Add tiny transformer LLM notebook

[cetagostini]

Summary

• Add a tiny decoder-only transformer LLM gallery notebook using PyTensor and xtensor named dimensions.
• Lower matmul-shaped xtensor dot contractions to matmul and keep outer products on the existing einsum fallback.
• Prefer static shape constants in tensordot reshape shapes and add focused regression coverage.

Test plan

• conda run -n pytensor-dev python -m ruff check pytensor/tensor/math.py pytensor/xtensor/rewriting/math.py tests/tensor/test_math.py tests/xtensor/test_math.py
• `conda run -n pytensor-dev python -m pytest tests/xtensor/test_math.py::test_dot
• tests/xtensor/test_math.py::test_dot_lowers_to_matmul
• tests/xtensor/test_math.py::test_dot_outer_product_falls_back_to_einsum
• tests/xtensor/test_math.py::test_dot_errors tests/xtensor/test_math.py::test_dot_vectorize
• tests/tensor/test_math.py::TestTensordot -q`

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[cetagostini]

Doing the notebook found a few things which were adding an overhead in xtensor, adjust and end-up being even faster now xtensor than plain tensor.

[ricardoV94]

This is nice, I don't want the random xtensor changes, we need to investigate why it was not simplifying in your case, may be another symptom of #2056 or something else, but shouldn't be done in a docs PR

[cetagostini]

@ricardoV94 follow some of your comments, and came up with this: #2164

[twiecki]

That's cool!

Reply via ReviewNB