torch_musa Release v2.9.1

Release Note

Hi all, torch_musa v2.9.1 is now available. Starting with this release, torch_musa officially integrates with Release SDK 5.1.0 components, providing full CUDA-aligned operator coverage across tensor types, including dense, quantized, sparse, sparsecsr, and nested tensors. This release also fixes several issues.

We also made mutlass as a third-party repository of torch_musa for implementing high-performance matmul kernels.

Additionally, the -march=native option has been removed from MUSAExtension to improve cross-platform compatibility.

Build torch_musa v2.9.1 on MUSA platform with MUSA SDK>= 5.1.0 please.

Enhancements

• Fixed ternary with ambiguous out shape, resize instead
• Random distribution kernels support CUDA hardware configurations for numerical consistency
• Fixed lazy conjugation handling in bmm
• Removed tensor's dim check
• Fixed int8(-1) == 255 behaves inconsistently with CUDA
• Support Flash SDPA with dropout & headdim>=256

New Features

• Migrated MUDNN invocation to C APIs
• Aligned torch CUDA native allocator and accelerator APIs
• Comm operators in synchronous mode run on the default stream and return nullptr (the original implementation returned a work handle)
• ProcessGroup supports context mode. All communication operations within the context are executed in the specified process group by default
• A new dynamic configuration API for ProcessGroupMCCL timeout is added. The timeout duration can be adjusted dynamically via backend._set_default_timeout
• Non-blocking APIs are enabled during MCCL initialization, which completely resolves permanent program hangs caused by communication issues
• The new lazy-hsdp-allreduce feature is added to improve communication efficiency

Known && blocked issues

• _cslt_compress, _cslt_sparse_mm, and _cslt_sparse_mm_search depend on MUSPARSELT. torch_musa has completed the integration, but these operators still fail at runtime

Please feel free to contact us with any issues or questions.