Wei Wang

Reference: https://github.com/pytorch/builder/pull/1376/ cc @ptrblck @atalman @malfet For Windows: cudnn 9.0.0 was incorrectly specified initially but later cudnn 8.9 was used. Magma: re-using git clone https://github.com/ptrblck/magma_win.git magma

Add CUDA 12.4 workflows

13

Reference: https://github.com/pytorch/pytorch/pull/98492 cc @albanD @ptrblck @atalman @malfet

[CUDA] [CI]: Enable CUDA 12.4 CI

12

Reference PR: https://github.com/pytorch/pytorch/pull/93406 cc @atalman @malfet @ptrblck @eqy

[CUDA] [CI] Add cu124 docker images

1

Fixes issues encountered in https://github.com/pytorch/pytorch/pull/121956 cc @ptrblck @Aidyn-A @atalman @malfet

PyTest 8.2.0 in compatible with certain versions of pytorch, especially DDP tests

1

_ ERROR collecting test/distributed/_composable/fsdp/test_fully_shard_comm.py __ Traceback (most recent call last): File "/usr/local/lib/python3.10/dist-packages/_pytest/runner.py", line 341, in from_call result: Optional[TResult] = func() File "/usr/local/lib/python3.10/dist-packages/_pytest/runner.py", line 389, in collect return list(collector.collect()) File "/usr/local/lib/python3.10/dist-packages/_pytest/unittest.py",...

[Do not merge] Test for "Cuda124 arm ci"

1

Will this fixes issues in #126174 ? cc @tinglvv

and make error message more clear when apex.transformer is explicitly called on unsupported platform In some configurations, pytorch may not be built with distributed support. So we avoid the auto...

Enhance cuda upgrade doc

1

This PR documents all cuda 12.4 related PRs across pytorch/builder (majority), pytorch/pytorch, pytorch/test-infra. cc @atalman @malfet @ptrblck @tinglvv

upgrade cudnn to v9.1.0.70 for cu118/121/124

1

also: Remove cuda117 related TODO as it is already deprecated Supporting: https://github.com/pytorch/builder/pull/1822 and https://github.com/pytorch/pytorch/pull/123475 Reference PR: https://github.com/pytorch/test-infra/pull/1523 cc @ptrblck @eqy @atalman @malfet

See PR comments in #1984 for details. Summary: (all aarch64 specific) CUDA 12.4.0 wheel (e.g. pick 08/01/2024 nightly) works with conda numpy, 1.x as well as 2.x CUDA 12.4.1 wheel...