{ai}[foss/2024a] PyTorch v2.9.1 w/ CUDA 12.6.0

[Flamefire]

(created using eb --new-pr)

Early draft. Compiles but not tested

Requires:

• {ai}[foss/2024a] PyTorch v2.7.1 w/ CUDA 12.6.0 #23923
• Fix issues with PyTorch test-results (XML files) parsing and add tests easybuild-easyblocks#3803
• {lib}[gfbf/2024a] nvidia-cutlass v3.8.0.0 w/ CUDA 12.6.0 #23606
• {tools}[GCCcore/13.3.0] tlparse v0.4.0 #24882

[github-actions]

Diff of new easyconfig(s) against existing ones is too long for a GitHub comment. Use --review-pr (and --review-pr-filter / --review-pr-max) locally.

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 0 out of 1 (total: 27 hours 27 mins 43 secs) (1 easyconfigs in total)
c144 - Linux Rocky Linux 9.6, x86_64, AMD EPYC 9334 32-Core Processor (zen4), 4 x NVIDIA NVIDIA H100, 580.65.06, Python 3.9.21
See https://gist.github.com/Flamefire/50cdd13305fd9a33c6140c223aeab6cd for a full test report.

[boegel]

Test report by @boegel
Using easyblocks from PR(s) easybuilders/easybuild-easyblocks#3803
FAILED
Build succeeded for 4 out of 6 (total: 8 mins 45 secs) (6 easyconfigs in total)
node3907.accelgor.os - Linux RHEL 9.6, x86_64, AMD EPYC 7413 24-Core Processor (zen3), 1 x NVIDIA NVIDIA A100-SXM4-80GB, 580.95.05, Python 3.9.21
See https://gist.github.com/boegel/e06f98f956452bcb8a132f816e55927b for a full test report.

[boegel]

Checksum of easybuild/easyconfigs/t/Triton/triton_test.py in develop branch is 02a3390a5dbe27385358ab319cf10972cd8b51aca599a6809efea612a90ecdba ?

See also https://github.com/easybuilders/easybuild-easyconfigs/pull/23120/files#diff-281b568c9515da9719c383a4978f7626f1fae4773e9c38132ee7160b18771e6bR141

[Flamefire]

oh, yes. The upload failed because --update-pr couldn't handle Python files (tried to parse as easyconfig to find destination folder).
Will fix on Monday

[Flamefire]

update added

[boegel]

I'm also seeing a crash with the triton_test.py script:

== FAILED: Installation ended unsuccessfully: Sanity check failed: sanity check command TRITON_HOME=$TMPDIR/eb-triton_home python
/software/Triton/3.5.0-gfbf-2024a-CUDA-12.6.0/test/triton_test.py 8.0 failed with exit code 1 (output: Traceback (most recent call last):
  File "/software/Triton/3.5.0-gfbf-2024a-CUDA-12.6.0/test/triton_test.py", line 13, in <module>
    src = triton.compiler.ASTSource(
          ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/software/Triton/3.5.0-gfbf-2024a-CUDA-12.6.0/lib/python3.12/site-packages/triton/compiler/compiler.py", line 67, in __init__
    for k in self.signature.keys():
             ^^^^^^^^^^^^^^^^^^^
AttributeError: 'str' object has no attribute 'keys'

[Flamefire]

That's why the checksum changed: Breaking change in Triton 3.5 and I updated the test script accordingly. It is in #24793

[Flamefire]

I had to use tlparse 0.4.0 (also separate PR in #24882) as the older one isn't compatible with PyTorch output, see pytorch/pytorch@92c2dae

The lowest tlparse version that works is 0.3.42.

Not sure if this causes conflicts in EB. The alternative is to drop this dependency as it is optional

[Flamefire]

Rebased to remove EasyConfigs present in develop from this branch.

Also added 2 more patches to avoid remaining failures.

[Flamefire]

Test report by @Flamefire
FAILED
Build succeeded for 4 out of 5 (total: 4 mins 48 secs) (5 easyconfigs in total)
c144 - Linux Rocky Linux 9.6, x86_64, AMD EPYC 9334 32-Core Processor (zen4), 4 x NVIDIA NVIDIA H100, 580.65.06, Python 3.9.21
See https://gist.github.com/Flamefire/6f2e5c7a020ec72dff9d2f5c8220fba5 for a full test report.

[Flamefire]

Test report by @Flamefire
SUCCESS
Build succeeded for 5 out of 5 (total: 24 hours 42 mins 21 secs) (5 easyconfigs in total)
c144 - Linux Rocky Linux 9.6, x86_64, AMD EPYC 9334 32-Core Processor (zen4), 4 x NVIDIA NVIDIA H100, 580.65.06, Python 3.9.21
See https://gist.github.com/Flamefire/a6bd885643124f5fac4864060e0e18cd for a full test report.