Fix GLM4 alignment issue (#2723)

* Fix GLM4 alignment issue

* Cleanups.

---------

Co-authored-by: Laurent <laurent.mazare@gmail.com>

candle-examples/Cargo.toml

@@ -50,7 +50,7 @@ tracing = { workspace = true }
 tracing-chrome = { workspace = true }
 tracing-subscriber = { workspace = true }
 # Necessary to disambiguate with tokio in wasm examples which are 1.28.1
-tokio = "1.29.1"
+tokio = "1.43.0"
 
 [build-dependencies]
 anyhow = { workspace = true }

candle-examples/examples/glm4/main.rs

@@ -1,12 +1,10 @@
-use candle_transformers::models::glm4::*;
-use clap::Parser;
-
 use candle::{DType, Device, Tensor};
 use candle_nn::VarBuilder;
 use candle_transformers::generation::LogitsProcessor;
+use candle_transformers::models::glm4::*;
+use clap::Parser;
 use hf_hub::{Repo, RepoType};
 use tokenizers::Tokenizer;
-
 struct TextGeneration {
     model: Model,
     device: Device,
@@ -19,7 +17,8 @@ struct TextGeneration {
 impl TextGeneration {
     #[allow(clippy::too_many_arguments)]
     fn new(model: Model, tokenizer: Tokenizer, args: Args, device: &Device, dtype: DType) -> Self {
-        let logits_processor = LogitsProcessor::new(args.seed, args.temperature, args.top_p);
+        let logits_processor =
+            LogitsProcessor::new(args.seed, Some(args.temperature), Some(args.top_p));
         Self {
             model,
             tokenizer,
@@ -125,12 +124,12 @@ struct Args {
     verbose: bool,
 
     /// The temperature used to generate samples.
-    #[arg(long)]
-    temperature: Option<f64>,
+    #[arg(long, default_value_t = 0.8)]
+    temperature: f64,
 
     /// Nucleus sampling probability cutoff.
-    #[arg(long)]
-    top_p: Option<f64>,
+    #[arg(long, default_value_t = 0.8)]
+    top_p: f64,
 
     /// The seed to use when generating random samples.
     #[arg(long, default_value_t = 299792458)]
@@ -147,7 +146,7 @@ struct Args {
     revision: Option<String>,
 
     #[arg(long)]
-    weight_file: Option<String>,
+    weight_path: Option<String>,
 
     #[arg(long)]
     tokenizer: Option<String>,
@@ -172,9 +171,7 @@ fn main() -> anyhow::Result<()> {
     );
     println!(
         "temp: {:.2} repeat-penalty: {:.2} repeat-last-n: {}",
-        args.temperature.unwrap_or(0.6),
-        args.repeat_penalty,
-        args.repeat_last_n
+        args.temperature, args.repeat_penalty, args.repeat_last_n
     );
 
     let start = std::time::Instant::now();
@@ -203,15 +200,23 @@ fn main() -> anyhow::Result<()> {
             .get("tokenizer.json")
             .map_err(anyhow::Error::msg)?,
     };
-    let filenames = match args.weight_file.as_ref() {
-        Some(weight_file) => vec![std::path::PathBuf::from(weight_file)],
-        None => candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?,
+    let config_filename = match &args.weight_path {
+        Some(path) => std::path::Path::new(path).join("config.json"),
+        _ => repo.get("config.json")?,
     };
+
+    let filenames = match &args.weight_path {
+        Some(path) => {
+            candle_examples::hub_load_local_safetensors(path, "model.safetensors.index.json")?
+        }
+        _ => candle_examples::hub_load_safetensors(&repo, "model.safetensors.index.json")?,
+    };
+
     println!("retrieved the files in {:?}", start.elapsed());
     let tokenizer = Tokenizer::from_file(tokenizer_filename).expect("Tokenizer Error");
 
     let start = std::time::Instant::now();
-    let config = Config::glm4();
+    let config: Config = serde_json::from_slice(&std::fs::read(config_filename)?)?;
     let device = candle_examples::device(args.cpu)?;
     let dtype = if device.is_cuda() {
         DType::BF16

candle-examples/src/lib.rs

@@ -4,7 +4,6 @@ pub mod coco_classes;
 pub mod imagenet;
 pub mod token_output_stream;
 pub mod wav;
-
 use candle::utils::{cuda_is_available, metal_is_available};
 use candle::{Device, Result, Tensor};
 
@@ -147,3 +146,28 @@ pub fn hub_load_safetensors(
         .collect::<Result<Vec<_>>>()?;
     Ok(safetensors_files)
 }
+
+pub fn hub_load_local_safetensors<P: AsRef<std::path::Path>>(
+    path: P,
+    json_file: &str,
+) -> Result<Vec<std::path::PathBuf>> {
+    let path = path.as_ref();
+    let jsfile = std::fs::File::open(path.join(json_file))?;
+    let json: serde_json::Value = serde_json::from_reader(&jsfile).map_err(candle::Error::wrap)?;
+    let weight_map = match json.get("weight_map") {
+        None => candle::bail!("no weight map in {json_file:?}"),
+        Some(serde_json::Value::Object(map)) => map,
+        Some(_) => candle::bail!("weight map in {json_file:?} is not a map"),
+    };
+    let mut safetensors_files = std::collections::HashSet::new();
+    for value in weight_map.values() {
+        if let Some(file) = value.as_str() {
+            safetensors_files.insert(file);
+        }
+    }
+    let safetensors_files: Vec<_> = safetensors_files
+        .into_iter()
+        .map(|v| path.join(v))
+        .collect();
+    Ok(safetensors_files)
+}