Upgrade Modal runtime to CUDA 13.3 with nvmath and cuBLASDx

[msaroufim]

Summary

• upgrade the shared Modal base image from CUDA 12.9.1 to CUDA 13.3.0
• add cuda-tile==1.4.0, nvmath-python[cu13-dx]==0.9.0, and pinned cuBLASDx/MathDx packages and headers
• align the CUTLASS C++ headers with CUTLASS DSL 4.5.2
• remove the incompatible cuNumeric CUDA 12 dependency stack
• add image-build checks for Python dependencies, cuTile's assembler, and a cuBLASDx C++ include compile

Why

Kernelbot jobs need cuTile, nvmath-python, and cuBLASDx in the shared Modal runtime. The old CUDA 12.9/cuNumeric dependency set cannot provide a clean supported CUDA 13 nvmath stack. CUDA 13.3 also supplies tileiras, avoiding the cuda-tile[tileiras] toolkit constraint that conflicts with the CUDA 13.0 dependency set used by the PyTorch 2.12 wheel.

Impact

New Modal jobs can import and execute cuTile and nvmath APIs and compile submissions against cuBLASDx headers. The Modal host driver remains provider-managed; the current R580.95.05 driver passed the runtime checks.

Validation

• uv run --extra dev ruff check src/runners/modal_runner.py
• uv run --extra dev python -m py_compile src/runners/modal_runner.py
• git diff --check
• rebased Modal pytest deployment: 2 T4 integration tests passed
• H100 smoke: cuTile kernel, nvmath host/device matmul, and cuBLASDx SM90 compile passed
• production Modal main deployment completed
• production A100 test submission 860004 passed and reported torch 2.12.0+cu130

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
src/libkernelbot
utils.py
Project Total

This report was generated by python-coverage-comment-action