Triton pin update for PyTorch 2.8 / Triton 3.4

[iupaikov-amd]

Creating an issue to track tasks related to possible triton commit bump for PyTorch release/2.8.

Creating this to assess the possibility of bumping the triton pin for upcoming pytorch/2.8 release. Going to link any issues related to this here.

Please list any associated issues and PRs in the comments for tracking purposes.

workflow testing PRs

• [release] Triton pin update to 3.4 #156664
• [TEST] Triton 3.4.0 pin update #156186
• [TESTING] [DO NOT MERGE] Updated triton commit pin #155117
• hud release/2.8 : https://hud.pytorch.org/hud/pytorch/pytorch/release%2F2.8/1?per_page=50&mergeEphemeralLF=true

Currently testing triton commit: triton-lang/triton@6564840

1146 total failing tests:

High priority general issues

• 1007 Triton has removed the experimental descriptor API #154162 (@davidberard98)
• 1+ [Upstream Triton] 'tt.expand_dims' op failed to infer returned types (test_comprehensive_nanquantile_cuda_float64) #154933 (@davidberard98)
• 1+ [Upstream Triton] nvcc / ptx mismatch w/ AOTI #154938 (@davidberard98)
• [Upstream Triton] AOTI support w/ new TMA API #155047
• 1 [Upstream Triton] AssertionError: False is not true test_inductor_profiling_triton_hooks #154223 -> [triton pin][tests] update inductor/profiler launch_(enter|exit)_hooks tests #154894 (@davidberard98)
• Advance version 3.3.0->3.4.0 triton-lang/triton#7031 (@atalman)
• Move inductor workflows focal (ubuntu 20.04) -> jammy (ubuntu 22.04) #154153 (@atalman)
• Move all CI/CD workflows from focal to jammy #154157 (@atalman)
• [Inductor] Set Triton Allocator Function For Use with New TMA API in Inductor #155373 (@NikhilAPatel)]
• [Upstream Triton] Support new host-side TMA API in user-defined triton kernels #155574 (@davidberard98)
• [cherry-picked into 3.4 branch, awaiting pytorch pin update] [Upstream Triton] persistent mm + tma accuracy failures #156028 (@davidberard98)
• [inductor][triton pin] Support triton builtins after #7054 #156031 (@davidberard98)
• [NVIDIA] L2 cache hints only for sm >= 80 triton-lang/triton#7219 (@davidberard98)
• [cherry-picked into 3.4 branch, awaiting pytorch pin update] several transformers tests fail with torch 2.8 RC but pass with torch 2.7.1 on T4 (but both pass on A10) #157276
• [PT2.8 potential blocker] - HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION in unfold kernel ROCm/triton#829 (@scxiao)
• [PT2.9 potential blocker] [Triton 3.5] - Memory access fault int64 indexing ROCm/triton#830 (@makslevental)
• [revert: [Revert] [Coalesce] Fix the default order to be row major (#5707) #7143 triton-lang/triton#7380 is in 3.4.x branch, awaiting pytorch pin update] [triton 3.5] Failing Pytorch UT: Numerical issues in test_comprehensive___rpow___cuda_float32 ROCm/triton#832

General issues

• 8 [Upstream Triton] AssertionError: Scalars are not equal! inductor/test_codecache.py #154250 (@davidberard98)

AMD specific

• 5 [Upstream Triton] [ROCm] Accuracy issues in inductor/test_torchinductor_opinfo #155803
• 3 [Upstream Triton] [ROCm] AssertionError: Tensor-likes are not close! test_upsample_nearest1d_dynamic_shapes_cuda #154215
• 1 [Upstream Triton] AssertionError: False is not true test_pad_3d_tensor #154224 (can't repro on h100)

Follow-ups

• [user triton][dynamo Fix TMA Descriptor reconstruction #155856
• [Upstream Triton] Handle user-specified triton.set_allocator function #155584 (not urgent for release)

cc @chauhang @penguinwu @bertmaher @int3 @davidberard98 @nmacchioni @chenyang78 @embg @peterbell10 @aakhundov

[iupaikov-amd]

Would really appreciate if someone could take a look at why cuda builds fail for testing PR. There're some folder structure changes in triton so install_triton script fails for some reason for cuda, but not for amd. I would like to get a full CI run and do a cross-comparison.

[malfet]

@iupaikov-amd can you please create a PR that updates the pin and see what will fail rather than creating numerous issues, because it feels like some of them might be a local setup problems

[desertfire]

@davidberard98

[davidberard98]

re: cuda build failures, I'm putting up #154635. I'm not sure right now whether this fixes all the issues or just some of them.

[davidberard98]

@anmyachev's #154905 fixes some recent _unwrap_if_constexpr failures

[iupaikov-amd]

Updated triton commit hash to 65648400d91f074770224ea38b732d7cba934f12 to try and fix cuda build issues in testing PR.