vllm v0.26.0 compat

[Joey-gvwal]

Summary

vLLM v0.26.0 (via vllm-musa) touches two torch surfaces that torch_musa 2.9
does not provide as-is:

1. torch.accelerator.get_memory_info() — a torch 2.10+ API, absent on
   torch_musa 2.9 (engine-core startup hits AttributeError).
2. The kernels vLLM relocated into csrc/libtorch_stable/ use upstream CUDA
   spellings (cudaDeviceProp, cudaStream_t, cublasHandle_t,
   aoti_torch_get_current_cuda_stream, …), but torch_musa's stable ABI only
   exposes the musa* names.

These shims let both build/run unchanged on MUSA, keeping the CUDA→MUSA mapping
in torchada instead of downstream forks.

Changes

• _patch.py: map torch.accelerator.get_memory_info →
  torch.musa.mem_get_info (same (free, total) contract).
• csrc/stable_compat/: add #ifndef-guarded CUDA→MUSA aliases so
  libtorch-stable kernels compile against upstream names —
  • cuda_runtime.h: cudaDeviceProp, cudaError_t, cudaSuccess,
    cudaStream_t, cudaGet{DeviceCount,DeviceProperties,Device,ErrorString}
  • cublas_v2.h: cublasHandle_t → mublasHandle_t
  • torchada_stable_box.h: forward aoti_torch_get_current_cuda_stream →
    aoti_torch_get_current_musa_stream; define TORCH_UTILS_CHECK
    Guards make each alias a no-op where an mcc porting layer already provides
    the name.

Testing

Rebuilt vllm-musa's stable extension (_C_stable_libtorch) against this
torchada on MTT S5000 (torch_musa 2.9.1). All 4 extensions import; smokes pass:
FP8 quant (Qwen3-8B/32B-FP8), MLA (DeepSeek-V2-Lite-FP8, FLASHMLA), and
multi-GPU stable custom_all_reduce (Qwen3-32B-FP8, tp2).