10 Commits (main)

Author SHA1 Message Date
BlinkDL b4925900e7 misc 3 years ago
BlinkDL dc26998708 torch jit 3 years ago
BlinkDL 23f64aeebc misc improvements 3 years ago
BlinkDL 2567c8c904 rescale to avoid FP16 overflow 3 years ago
BlinkDL aef9f6f7ef fp16 3 years ago
BlinkDL 8a4a41a3aa better 3 years ago
BlinkDL daed379db2 bf16 inference - 15G VRAM for 7b model 3 years ago
BlinkDL c7155525bb better 3 years ago
BlinkDL fc3bc1eb0e good for UTF-8 inference (such as CJK) 3 years ago
BlinkDL 3a7e6a6aa3 faster inference 3 years ago