-
Replace with cleaned vector-only implementation
- Remove obsolete cube/ duplicate (cube was identical to vector after Cube matmul fallback)
- Remove dead matmul tiling code and TCubeTiling fields
- Remove OPS_LOG_I diagnostic logs
- Add support for arbitrary headDim via hdPad + DataCopyPad tail
- Add fp32 scalar QK dot product and PV accumulation
- Dynamic blockK reduction (64->4) for tight UB budgets
- Updated README and tests
-
v2.0: Full float32 precision pipeline, headDim=64 support, 448/448 production tests passed
-
Delete tests/test_aclnn.cpp with huggingface_hub
-
Upload folder using huggingface_hub
