fix(cpp_extension): accept torch 2.6+ include_paths/library_paths call signatures

[yeahdongcn]

What / Why

PyTorch 2.10 added torch_include_dirs to torch.utils.cpp_extension.include_paths, and torch_include_dirs and cross_target_platform to library_paths. Its Inductor C++ builder calls include_paths(device_type, torch_include_dirs) positionally and library_paths(device_type, torch_include_dirs=..., cross_target_platform=...).

On MUSA, torchada replaces both helpers with versions that take (cuda, device_type). The positional call therefore hands device_type=False to include_paths, which fails with 'bool' object has no attribute 'lower', and library_paths rejects the two keywords. Every CPU kernel that Inductor compiles with a cold cache goes through these calls, so after import torchada a plain CPU torch.compile with an empty TORCHINDUCTOR_CACHE_DIR fails.

vLLM-Omni's MAGI-2 test suites on MUSA hit this: their CPU torch.compile tests run with a fresh Inductor cache and failed in this call. With this change installed they no longer fail there.

Change

• include_paths(device_type=None, torch_include_dirs=True, *, cuda=None) and library_paths(device_type=None, torch_include_dirs=True, cross_target_platform=None, *, cuda=None) follow PyTorch's signature, positionally and by keyword.
• A bool in the device_type position is the PyTorch < 2.6 positional cuda argument, which torch 2.5's builder passes. cuda= stays as a keyword.
• On MUSA, torch_include_dirs=False leaves out paths under PyTorch's own include and lib directories. torch_musa's directories and torchada's stable-ABI compat headers stay.
• Off MUSA, torch_include_dirs and cross_target_platform are passed through when the installed torch accepts them.
• On torch 2.6-2.9 the builder passes only device_type, positionally. With the old signature that string landed in cuda and counted as true, so CPU builds also received the MUSA include and library paths from that device-options call. That call now returns the CPU set (no MUSA library directories).
• cuda is now keyword-only, so torchada's own two-positional form include_paths(cuda, device_type) is no longer accepted. Nothing in this repository used it.

Verification

New tests in tests/test_cuda_patching.py::TestCppExtensionPaths (MUSA only):

• test_paths_accept_the_positional_torch_signature: the positional torch 2.10+ calls return the same paths as the keyword calls, and a positional bool matches cuda=;
• test_paths_without_torch_dirs: with torch_include_dirs=False no returned path lies under the torch package, and the result is a subset of the default;
• test_inductor_cpu_compile_with_a_cold_cache: a CPU torch.compile(fullgraph=True) in a fresh interpreter with an empty TORCHINDUCTOR_CACHE_DIR builds and returns the expected values.

Results on MTT S5000 (torch 2.11.0.post2, torch_musa 2.11.0.post2+musa5.2.0):

• A cold-cache CPU torch.compile fails with stock torchada and passes with this change.
• pytest tests/, with this change applied on top of fix(musa): keep torch.isfinite asynchronous on MUSA floating tensors #120 (the torchada tree used for the vLLM-Omni runs): 634 passed, 19 skipped, 2 failed. The new tests pass. The two failures are TestInductorTemplateHeuristics::test_copies_only_cuda_triton_heuristics_and_clears_cache and ::test_is_idempotent_and_preserves_cache_without_changes (KeyError: ('triton::bmm', 'musa', None) and the same for triton::mm), and they fail the same way on the unmodified source in the same image.

Not covered / notes

• Only torch 2.11 was run. The torch 2.6-2.9 CPU path change above was not run on a torch_musa 2.9 stack.
• library_paths(..., torch_include_dirs=False) is only reached by libtorch-free AOTInductor builds (aot_inductor.link_libtorch=False), which were not exercised on MUSA beyond the unit test.
• On MUSA, library_paths still returns [] for a CPU build, as before. Inductor adds TORCH_LIB_PATH itself.
• The version is not bumped here, to avoid conflicting with fix(musa): keep torch.isfinite asynchronous on MUSA floating tensors #120; it would be 0.1.92 if this lands after fix(musa): keep torch.isfinite asynchronous on MUSA floating tensors #120.

AI assistance: Claude Code drafted the change, the tests and this description and ran the MUSA validation listed above.