Lightning-AI/lightning-thunder

quantization: process tensors on meta device directly, maybe implement CPU quantization (if it is easy)

Open

#1111 aperta il 6 set 2024

Vedi su GitHub
 (4 commenti) (0 reazioni) (0 assegnatari)Python (114 fork)github user discovery
good first issuetransforms

Metriche repository

Star
 (1460 star)
Metriche merge PR
 (Metriche PR in attesa)

Descrizione

Currently the BitsAndBytesLinearQuant4bit for submodule always calls bitsandbytes.functional.quantize_4bit. This is somewhat touchy for CPU tensors because quantize_4bit only works on GPU tensors but it is outright not so nice for meta tensors, where we only would need to get the right shapes.

https://github.com/Lightning-AI/lightning-thunder/blob/e64d347def39bb47101efafe4177adf9f77a63ec/thunder/transforms/quantization.py#L93-L103

Guida contributor