MPT fixes. (#1117)

* MPT fixes.

* Another couple fixes.

* Another shape fix.

candle-examples/examples/replit-code/main.rs

@@ -215,7 +215,7 @@ fn main() -> Result<()> {
     let config = Config::replit_code_v1_5_3b();
     let device = candle_examples::device(args.cpu)?;
     let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[filename], DType::F32, &device)? };
-    let model = Model::new(&config, vb)?;
+    let model = Model::new(&config, vb.pp("transformer"))?;
     println!("loaded the model in {:?}", start.elapsed());
 
     let mut pipeline = TextGeneration::new(

candle-transformers/src/models/mpt.rs

@@ -1,5 +1,5 @@
 #![allow(unused)]
-use crate::models::with_tracing::{linear, Embedding as E, Linear};
+use crate::models::with_tracing::{linear_no_bias, Embedding as E, Linear};
 /// MPT model used by replit-code-v1_5-3b
 /// https://huggingface.co/replit/replit-code-v1_5-3b/blob/main/modeling_mpt.py
 use candle::{DType, Device, IndexOp, Module, Result, Tensor, D};
@@ -57,11 +57,11 @@ struct GroupedQueryAttention {
 
 impl GroupedQueryAttention {
     fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
-        let wqkv_size = cfg.d_model + 2 * cfg.kv_n_heads;
-        let wqkv = linear(cfg.d_model, wqkv_size, vb.pp("Wqkv"))?;
         let head_dim = cfg.d_model / cfg.n_heads;
+        let wqkv_size = cfg.d_model + 2 * cfg.kv_n_heads * head_dim;
+        let wqkv = linear_no_bias(cfg.d_model, wqkv_size, vb.pp("Wqkv"))?;
         let softmax_scale = 1f64 / (head_dim as f64).sqrt();
-        let out_proj = linear(cfg.d_model, cfg.d_model, vb.pp("out_proj"))?;
+        let out_proj = linear_no_bias(cfg.d_model, cfg.d_model, vb.pp("out_proj"))?;
         let attn_bias = build_alibi_bias(cfg)?.to_device(vb.device())?;
         Ok(Self {
             wqkv,
@@ -155,8 +155,8 @@ struct Ffn {
 impl Ffn {
     fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
         let hidden = cfg.d_model * cfg.expansion_ratio;
-        let down_proj = linear(cfg.d_model, hidden, vb.pp("down_proj"))?;
-        let up_proj = linear(hidden, cfg.d_model, vb.pp("up_proj"))?;
+        let up_proj = linear_no_bias(cfg.d_model, hidden, vb.pp("up_proj"))?;
+        let down_proj = linear_no_bias(hidden, cfg.d_model, vb.pp("down_proj"))?;
         Ok(Self { up_proj, down_proj })
     }
 }
@@ -177,8 +177,12 @@ struct MPTBlock {
 
 impl MPTBlock {
     fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
-        let norm1 = layer_norm(cfg.d_model, 1e-5, vb.pp("norm_1"))?;
-        let norm2 = layer_norm(cfg.d_model, 1e-5, vb.pp("norm_2"))?;
+        let ln_cfg = candle_nn::LayerNormConfig {
+            affine: false,
+            ..Default::default()
+        };
+        let norm1 = layer_norm(cfg.d_model, ln_cfg, vb.pp("norm_1"))?;
+        let norm2 = layer_norm(cfg.d_model, ln_cfg, vb.pp("norm_2"))?;
         let attn = GroupedQueryAttention::new(cfg, vb.pp("attn"))?;
         let ffn = Ffn::new(cfg, vb.pp("ffn"))?;
         Ok(Self {
@@ -212,7 +216,7 @@ fn build_alibi_bias(cfg: &Config) -> Result<Tensor> {
         alibi_bias.reshape((1, 1, 1, seq_len))?
     };
     let mut n_heads2 = 1;
-    while 2 * n_heads2 <= cfg.n_heads {
+    while n_heads2 < cfg.n_heads {
         n_heads2 *= 2
     }
     let slopes = (1..=n_heads2)
@@ -230,8 +234,8 @@ fn build_alibi_bias(cfg: &Config) -> Result<Tensor> {
             .cloned()
             .collect::<Vec<f32>>()
     };
-    let slopes = Tensor::new(slopes, &Device::Cpu)?;
-    alibi_bias.broadcast_mul(&slopes)
+    let slopes = Tensor::new(slopes, &Device::Cpu)?.reshape((1, (), 1, 1))?;
+    alibi_bias.to_dtype(DType::F32)?.broadcast_mul(&slopes)
 }
 
 #[derive(Debug)]
@@ -250,7 +254,11 @@ impl Model {
             let block = MPTBlock::new(cfg, vb_b.pp(i))?;
             blocks.push(block)
         }
-        let norm_f = candle_nn::layer_norm(cfg.d_model, 1e-5, vb.pp("norm_f"))?;
+        let ln_cfg = candle_nn::LayerNormConfig {
+            affine: false,
+            ..Default::default()
+        };
+        let norm_f = candle_nn::layer_norm(cfg.d_model, ln_cfg, vb.pp("norm_f"))?;
         Ok(Self {
             wte,
             blocks,
@@ -270,6 +278,7 @@ impl Model {
             xs = block.forward(&xs, mask.as_ref())?
         }
         xs.narrow(1, seq_len - 1, 1)?
+            .squeeze(1)?
             .matmul(&self.wte.embeddings().t()?)?
             .squeeze(1)
     }