FineGrainedFP8Config(dequantize=true) materializes full BF16 model per rank before sharding

May 1, 2026 · #2114
View on GitHub
Python Difficulty: Medium

Parent Repository

NVIDIA-NeMo/Automodel

Python repository

503 150
All Issues Back to Automodel

Sign in required

Authenticate to use favourites & bookmarks

5