Greetings from AWS,
I work with the PyTorch team at AWS, we are interested in testing & validating Inductor’s support for the GPU hardware we offer to customers on AWS.
I went through the integration tests & see that A10 (every PR) & A100 (every night) GPU’s are currently being tested as part of the integration tests. We would like to extend the nightly testing harness’s test matrix to run tests on the supported GPU’s (T4, A10, T4G, V100, A100) offered by AWS. If the community is interested in this we can help with offering the needed hardware on AWS, extending the test harness to run tests on these GPU’s & fixing issues in the running the tests.
The following table captures, current status on Inductor’s GPU support and testing.
| GPU | CUDA | Supported | Tested |
|---|---|---|---|
| M60 | 5 | No | No |
| T4 | 7.5 | Yes | No |
| A10 | 8.6 | Yes | Yes |
| T4G | 7.5 | Yes | No |
| V100 | 7 | Yes | No |
| A100 | 8 | Yes | Yes |
We look forward to working with PyTorch community to have Inductor well supported on all the AWS GPU hardware.