Reported in #18.

Fixes #17.

- Add upstream patch for python 3.13 API change https://github.com/tensorflow/tensorflow/commit/846468dd25153e9a41f0a9da1ba1604ce819628d
- Add additional patch for python 3.13 API changes:
  - _PyArg_NoKeywords related upstream commit https://github.com/python/cpython/commit/be5e8a010341c4d2d28ef53a1baed402ee06466e.
  - _PyObject_VisitManagedDict/_PyObject_ClearManagedDict related upstream commit https://github.com/python/cpython/commit/fc2cb86d210555d509debaeefd370d5331cd9d93.
- Add patch adding hermetic PYTHON 3.13 requirements lock file based on upstream commit https://github.com/tensorflow/tensorflow/pull/82768/commits/d5c90f7118ad5a8d3a9e80cd82ced94833e0ea38.
- Fixes #16.

Still built with python-3.12.7-1 and python-installer-0.7.0-8 because of
missing Python 3.13 support.

- `-Wno-gnu-offsetof-extensions` becomes `-Wno-c23-extensions`.
- Add patch to have bazel apply patch to bundled cutlas. The patch being applied is https://github.com/NVIDIA/cutlass/commit/82f5075946e2569589439d500733b700a3141374.

- Add upstream patches supporting hermetic CUDA 12.8.0 and CUDNN 9.7.0.
  - PR #18248: cuda 12.6.2
  - https://github.com/tensorflow/tensorflow/commit/8ca77eb4bee16d6efe9a2cfda12ede448eb7bc72.patch
  - [PATCH] Add support for CUDA 12.6.3 and CUDNN 9.5.1/9.6.0.
  - https://github.com/tensorflow/tensorflow/commit/521421ea776003c21b9db6caeec1af7258f7e52e.patch
  - PR #20948: [GPU] Fix NCCL with CUDA 12.6.3.
  - https://github.com/tensorflow/tensorflow/commit/37c02230d750b1b593c498a2aaaec59b1875585c.patch
  - Add support for CUDA 12.8.0 and CUDNN 9.7.0.
  - https://github.com/tensorflow/tensorflow/commit/7ede78237dfbc5ddef63b01ddda6f6bad2458710.patch
  - Add NCCL dict entry for CUDA 12.8.0.
  - https://github.com/tensorflow/tensorflow/commit/8a0fbdaebc6209f6b90a8d9cfc83055253d69d1e.patch
- Add custom patch for build failure based on https://github.com/tensorflow/tensorflow/commit/5467ee993e1d3e4709c1e99f3a15a978325ae536.
- Update `HERMETIC_CUDA_COMPUTE_CAPABILITIES` compute capabilities.
  - Add  `sm_90a`, `sm_100`, `sm_100a`, `sm_101`, `sm_101a`, `sm_120` and `sm_120a`.
  - Replace `compute_90` with `compute120`.

Warn users about broken package with python 3.13

Both packages were moved into python-keras and tensorflow.
https://github.com/keras-team/keras-applications/commit/06fbeb0f16e1304f239b2296578d1c50b15a983a
https://github.com/keras-team/keras-preprocessing/commit/3e380065d4afc7347aaee8d89325a16b22158438
Related python-keras!1.

Bazel hermetic python build only uses python itself. python-installer is needed to install the generated wheel in $pkgdir.
Implements #15.

Prevents build failure with: TF wheel shouldn't be built with CUDA dependencies. Please provide `--config=cuda_wheel` for bazel build command. If you absolutely need to add CUDA dependencies, provide `--@local_config_cuda//cuda:override_include_cuda_libs=true`.

Rename TF_CUDA_VERSION to HERMETIC_CUDA_VERSION and converted to required triplet form.
Rename TF_CUDNN_VERSION to HERMETIC_CUDNN_VERSION and converted to required triplet form.
Rename TF_CUDA_COMPUTE_CAPABILITIES to HERMETIC_CUDA_COMPUTE_CAPABILITIES.

Fixes #11.

Fixes #13 and #12.

loqs authored 10 months ago and Sven-Hendrik Haase committed 10 months ago

Fixes #7