Added
- Add Jupyter notebook tutorial for using amd-flashinfer on ROCm (#213) @diptorupd
- Add ROCm profiler module (flashinfer/rocm_profiler/) and FA2 single-prefill benchmark driver using rocprofv3 (#205) @diptorupd
- Gate torch.compile integration behind FLASHINFER_USE_TORCH_CUSTOM_OPS, with HIP pytest coverage (#210) @demandal25
Changed
- FA2 prefill on HIP: improve occupancy and throughput on CDNA3 via LDS-aware CTA_TILE_Q selection and shared-memory budget capping (#209) @diptorupd
- Add k128B_16Row swizzle mode for prefill shared memory to reduce LDS bank conflicts on CDNA3 (#207) @diptorupd
Fixed
- HIP FA2 on CDNA3: correct bfloat16 row-sum MFMA intrinsic selection and MFMA C/D row indexing for custom masks with GQA (#214) @subhajitdchow
- Editable installs: ensure JIT sees current headers under include/ with scikit-build-core redirect mode (symlink at install + get_include_paths fix) (#208) @diptorupd
Maintenance
- Refresh README @demandal25
Contributors: @diptorupd, @demandal25, @subhajitdchow
Summary: This point release focuses on ROCm prefill correctness and performance, developer tooling, and install ergonomics. It fixes silent bf16 softmax scaling and custom-mask/GQA mis-indexing in HIP FA2, improves CDNA3 occupancy and LDS bank behavior for FA2 prefill, adds an opt-in torch.compile path, a rocprofv3-based profiling kit, and a notebook tutorial. Editable installs now pick up header edits without reinstalling.
Contributors
diptorupd, subhajitdchow, and demandal25