What's Changed
- Fix SDPA fused backward for non-positive sm_scale. by @anhminhnguyenhoang in #245
- Widen the gfx1100 attn_fwd tuning space to PADDED_HEAD=True too by @sstamenk in #250
- Numerical Error Lessons from the Triton Attention Kernels, and a Test that Catches Them by @xinyazhang in #249
- PostgreSQL based Generic Distributed Execution Framework (pgDEF) by @xinyazhang in #252
- Add gfx117X target support by @mstankov-amd in #241
- Pin gfx117x Triton compiler with altwheel config by @arsic3 in #254
- Add a utility tool aotriton.unaks2 by @xinyazhang in #256
- Auto-detect ROCM_PATH and CMAKE_PREFIX_PATH via rocm-sdk by @xinyazhang in #258
- Human-Readable Tuning Database by @xinyazhang in #257
- Include in generated op/shim/flyc sources by @k-artem in #262
- Enable gfx1250/gfx1250-strict with tuned database. by @xinyazhang in #264
- In gfx1201's FlyDSL kernel, replace bf16_trunc_pack_v8 with bf16_rne_pack_v8 by @xinyazhang in #266
- Fix tuning db for gfx950 and gfx1100 by @xinyazhang in #267
- Add VarlenBits::principal_diagonal_choice to anchor sliding windows at the bottom-right diagonal by @xinyazhang in #268
- Perfmon: Cross-Version Performance Monitor and Human-Readable Performance Database by @xinyazhang in #261
- End of 0.15b: Tuning at GPU-Filling Sizes and the gfx1100 Tuning Database Rebuild by @xinyazhang in #270
New Contributors
- @anhminhnguyenhoang made their first contribution in #245
- @mstankov-amd made their first contribution in #241
- @arsic3 made their first contribution in #254
- @k-artem made their first contribution in #262
Full Changelog: 0.14b...0.15b