuid()
35
- to enable subclasses to only have to implement uuid.
36
- """
37
- return self.uuid()
38
-
39
- def __setstate__(self, state: Any) -> NoReturn:
40
- raise ValueError(
41
- "Cannot unpickle CustomGraphPass because pickling"
42
- " is used for cache key uuid. Use torch>=2.6 with"
43
- " native uuid support for custom passes."
44
- )