huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 02:38:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Better handling of dtypes in llama. (#243)

Browse Source
This commit is contained in:
Laurent Mazare
2023-07-26 08:28:33 +01:00
committed by GitHub
parent 471855e2ee
commit e40b150bbe
2 changed files with 12 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
candle-examples/examples/llama/main.rs
View File

@ -128,8 +128,8 @@ fn main() -> Result<()> {
let device = candle_examples::device(args.cpu)?;
let config = Config::config_7b(args.use_flash_attn);
let cache = model::Cache::new(!args.no_kv_cache, &config, &device)?;
let dtype = if args.use_f32 { DType::F32 } else { DType::F16 };
let cache = model::Cache::new(!args.no_kv_cache, dtype, &config, &device)?;
let (llama, tokenizer_filename) = match args.npy {
Some(filename) => {
let vb = VarBuilder::from_npz(filename, dtype, &device)?;