update doc to use PJRT_DEVICE=CUDA instead of PJRT_DEVICE=GPU

[vanbasten23]

#5675 discovered that runtime has replaced PJRT_DEVICE GPU with cuda and rocm. This PR updates our GPU documentation.

Test via:

• GPU_NUM_DEVICES=1 PJRT_DEVICE=GPU python pytorch/xla/test/test_train_mp_imagenet.py --fake_data
• PJRT_DEVICE=CUDA torchrun --nnodes 1 --nproc-per-node 4 pytorch/xla/test/test_train_mp_imagenet.py --fake_data --pjrt_distributed --batch_size=128 --num_epochs=1
• GPU CI

[zpcore]

Thanks for the update. Do we also want to remove the GPU from the device list here: https://github.com/pytorch/xla/blob/1f0e972a6dfc5ed3205b23d0fccac8ebb128584f/torch_xla/core/xla_model.py#L92C1-L92C1? There are several places still mention GPU choice in xla_model.py file.

[JackCaoG]

Hey @vanbasten23 can we make sure PJRT_DEVICE=GPU continue to work and add warning message to let user to set it to CUDA.

[vanbasten23]

Thanks for the update. Do we also want to remove the GPU from the device list here: https://github.com/pytorch/xla/blob/1f0e972a6dfc5ed3205b23d0fccac8ebb128584f/torch_xla/core/xla_model.py#L92C1-L92C1? There are several places still mention GPU choice in xla_model.py file.

Yeah, that's a good point. I plan to remove it in a follow-up PR

[vanbasten23]

Hey @vanbasten23 can we make sure PJRT_DEVICE=GPU continue to work and add warning message to let user to set it to CUDA.

FWIW, it's already causing some problem: if I use our nightly (11/1/23), I see:

root@xiowei-gpu-1:/# PJRT_DEVICE=GPU python
Python 3.8.18 (default, Oct 12 2023, 10:35:13)
[GCC 10.2.1 20210110] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch_xla.core.xla_model as xm
>>> import torch, torch_xla
>>> xm.get_xla_supported_devices(devkind='GPU')
# it returns nothing

On the contrary, r2.1 returns ['xla:0']. Let me see if there is an easy fix

[zpcore]

Hey @vanbasten23 can we make sure PJRT_DEVICE=GPU continue to work and add warning message to let user to set it to CUDA.

FWIW, it's already causing some problem: if I use our nightly (11/1/23), I see:

root@xiowei-gpu-1:/# PJRT_DEVICE=GPU python
Python 3.8.18 (default, Oct 12 2023, 10:35:13)
[GCC 10.2.1 20210110] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch_xla.core.xla_model as xm
>>> import torch, torch_xla
>>> xm.get_xla_supported_devices(devkind='GPU')
# it returns nothing

On the contrary, r2.1 returns ['xla:0']. Let me see if there is an easy fix

Yes, that's what we saw when run on GPU recently.
We added devkind = "CUDA" if devkind == "GPU" else devkind before xla_devices = _DEVICES.value line for a temporary fix in order to run on GPU.

[vanbasten23]

can we make sure PJRT_DEVICE=GPU continue to work and add warning message to let user to set it to CUDA.

hi @JackCaoG , I've added warning messages and fix the currently known bug.
Regarding "make sure PJRT_DEVICE=GPU continue to work", we're already in a mixed state where GPU CI uses CUDA but our public documentation uses GPU. IMO, the only way to make sure of that is to run GPU CI with PJRT_DEVICE=GPU but that sounds too much overhead. How about we just do our best effort?

[JackCaoG]

@vanbasten23 just have one test run with PJRT_DEVICE=GPU(explcity set env var when running test or something) and rest run with PJRT_DEVICE=CUDA(default)

[vanbasten23]

GPU CI test TestDynamicShapeModels is failing:

test_backward_pass_with_dynamic_input (__main__.TestDynamicShapeModels) ... 2023-11-03 17:22:04.313597: E external/xla/xla/status_macros.cc:54] INTERNAL: RET_CHECK failure (external/xla/xla/service/dynamic_padder.cc:1989) op_support != OpDynamismSupport::kNoSupport Dynamic input unexpectedly found for unsupported instruction: %add = f32[<=10,1]{1,0} add(f32[<=10,1]{1,0} %broadcast.2, f32[10,1]{1,0} %exponential)

and it's likely due to the recent pin update. A likely culprit is openxla/xla@33bcc66.

The test has not been run since last pin update because

xla/test/ds/test_dynamic_shape_models.py

Line 46 in a93e14b

not xm.get_xla_supported_devices("GPU") and

would silently fail and it's always been skipped.

Luckily, another cl is trying to revert openxla/xla@33bcc66. I'll skip this test on GPU for now and only let it run on TPU (on TPU, the test succeeds.).

[vanbasten23]

Thanks for the review!