Repository navigation
Potential float mismatch between training inputs and model params #69
Description
Activity
- linked a pull request that will close this issueAlign `fit_vae` input casting with model parameter dtype #74
on Sep 15, 2026
{{ message }}
Possible dtype mismatch. Model parameters are bfloat16 but training inputs are float32.
I believe that the .float() linked below casts to float32 and breaks anytime model params were created with --bfloat16.
embedding-kit/src/embkit/optimize/__init__.py
Line 322 in c41ffd5