Skip to content

【Paddle Tensor No.27】 paddle.to_tensor 适配 __cuda_array_interface__#69913

Merged
HydrogenSulfate merged 13 commits intoPaddlePaddle:developfrom
aquagull:__cuda_array_interface__
Dec 6, 2024
Merged

【Paddle Tensor No.27】 paddle.to_tensor 适配 __cuda_array_interface__#69913
HydrogenSulfate merged 13 commits intoPaddlePaddle:developfrom
aquagull:__cuda_array_interface__

Conversation

@aquagull
Copy link
Contributor

@aquagull aquagull commented Dec 3, 2024 •
edited
Loading

PR Category

User Experience

PR Types

New features

Description

paddle.to_tensor 适配 __cuda_array_interface__，对实现了__cuda_array_interface__的非paddle.Tensor对象（如torch.Tensor、cupy.NdArray）均可通过其__cuda_array_interface__接口，以零拷贝的方式转换为paddle.Tensor

def check_tensor_conversion(self, x, check_version = True):
    """Helper function to check tensor conversion."""
    paddle_tensor = paddle.to_tensor(x, place=paddle.CUDAPlace(0))

    self.assertEqual(x.__cuda_array_interface__["typestr"], paddle_tensor.__cuda_array_interface__["typestr"])
    self.assertEqual(x.__cuda_array_interface__["shape"], paddle_tensor.__cuda_array_interface__["shape"])
    self.assertEqual(x.__cuda_array_interface__["strides"], paddle_tensor.__cuda_array_interface__["strides"])
    self.assertEqual(x.__cuda_array_interface__["data"], paddle_tensor.__cuda_array_interface__["data"])
    if check_version:
        self.assertEqual(x.__cuda_array_interface__["version"], paddle_tensor.__cuda_array_interface__["version"])

# 测试cupy
def test_cuda_tensor_cupy(self):
    if not torch.cuda.is_available():
        return
    shapes = [[2], [2, 3], [2, 3, 4], [2, 3, 4, 5]]
    dtypes = [cp.float32, cp.float64, cp.int32, cp.int64]

    for shape in shapes:
        for dtype in dtypes:
            x = cp.random.random(tuple(shape)).astype(dtype)
            self.check_tensor_conversion(x, False)

# 测试pytorch
def test_cuda_array_torch(self):
    if not paddle.is_compiled_with_cuda():
        return  
    shapes = [[2], [2, 3], [2, 3, 4], [2, 3, 4, 5]] 
    dtypes = [np.float32, np.float64, np.int32, np.int64] 

    for shape in shapes:
        for dtype in dtypes:
            x = torch.tensor(np.random.rand(*shape).astype(dtype), device='cuda')
            self.check_tensor_conversion(x)
# 测试0-size tensor
def test_cuda_array_exception(self):
    def test_cuda_array_zero_size():
        shape = [0]
        x = torch.tensor(np.random.rand(*shape), device='cuda')
        self.check_tensor_conversion(x)
    self.assertRaises(ValueError, test_cuda_array_zero_size)
# 测试内存
def test_for(self):
        for i in range(100):  
            x_torch = torch.randn([2048, 1024], device='cuda')  
            x_paddle = paddle.to_tensor(x_torch) 
            print(f"{i} paddle mem: {paddle.device.cuda.max_memory_allocated() / (1 << 20):.2f} MB, torch mem: {torch.cuda.max_memory_allocated() / (1 << 20):.2f} MB")
# “测试内存”输出
0 paddle mem: 0.00 MB, torch mem: 8.00 MB
1 paddle mem: 0.00 MB, torch mem: 16.00 MB
2 paddle mem: 0.00 MB, torch mem: 16.00 MB
3 paddle mem: 0.00 MB, torch mem: 16.00 MB
4 paddle mem: 0.00 MB, torch mem: 16.00 MB
5 paddle mem: 0.00 MB, torch mem: 16.00 MB
6 paddle mem: 0.00 MB, torch mem: 16.00 MB
7 paddle mem: 0.00 MB, torch mem: 16.00 MB
8 paddle mem: 0.00 MB, torch mem: 16.00 MB
9 paddle mem: 0.00 MB, torch mem: 16.00 MB
10 paddle mem: 0.00 MB, torch mem: 16.00 MB
11 paddle mem: 0.00 MB, torch mem: 16.00 MB
12 paddle mem: 0.00 MB, torch mem: 16.00 MB
13 paddle mem: 0.00 MB, torch mem: 16.00 MB
14 paddle mem: 0.00 MB, torch mem: 16.00 MB
15 paddle mem: 0.00 MB, torch mem: 16.00 MB
16 paddle mem: 0.00 MB, torch mem: 16.00 MB
17 paddle mem: 0.00 MB, torch mem: 16.00 MB
18 paddle mem: 0.00 MB, torch mem: 16.00 MB
19 paddle mem: 0.00 MB, torch mem: 16.00 MB
20 paddle mem: 0.00 MB, torch mem: 16.00 MB
21 paddle mem: 0.00 MB, torch mem: 16.00 MB
22 paddle mem: 0.00 MB, torch mem: 16.00 MB
23 paddle mem: 0.00 MB, torch mem: 16.00 MB
24 paddle mem: 0.00 MB, torch mem: 16.00 MB
25 paddle mem: 0.00 MB, torch mem: 16.00 MB
26 paddle mem: 0.00 MB, torch mem: 16.00 MB
27 paddle mem: 0.00 MB, torch mem: 16.00 MB
28 paddle mem: 0.00 MB, torch mem: 16.00 MB
29 paddle mem: 0.00 MB, torch mem: 16.00 MB
30 paddle mem: 0.00 MB, torch mem: 16.00 MB
31 paddle mem: 0.00 MB, torch mem: 16.00 MB
32 paddle mem: 0.00 MB, torch mem: 16.00 MB
33 paddle mem: 0.00 MB, torch mem: 16.00 MB
34 paddle mem: 0.00 MB, torch mem: 16.00 MB
35 paddle mem: 0.00 MB, torch mem: 16.00 MB
36 paddle mem: 0.00 MB, torch mem: 16.00 MB
37 paddle mem: 0.00 MB, torch mem: 16.00 MB
38 paddle mem: 0.00 MB, torch mem: 16.00 MB
39 paddle mem: 0.00 MB, torch mem: 16.00 MB
40 paddle mem: 0.00 MB, torch mem: 16.00 MB
41 paddle mem: 0.00 MB, torch mem: 16.00 MB
42 paddle mem: 0.00 MB, torch mem: 16.00 MB
43 paddle mem: 0.00 MB, torch mem: 16.00 MB
44 paddle mem: 0.00 MB, torch mem: 16.00 MB
45 paddle mem: 0.00 MB, torch mem: 16.00 MB
46 paddle mem: 0.00 MB, torch mem: 16.00 MB
47 paddle mem: 0.00 MB, torch mem: 16.00 MB
48 paddle mem: 0.00 MB, torch mem: 16.00 MB
49 paddle mem: 0.00 MB, torch mem: 16.00 MB
50 paddle mem: 0.00 MB, torch mem: 16.00 MB
51 paddle mem: 0.00 MB, torch mem: 16.00 MB
52 paddle mem: 0.00 MB, torch mem: 16.00 MB
53 paddle mem: 0.00 MB, torch mem: 16.00 MB
54 paddle mem: 0.00 MB, torch mem: 16.00 MB
55 paddle mem: 0.00 MB, torch mem: 16.00 MB
56 paddle mem: 0.00 MB, torch mem: 16.00 MB
57 paddle mem: 0.00 MB, torch mem: 16.00 MB
58 paddle mem: 0.00 MB, torch mem: 16.00 MB
59 paddle mem: 0.00 MB, torch mem: 16.00 MB
60 paddle mem: 0.00 MB, torch mem: 16.00 MB
61 paddle mem: 0.00 MB, torch mem: 16.00 MB
62 paddle mem: 0.00 MB, torch mem: 16.00 MB
63 paddle mem: 0.00 MB, torch mem: 16.00 MB
64 paddle mem: 0.00 MB, torch mem: 16.00 MB
65 paddle mem: 0.00 MB, torch mem: 16.00 MB
66 paddle mem: 0.00 MB, torch mem: 16.00 MB
67 paddle mem: 0.00 MB, torch mem: 16.00 MB
68 paddle mem: 0.00 MB, torch mem: 16.00 MB
69 paddle mem: 0.00 MB, torch mem: 16.00 MB
70 paddle mem: 0.00 MB, torch mem: 16.00 MB
71 paddle mem: 0.00 MB, torch mem: 16.00 MB
72 paddle mem: 0.00 MB, torch mem: 16.00 MB
73 paddle mem: 0.00 MB, torch mem: 16.00 MB
74 paddle mem: 0.00 MB, torch mem: 16.00 MB
75 paddle mem: 0.00 MB, torch mem: 16.00 MB
76 paddle mem: 0.00 MB, torch mem: 16.00 MB
77 paddle mem: 0.00 MB, torch mem: 16.00 MB
78 paddle mem: 0.00 MB, torch mem: 16.00 MB
79 paddle mem: 0.00 MB, torch mem: 16.00 MB
80 paddle mem: 0.00 MB, torch mem: 16.00 MB
81 paddle mem: 0.00 MB, torch mem: 16.00 MB
82 paddle mem: 0.00 MB, torch mem: 16.00 MB
83 paddle mem: 0.00 MB, torch mem: 16.00 MB
84 paddle mem: 0.00 MB, torch mem: 16.00 MB
85 paddle mem: 0.00 MB, torch mem: 16.00 MB
86 paddle mem: 0.00 MB, torch mem: 16.00 MB
87 paddle mem: 0.00 MB, torch mem: 16.00 MB
88 paddle mem: 0.00 MB, torch mem: 16.00 MB
89 paddle mem: 0.00 MB, torch mem: 16.00 MB
90 paddle mem: 0.00 MB, torch mem: 16.00 MB
91 paddle mem: 0.00 MB, torch mem: 16.00 MB
92 paddle mem: 0.00 MB, torch mem: 16.00 MB
93 paddle mem: 0.00 MB, torch mem: 16.00 MB
94 paddle mem: 0.00 MB, torch mem: 16.00 MB
95 paddle mem: 0.00 MB, torch mem: 16.00 MB
96 paddle mem: 0.00 MB, torch mem: 16.00 MB
97 paddle mem: 0.00 MB, torch mem: 16.00 MB
98 paddle mem: 0.00 MB, torch mem: 16.00 MB
99 paddle mem: 0.00 MB, torch mem: 16.00 MB

@paddle-bot
Copy link

paddle-bot bot commented Dec 3, 2024

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

@paddle-bot paddle-bot bot added the contributor External developers label Dec 3, 2024
@aquagull aquagull changed the title 【Paddle Tensor No.27】 【Paddle Tensor No.27】 适配__cuda_array_interface__ Dec 3, 2024
HydrogenSulfate
HydrogenSulfate reviewed Dec 3, 2024
View reviewed changes
@HydrogenSulfate HydrogenSulfate changed the title 【Paddle Tensor No.27】 适配__cuda_array_interface__ 【Paddle Tensor No.27】 to_tensor 适配 __cuda_array_interface__ Dec 3, 2024
@HydrogenSulfate HydrogenSulfate changed the title 【Paddle Tensor No.27】 to_tensor 适配 __cuda_array_interface__ 【Paddle Tensor No.27】 paddle.to_tensor 适配 __cuda_array_interface__ Dec 3, 2024
@luotao1 luotao1 added the HappyOpenSource 快乐开源活动issue与PR label Dec 3, 2024
HydrogenSulfate
HydrogenSulfate reviewed Dec 4, 2024
View reviewed changes
@luotao1 luotao1 mentioned this pull request Dec 4, 2024
Closed
@HydrogenSulfate
Copy link
Contributor

HydrogenSulfate commented Dec 5, 2024

@aquagull 和pytorch、cupy的单测贴一下到PR描述中

HydrogenSulfate
HydrogenSulfate reviewed Dec 5, 2024
View reviewed changes
@aquagull
Copy link
Contributor Author

aquagull commented Dec 5, 2024

@aquagull 和pytorch、cupy的单测贴一下到PR描述中

已补上

SigureMo
SigureMo reviewed Dec 6, 2024
View reviewed changes
HydrogenSulfate
HydrogenSulfate reviewed Dec 6, 2024
View reviewed changes
SigureMo
SigureMo approved these changes Dec 6, 2024
View reviewed changes
Copy link
Member

@SigureMo SigureMo left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTMeow 🐾

HydrogenSulfate
HydrogenSulfate approved these changes Dec 6, 2024
View reviewed changes
Copy link
Contributor

@HydrogenSulfate HydrogenSulfate left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@HydrogenSulfate HydrogenSulfate merged commit 6eba01b into PaddlePaddle:develop Dec 6, 2024
@HydrogenSulfate HydrogenSulfate mentioned this pull request Dec 9, 2024
Closed
@aquagull aquagull deleted the __cuda_array_interface__ branch December 13, 2024 09:20
@HydrogenSulfate HydrogenSulfate mentioned this pull request Mar 3, 2025
Merged
gmarkall pushed a commit to NVIDIA/numba-cuda that referenced this pull request Mar 3, 2025
@HydrogenSulfate
Add paddle to interoperability chapter (#144)
964248d 
We support `paddle.Tensor.__cuda_array_interface__`(PaddlePaddle/Paddle#68192) and `paddle.to_tensor(tenosr_with_CAI_impl)`(PaddlePaddle/Paddle#69913), so we can convert paddle Tensor from/to other DL frameworks via CAI(based on CAI version 2).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@HydrogenSulfate HydrogenSulfate HydrogenSulfate approved these changes

@SigureMo SigureMo SigureMo approved these changes

@zrr1999 zrr1999 Awaiting requested review from zrr1999

@gouzil gouzil Awaiting requested review from gouzil

Assignees

@luotao1 luotao1

@HydrogenSulfate HydrogenSulfate

Labels

contributor External developers HappyOpenSource 快乐开源活动issue与PR

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@aquagull @HydrogenSulfate @SigureMo @luotao1