v5 Transformers

#34
by AntonV HF Staff - opened
AntonV changed pull request status to open
Nomic AI org

Just tested this, and I'm getting identical performance on NanoMSMARCO using Sentence Transformers using:

  • trust_remote_code=True on main
  • trust_remote_code=True on this PR
  • No trust_remote_code on this PR, i.e. with native transformers

My only comment is that the last of the 3 gets 2 warnings:

Unrecognized keys in `rope_parameters` for 'rope_type'='dynamic': {'rope_theta'}
model.safetensors: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 547M/547M [00:05<00:00, 94.0MB/s]
Loading weights: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 112/112 [00:00<00:00, 10455.19it/s]
Unrecognized keys in `rope_parameters` for 'rope_type'='dynamic': {'rope_theta'}
0.6025776932803806

But I assume this is possible for the frankenconfig that supports both old and new. Nice work @AntonV

  • Tom Aarsen

Before I forget it, the warning should be resolved now btw. Was not due to the config but something missing in our codebase

hnomic changed pull request status to merged

Sign up or log in to comment