Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Override cast_to_fp8 in te.module.linear (#120)
**Description** Fix a bug in TE integration. Currently we only override te.cpp_extensions.cast_to_fp8 with our own cat_to_fp8 in msamp.te.extension. We also need to override te.module.linear.cast_to_fp8. Otherwise, it will use the original function which does not support ScalingTensor and will raise an exception in Megatron-LM **Major Revision** - Override cast_to_fp8 in te.module.linear
- Loading branch information