Pre-tokenized evaluation mode for llama2.c. (#291)

candle-examples/examples/llama2-c/main.rs

@@ -215,8 +215,13 @@ struct Args {
     #[arg(long, default_value = "")]
     prompt: String,
 
+    /// A directory with the pre-tokenized dataset in the format generated by the tinystories.py
+    /// script from llama2.c https://github.com/karpathy/llama2.c
     #[arg(long)]
-    eval_file: Option<String>,
+    pretokenized_dir: Option<String>,
+
+    #[arg(long, default_value_t = 32)]
+    batch_size: usize,
 }
 
 fn main() -> anyhow::Result<()> {
@@ -243,13 +248,7 @@ fn main() -> anyhow::Result<()> {
 
     match args.task {
         Task::Inference => run_inference(tokenizer, &config_path, args)?,
-        Task::Evaluation => {
+        Task::Evaluation => run_eval(tokenizer, &config_path, args)?,
-            if let Some(eval_file) = &args.eval_file {
-                run_eval_file(eval_file.into(), &config_path, args)?
-            } else {
-                run_eval(tokenizer, &config_path, args)?
-            }
-        }
         Task::Training => todo!(),
     }
     Ok(())
@@ -278,7 +277,6 @@ fn run_eval_file(
     println!("dataset loaded: {} tokens", tokens.len());
 
     let seq_len = model.config.seq_len;
-    let batch_size = 32;
     let mut inputs = vec![];
     let mut targets = vec![];
     for start_idx in (0..tokens.len()).step_by(seq_len) {
@@ -290,7 +288,7 @@ fn run_eval_file(
         let targets_ = Tensor::new(&tokens[1..], &device)?;
         inputs.push(inputs_);
         targets.push(targets_);
-        if inputs.len() >= batch_size {
+        if inputs.len() >= args.batch_size {
             let inp = Tensor::stack(&inputs, 0)?;
             let tgt = Tensor::stack(&targets, 0)?;
             let logits = model.forward(&inp, 0)?;
@@ -314,6 +312,8 @@ fn run_eval(tokenizer: Tokenizer, config_path: &std::path::PathBuf, args: Args)
     let cache = model::Cache::new(false, &config, vb.pp("rot"))?;
     let model = Llama::load(vb, &cache, config)?;
 
+    let tokens = match args.pretokenized_dir {
+        None => {
             let api = hf_hub::api::sync::Api::new()?;
             let model_id = "roneneldan/TinyStories"; // TODO: Make this configurable.
             println!("loading the evaluation dataset from {}", model_id);
@@ -323,23 +323,44 @@ fn run_eval(tokenizer: Tokenizer, config_path: &std::path::PathBuf, args: Args)
             let file = std::io::BufReader::new(file);
             let mut tokens = vec![];
             for line in file.lines() {
-        let line = line?.replace("<|endoftext|>", "");
+                let line = line?.replace("<|endoftext|>", "<s>");
                 let line = tokenizer.encode(line, false).map_err(E::msg)?;
                 tokens.push(line.get_ids().to_vec())
             }
-    let tokens = tokens.concat();
+            tokens.concat()
+        }
+        Some(pretokenized_dir) => {
+            let path = std::path::PathBuf::from(pretokenized_dir).join("data00.bin");
+            let bytes = std::fs::read(path)?;
+            // Tokens are encoded as u16.
+            let mut tokens = vec![0u16; bytes.len() / 2];
+            std::io::Cursor::new(bytes).read_u16_into::<LittleEndian>(&mut tokens)?;
+            tokens.into_iter().map(|u| u as u32).collect::<Vec<u32>>()
+        }
+    };
     println!("dataset loaded and encoded: {} tokens", tokens.len());
-    let seq_len = 256;
+
+    let seq_len = model.config.seq_len;
+    let mut inputs = vec![];
+    let mut targets = vec![];
     for start_idx in (0..tokens.len()).step_by(seq_len) {
         if start_idx + seq_len + 1 > tokens.len() {
             break;
         }
         let tokens = &tokens[start_idx..start_idx + seq_len + 1];
-        let inputs = Tensor::new(&tokens[..seq_len], &device)?.unsqueeze(0)?;
+        let inputs_ = Tensor::new(&tokens[..seq_len], &device)?;
-        let targets = Tensor::new(&tokens[1..], &device)?;
+        let targets_ = Tensor::new(&tokens[1..], &device)?;
-        let logits = model.forward(&inputs, 0)?.squeeze(0)?;
+        inputs.push(inputs_);
-        let loss = candle_nn::loss::cross_entropy(&logits, &targets)?;
+        targets.push(targets_);
-        println!("{start_idx} {}", loss.to_vec0::<f32>()?);
+        if inputs.len() >= args.batch_size {
+            let inp = Tensor::stack(&inputs, 0)?;
+            let tgt = Tensor::stack(&targets, 0)?;
+            let logits = model.forward(&inp, 0)?;
+            let loss = candle_nn::loss::cross_entropy(&logits.flatten_to(1)?, &tgt.flatten_to(1)?)?;
+            println!("{}", loss.to_vec0::<f32>()?);
+            inputs.clear();
+            targets.clear();
+        }
     }
     Ok(())
 }