huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 19:58:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add the quantized mpt model. (#1123)

Browse Source 
* Add the quantized mpt model.

* Support the quantized model for replit-code.
This commit is contained in:
Laurent Mazare
2023-10-18 16:29:38 +01:00
committed by GitHub
parent cb034506cd
commit 86e7d539d2
5 changed files with 247 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
candle-transformers/src/quantized_nn.rs
View File

@ -59,6 +59,11 @@ pub fn layer_norm(size: usize, eps: f64, vb: VarBuilder) -> Result<candle_nn::La
Ok(candle_nn::LayerNorm::new(weight, bias, eps))
}
pub fn layer_norm_no_bias(size: usize, eps: f64, vb: VarBuilder) -> Result<candle_nn::LayerNorm> {
let weight = vb.get(size, "weight")?.dequantize(vb.device())?;
Ok(candle_nn::LayerNorm::new_no_bias(weight, eps))
}
pub fn linear_no_bias(in_dim: usize, out_dim: usize, vb: VarBuilder) -> Result<Linear> {
let weight = QMatMul::new(in_dim, out_dim, vb)?;
Ok(Linear { weight, bias: None })