Candle vs. PyTorch performance

I'm running https://github.com/huggingface/candle/tree/main/candle-examples/examples/llava vs. https://github.com/fpgaminer/joycaption/blob/main/scripts/batch-caption.py on a Mac m1.

Seeing significant performance difference, Candle seems much slower.
I enabled accelerate and metal features.

Would love some pointers how to improve it.

Start with candle-examples/examples/llava and compare its invocation and enabled features with scripts/batch-caption.py from joycaption. Reproduce the Mac M1 performance comparison, then profile or benchmark both paths to identify the cause of the difference. Done means a confirmed bottleneck and an actionable fix or documented limitation.

Written by the indexing model from the issue text.