This is a patch release containing the following changes to v3.14.1:
- Added support for
f8destination in layer normalization forward propagation on processors with Intel AVX-512 instruction set support (f1b8598, 6cfd004, c096cad) - Fixed runtime errors and crashes in matmul and convolution with
f8_e4m3/f8_e5m2binary post-ops on processors with Intel AVX-512 instruction set support (d16b6bf, a19e9d7, 63045f7, 847802f, 87dc209, f4bb731, 9a1412a) - Removed benchdnn test cases with
mishpost-op and dynamic scales due to false positive correctness issues (1d6723b) - Fixed correctness issue in SDPA subgraph with query length
16on Intel GPUs (4db2846)