Lorenzo Liusso comments

Repositories
Issues
Comments

Results 3 comments of


                                            Lorenzo Liusso

autocast bfloat16 to float32 for llama on clang

It's not an LLVM issue; fix_bf16() was not casting correctly. However, I believe this is just a workaround, and a more robust solution should be proposed.

autocast bfloat16 to float32 for llama on clang

Do we want to keep `fix_bf16`?

clang backend fails using version 18.1.3

seed = 1728209481 opened device CLANG from pid:31525 opened device PYTHON from pid:31525 assign