AaravGaur

3 ms·

1. 1.▲
   DeltaTensors: Git for model fine-tunes that also saves you storage

   1 points by AaravGaur 23d ago|0 comments
2. 2.▲
   Deltatensors – store model fine-tunes as compressed weight deltas(github.com)

   4 points by AaravGaur 4mo ago|1 comments
3. 3.▲
   by AaravGaur 4mo ago

   Near-lossless delta compression for fine-tuned neural network models. Compressed a Qwen-2.5-0.5b wikitext fine-tune by 3.2x. Instead of storing 50 fine-tunes of the same base model, store one base and 50 small .wdelta delta files. deltatens