From 86e7d539d253740d5a0579e6f53acc12e30d3e4c Mon Sep 17 00:00:00 2001
From: Laurent Mazare <laurent.mazare@gmail.com>
Date: Wed, 18 Oct 2023 16:29:38 +0100
Subject: Add the quantized mpt model. (#1123)

* Add the quantized mpt model.

* Support the quantized model for replit-code.
---
 candle-transformers/src/quantized_nn.rs | 5 +++++
 1 file changed, 5 insertions(+)

(limited to 'candle-transformers/src/quantized_nn.rs')

diff --git a/candle-transformers/src/quantized_nn.rs b/candle-transformers/src/quantized_nn.rs
index 1745327d..d71c3b60 100644
--- a/candle-transformers/src/quantized_nn.rs
+++ b/candle-transformers/src/quantized_nn.rs
@@ -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 })
-- 
cgit v1.2.3