| Filename | Latest commit message | Latest commit date |
|---|---|---|
2a30dcd360
stablehlo::limb_module: the tensor-engine member's limb layout as one StableHLO dot_general of bf16 operands, accumulated and returned in f32, the program a compiler maps to a matrix unit. stablehlo_gate writes it with the batch's limb weights and limbs and the exact coefficients, and a nine-bit-limb data set whose exact coefficients bf16 cannot reproduce (a run that matches them is not comparing). check.sh holds the limb module to "=!" (exact on eight-bit limbs, different on nine); both drivers read typed raw files, and floats are compared exactly (IREE's --expected_f32_threshold=0). Measured: exact under both compilers on the x86-64 and arm64 CPUs and on the discrete GPU through CUDA; the nine-bit limbs differ everywhere. Not exact: the IREE 3.12 HIP build for gfx1151, whose bf16 multiply returns non-integers (-14.0002 for -14), so the workflow shows that step and does not require it. Not built: the IREE Vulkan and Metal builds do not lower bf16 in this version. Co-Authored-By: Claude Opus 5.5 <[email protected]> |
||
| .. | ||
| check.sh | ||
| xla_run.py | ||