Ok.

candle-core/examples/llama/main.rs

@@ -15,7 +15,7 @@ use anyhow::{Error as E, Result};
 use clap::Parser;
 
 use candle::{DType, Device, Tensor};
-use candle_hub::{Repo, api::Api, RepoType};
+use candle_hub::{api::Api, Repo, RepoType};
 use std::collections::HashMap;
 use std::sync::{Arc, Mutex};
 
@@ -439,7 +439,10 @@ async fn main() -> Result<()> {
         .to_vec();
 
     let mut filenames = vec![];
-    for rfilename in ["model-00001-of-00002.safetensors", "model-00002-of-00002.safetensors"]{
+    for rfilename in [
+        "model-00001-of-00002.safetensors",
+        "model-00002-of-00002.safetensors",
+    ] {
         let filename = api.get(&repo, rfilename).await?;
         filenames.push(filename);
     }

candle-core/examples/llama/weights.rs

@@ -1,12 +1,15 @@
+use super::*;
+use candle::{Device, Result, Tensor};
+use half::f16;
 use memmap2::MmapOptions;
-use candle::{Device, Result, Shape, Tensor, WithDType};
+use safetensors::{
+    tensor::{Dtype, TensorView},
+    SafeTensors,
+};
 use std::fs::File;
 use std::path::PathBuf;
-use super::*;
-use safetensors::{SafeTensors, tensor::{Dtype, TensorView}};
-use half::f16;
 
-fn convert<'a>(view: TensorView<'a>, device: &Device) -> Result<Tensor>{
+fn convert(view: TensorView<'_>, device: &Device) -> Result<Tensor> {
     match view.dtype() {
         Dtype::F16 => {
             let v = view.data();
@@ -25,9 +28,8 @@ fn convert<'a>(view: TensorView<'a>, device: &Device) -> Result<Tensor>{
                 }
                 Tensor::from_slice(&c, view.shape(), device)
             }
-
         }
-        dt => todo!("Unhandled dtype {dt:?}")
+        dt => todo!("Unhandled dtype {dt:?}"),
     }
 }
 
@@ -37,7 +39,6 @@ pub struct VarBuilder<'a>{
     device: Device,
 }
 
-
 impl<'a> VarBuilder<'a> {
     pub fn new(safetensors: Vec<SafeTensors<'a>>, device: Device) -> Self {
         let mut routing = HashMap::new();
@@ -50,7 +51,7 @@ impl<'a> VarBuilder<'a>{
         Self {
             safetensors,
             device,
-            routing
+            routing,
         }
     }
 
@@ -60,7 +61,6 @@ impl<'a> VarBuilder<'a>{
         let view = self.safetensors[*index].tensor(tensor_name).unwrap();
         let tensor = convert(view, &self.device)?;
         Ok(tensor)
-
     }
 }
 
@@ -71,8 +71,10 @@ impl Linear{
     }
 
     fn load_multi(prefixes: &[&str], vb: &VarBuilder) -> Result<Self> {
-        let weights: Vec<_> = prefixes.iter().map(|p| vb.get(&format!("{p}.weight")).unwrap()).collect();
+        let weights: Vec<_> = prefixes
-        println!("shapes {:?}", weights.iter().map(|w| w.shape()).collect::<Vec<_>>());
+            .iter()
+            .map(|p| vb.get(&format!("{p}.weight")).unwrap())
+            .collect();
         let weight = Tensor::cat(&weights, 0)?;
         Ok(Self::new(weight))
     }
@@ -87,14 +89,21 @@ impl RmsNorm{
 
 impl CausalSelfAttention {
     fn load(prefix: &str, vb: &VarBuilder, cache: &Cache, config: &Config) -> Result<Self> {
-        let c_attn = Linear::load_multi(&[&format!("{prefix}.q_proj"), &format!("{prefix}.k_proj"), &format!("{prefix}.v_proj")], vb)?;
+        let c_attn = Linear::load_multi(
+            &[
+                &format!("{prefix}.q_proj"),
+                &format!("{prefix}.k_proj"),
+                &format!("{prefix}.v_proj"),
+            ],
+            vb,
+        )?;
         let o_proj = Linear::load(&format!("{prefix}.o_proj"), vb)?;
         Ok(Self::new(c_attn, o_proj, config.n_head, cache))
     }
 }
 
 impl Mlp {
-    fn load(prefix: &str, vb: &VarBuilder, config: &Config) -> Result<Self>{
+    fn load(prefix: &str, vb: &VarBuilder) -> Result<Self> {
         let c_fc1 = Linear::load(&format!("{prefix}.gate_proj"), vb)?;
         let c_fc2 = Linear::load(&format!("{prefix}.up_proj"), vb)?;
         let c_proj = Linear::load(&format!("{prefix}.down_proj"), vb)?;
@@ -105,24 +114,40 @@ impl Mlp{
 impl Block {
     fn load(prefix: &str, vb: &VarBuilder, cache: &Cache, config: &Config) -> Result<Self> {
         let attn = CausalSelfAttention::load(&format!("{prefix}.self_attn"), vb, cache, config)?;
-        let mlp = Mlp::load(&format!("{prefix}.mlp"), vb, config)?;
+        let mlp = Mlp::load(&format!("{prefix}.mlp"), vb)?;
         let input_layernorm = RmsNorm::load(&format!("{prefix}.input_layernorm"), vb)?;
-        let post_attention_layernorm = RmsNorm::load(&format!("{prefix}.post_attention_layernorm"), vb)?;
+        let post_attention_layernorm =
-        Ok(Self::new(input_layernorm, attn, post_attention_layernorm, mlp))
+            RmsNorm::load(&format!("{prefix}.post_attention_layernorm"), vb)?;
+        Ok(Self::new(
+            input_layernorm,
+            attn,
+            post_attention_layernorm,
+            mlp,
+        ))
     }
 }
 
 impl Llama {
-    pub fn load(device: &Device, filenames: &[PathBuf], cache: &Cache, config: &Config) -> Result<Self>{
+    pub fn load(
-        let handles: Vec<_> = filenames.iter().map(|f| {
+        device: &Device,
+        filenames: &[PathBuf],
+        cache: &Cache,
+        config: &Config,
+    ) -> Result<Self> {
+        let handles: Vec<_> = filenames
+            .iter()
+            .map(|f| {
                 let file = File::open(f).unwrap();
-            let buffer = unsafe { MmapOptions::new().map(&file).unwrap() };
+                unsafe { MmapOptions::new().map(&file).unwrap() }
-            buffer
+            })
-        }).collect();
+            .collect();
-        let tensors: Vec<_> = handles.iter().map(|h| {
+        let tensors: Vec<_> = handles
+            .iter()
+            .map(|h| {
                 let tensors = SafeTensors::deserialize(h).unwrap();
                 tensors
-        }).collect();
+            })
+            .collect();
 
         let vb = VarBuilder::new(tensors, device.clone());
 
@@ -130,15 +155,10 @@ impl Llama{
         let wte = Embedding::new(embedding);
         let lm_head = Linear::load("lm_head", &vb)?;
         let norm = RmsNorm::load("model.norm", &vb)?;
-        let blocks: Vec<_> = (0..config.n_layer).map(|i| Block::load(&format!("model.layers.{i}"), &vb, cache, config).unwrap()).collect();
+        let blocks: Vec<_> = (0..config.n_layer)
+            .map(|i| Block::load(&format!("model.layers.{i}"), &vb, cache, config).unwrap())
+            .collect();
 
-        Ok(Self::new(
+        Ok(Self::new(wte, blocks, norm, lm_head))
-            wte,
-            blocks,
-            norm,
-            lm_head
-        ))
     }
 }
-
-

candle-hub/src/api.rs

@@ -397,7 +397,6 @@ impl Api {
             let parallel_failures_semaphore = parallel_failures_semaphore.clone();
             let progress = progressbar.clone();
             handles.push(tokio::spawn(async move {
-                println!("Start {start:?} - {stop:?}");
                 let mut chunk = Self::download_chunk(&client, &url, &filename, start, stop).await;
                 let mut i = 0;
                 if parallel_failures > 0 {