nvidia

/

NVIDIA-Nemotron-Nano-9B-v2

like 419

Follow

NVIDIA 41.5k

Description

This PR adds support to modeling_nemotron.py for running inference on CPU. This is a cleaned up version of the edits I made while working on support in llama.cpp.

Changes

• Handle failed imports of rmsnorm_fn
• Add un-optimized implementation of MambaRMSNormGated.forward
• Fix NemotronHMamba2Mixer.torch_forward to use repeat_interleaved for B and C (see discussion here)