Extend libtorch-stable ABI compat for torch_musa 2.9 (vLLM v0.24.0 kernels)

[yeahdongcn]

Summary

torch_musa 2.9.0's torch::stable runtime predates the torch 2.10 stable-ABI
surface that vLLM v0.24.0's csrc/libtorch_stable/ kernels (and the equivalent
SGLang kernels) are written against. This extends torchada's libtorch-stable
compat layer so those kernels build and register on MUSA without patching the
kernels themselves.

What's added

• torch::stable::Device / DeviceType shim (device.h) with the ctors and
  == / is_cpu / is_cuda / is_privateuseone predicates the kernels use.
• torch::stable::Tensor accessors backported into tensor_struct.h at
  build time — sizes / strides / device / element_size /
  storage_offset, and the templated mutable_data_ptr<T> / const_data_ptr<T>
  — gated on TORCHADA_STABLE_ACCESSORS so a TU that does not force-include the
  box header is untouched.
• TORCH_BOX boxer (torchada_stable_box.h) synthesized over torch_musa's
  from / to conversions, including spreading std::tuple / std::pair
  returns across stack slots.
• torch::stable free functions torch_musa 2.9 lacks: contiguous,
  flatten, empty, from_blob.
• THO_DISPATCH_* shim (torch/headeronly/core/Dispatch.h) and a
  torch/csrc/stable/macros.h forward.
• CUDA→MUSA header shims for stable kernels that include them directly:
  cuda_runtime.h, cuda_bf16.h (+ __nv_bfloat16{,2} alias), cuda_fp16.h,
  cuda_fp8.h, and cublas_v2.h → mublas.h.
• Build wiring: include_paths() auto-appends the compat include dir on MUSA,
  and the header backport runs lazily at extension-build time (never at import,
  so a bare import torchada / inference run never mutates the torch headers).

Testing

Added TestStableCompatHeaders + TestStableHeaderBackport to
tests/test_cuda_patching.py, covering the compat headers, the TORCH_BOX
shim, and the accessor / inline-rewrite backport transforms (path helpers,
header content, cuda_* → musa_* redirects, injection injected / already
/ anchor-missing statuses + idempotency, inline-rewrite skip rules).

Full suite run in a MUSA container (MTT S5000, torch_musa 2.9.0 / musa 40305):

PYTHONPATH=src python -m pytest tests/test_cuda_patching.py
=> 201 passed, 4 skipped, 0 failed

Notes

• Targets torch_musa 2.9's ABI specifically. The torch 2.11 from_blob
  deleter overload and the fully-defaulted / layout-enum empty forms are
  intentionally not shimmed — torch_musa 2.9's AOTI C-shim has no
  deleter-capable create_tensor_from_blob, so a dropped deleter would leak the
  backing storage; kernels needing those forms stay out of the MUSA kernel set.
• The header backport surfaces a warning if the tensor_struct.h anchor is
  absent, instead of silently skipping and failing later in the compiler.

[augmentcode]

🤖 Augment PR Summary

Summary: This PR extends torchada’s libtorch-stable compatibility layer so vLLM v0.24.0 / SGLang stable-ABI kernels (written against newer torch::stable surfaces) can build and register on torch_musa 2.9 without patching kernel sources.

Changes:

• Added stable-compat header shims that redirect CUDA-only includes to MUSA equivalents (e.g. cuda_runtime.h, cuda_fp16.h, cuda_bf16.h, cublas_v2.h).
• Introduced torch::stable::Device/DeviceType shim and a torch/csrc/stable/macros.h forwarder for older torch_musa layouts.
• Shipped a THO_DISPATCH_* shim (torch/headeronly/core/Dispatch.h) and avoided namespace collisions with ATen’s ::detail.
• Expanded the force-include stable box header to backport missing stable free functions (contiguous, flatten, empty, from_blob) and add MUSA runtime/kernel-launch check macros used by upstream kernels.
• Enhanced the TORCH_BOX shim to support tuple/pair multi-value returns by spreading elements across StableIValue stack slots.
• Updated extension build wiring to append the stable_compat include dir on MUSA and to lazily backport missing torch::stable::Tensor accessors / inline definitions at extension-build time.
• Added unit tests validating header presence, path helpers, and idempotent/string-transform backport behavior.

Technical Notes: Header backporting is best-effort and gated so normal import torchada / inference runs do not mutate torch headers; the compat headers are appended last in include paths so future real torch_musa headers can take precedence.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 2 suggestions posted.

Comment augment review to trigger a new review at any time.