Reapply "[FRONTEND] added support for tuples (#5220)"

[anmyachev]

This reverts commit 492ea92.

Summary of changes:

• support for tuples
• constexprs are now also part of the signature. The format is: signature={..., 'o': "*fp32", '[Name of constexpr]': 'constexpr'}
• For ASTSource there is parameter name change: constants -> constexprs
• New nesting level has appeared for the main properties (defined in _add_common_properties): divisibility_16, equal_to_1 and equal_to_none.

Inductor CI:

• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12415670010 (failed)
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12420597752 (failed; new pytorch build did not happen)
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12428576515 (not enough space for building)
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/12435110524

[anmyachev]

On Triton's side I think I fixed everything.

For Torch Inductor, even considering that I made several changes for it, this is still not enough and the full set of inductor tests does not pass. Considering the number of changes, I believe it is good idea to split it into another PR.

[alexbaden]

For Torch Inductor, even considering that I made several changes for it, this is still not enough and the full set of inductor tests does not pass. Considering the number of changes, I believe it is good idea to split it into another PR.

I agree. I think we should remove the PyTorch patch from this PR and keep it in a separate branch since the inductor tests will fail either way.

[whitneywhtsang]

According to the discussion, are we removing all debug prints and PyTorch patches in this PR?

[anmyachev]

@alexbaden I can't completely remove the patch, because even the main tests will stop working. We have at least one test in the main set that uses PyTorch inductor:

intel-xpu-backend-for-triton/scripts/test-triton.sh

Line 341 in b9afb6c

ZE_AFFINITY_MASK=0 python pytorch/benchmarks/dynamo/huggingface.py --accuracy --float32 -dxpu -n10 --no-skip --dashboard --inference --freezing --total-partitions 1 --partition-id 0 --only AlbertForMaskedLM --backend=inductor --timeout=4800 --output=$(pwd -P)/inductor_log.csv

@whitneywhtsang I removed debug things from the Triton code, but I intend to leave them in the patch, because they will still be useful for fixing the inductor tests (at least they highlight places that are affected by the change in the Triton interface).

UPD: CI is not working now because there is no more space on the runners: /runner/_work/intel-xpu-backend-for-triton/intel-xpu-backend-for-triton/pytorch/third_party/ideep/mkl-dnn/src/gpu/intel/jit/gemm/gen_gemm_kernel_generator.cpp:27738:1: fatal error: error writing to /tmp/ccXBqNWT.s: No space left on device. cc @kwasd @pbchekin

[anmyachev]

@pbchekin @kwasd please ping me when it makes sense to try to restart CI

[pbchekin]

@pbchekin @kwasd please ping me when it makes sense to try to restart CI

Please restart.

[anmyachev]

@pbchekin @kwasd please ping me when it makes sense to try to restart CI

Please restart.

Unfortunately it doesn't work

[pbchekin]

Unfortunately it doesn't work

Ops, sorry for that. Let me fix the problem, merge #3053, then I restart the workflow.

[anmyachev]

Test with pip failed in the Inductor test suite, and this happened because pytorch is not rebuilt with the new patch in this workflow. So I'm going to merge this pull request, assuming that as soon as new PyTorch wheels become available for this flow, it should go through successfully.