foundation-model-stack / fms-hf-tuning

🚀 Collection of tuning recipes with HuggingFace SFTTrainer and PyTorch FSDP.
Apache License 2.0
28 stars 48 forks source link

fix: do not resize embedding layer by default #310

Closed kmehant closed 3 months ago

kmehant commented 3 months ago

Defaults multiple to 1 thereby not resizing the embedding unless user wishes to resize it by explicitly using this feature.

Motivation is to support vLLM inference for LoRA which needs merging of adapter when embedding layer is resized with no real tokens. Findings from @Ssukriti and team.