mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-19 23:50:51 +00:00
- rustfmt over the tree. - clippy 1.98: `is_multiple_of` for six modular checks, and `as_chunks` for the bf16/f32 row decode. - cpu_flash::packed: the first chunk against its own pass, and the cache path against the block pass, are different kernels. On x86 they agree to about 1e-7, not bit for bit, so the check is within 1e-6. A key read past the chunk would be off by the size of a score. - geglu: the fused erf is within about 3e-7 of the composite's (measured worst 2.8e-7 on an AMD EPYC 7571, near g = -0.06), so the bound is (½|g||u| + |y|)·4e-7. hanzo-ml and hanzo-nn: 409 tests pass on x86. The ROCm tests (rocm_moe 6, rocm_gdn 2, rocm_ops 1) pass on gfx1151. Co-authored-by: Hanzo Dev <[email protected]>