[DLPACK] Setting device_id in FromDLPACK

[HydrogenSulfate]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

Pcard-75624

Issued in NVIDIA/warp#318 (comment), device_id should be set in TensorFromDLPack when in GPU env.
cc @shi-eric

before:

import paddle

x_gpu2 = paddle.randn([2, 2]).to(device='gpu:2')

t_gpu2 = paddle.utils.dlpack.to_dlpack(x_gpu2)

x_gpu2_ref = paddle.utils.dlpack.from_dlpack(t_gpu2)

print(x_gpu2.place, x_gpu2_ref.place)
Place(gpu:2) Place(gpu:0) # device_id of x_gpu2_ref.place is not not set correctly

after fixing:

import paddle

x_gpu2 = paddle.randn([2, 2]).to(device='gpu:2')

t_gpu2 = paddle.utils.dlpack.to_dlpack(x_gpu2)

x_gpu2_ref = paddle.utils.dlpack.from_dlpack(t_gpu2)

print(x_gpu2.place, x_gpu2_ref.place)
Place(gpu:2) Place(gpu:2) # x_gpu2_ref.place is same as x_gpu2

[paddle-bot]

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.

[SigureMo]

LGTMeow