Read parameters from model's JSON file instead of hard-coding them, make max sequence length configurable.

Cargo.lock

@@ -800,6 +800,8 @@ dependencies = [
  "protobuf-parse",
  "rand",
  "rayon",
+ "serde",
+ "serde_json",
  "thiserror",
 ]
 

Cargo.toml

@@ -22,6 +22,8 @@ rayon = "1.7"
 clap = { version = "4.1", features = ["derive"] }
 indicatif = "0.17"
 colored = "2"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
 
 # We need protobuf compiler
 [build-dependencies]

src/rllama_main.rs

@@ -5,6 +5,7 @@ use crate::transformer::Transformer;
 use crate::unpickler;
 use clap::Parser;
 use colored::Colorize;
+use serde::{Deserialize, Serialize};
 use std::io::{Read, Write};
 
 #[derive(Parser)]
@@ -14,11 +15,17 @@ struct Cli {
     model_path: String,
     #[arg(long)]
     tokenizer_path: String,
+    #[arg(long)]
+    param_path: String,
+
     #[arg(long)]
     prompt: Option<String>,
     #[arg(long)]
     prompt_file: Option<String>,
 
+    #[arg(long)]
+    max_seq_len: Option<usize>,
+
     #[arg(long)]
     temperature: Option<f32>,
     #[arg(long)]
@@ -27,10 +34,21 @@ struct Cli {
     top_k: Option<i32>,
 }
 
+#[derive(Clone, Serialize, Deserialize)]
+struct ModelParams {
+    dim: usize,
+    multiple_of: usize,
+    n_heads: usize,
+    n_layers: usize,
+    norm_eps: f64,
+    vocab_size: i64,
+}
+
 pub fn main() -> Result<(), Box<dyn std::error::Error>> {
     let cli = Cli::parse();
     let model_path = cli.model_path;
     let tokenizer_path = cli.tokenizer_path;
+    let param_path = cli.param_path;
 
     let mut be_quiet: bool = false;
     if !colored::control::SHOULD_COLORIZE.should_colorize() {
@@ -46,6 +64,14 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
         };
     }
 
+    // Read ModelParams from param_path, we expect it to be JSON
+    let mut fs = std::fs::File::open(&param_path)?;
+    let mut bs = Vec::new();
+    fs.read_to_end(&mut bs)?;
+    std::mem::drop(fs);
+    let params: ModelParams = serde_json::from_slice(&bs)?;
+    pln!("Loaded model parameters from {}.", param_path);
+
     let prompt: String = match (cli.prompt, cli.prompt_file) {
         (Some(prompt), None) => {
             pln!("Using prompt: {}", prompt);
@@ -84,19 +110,20 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
     pln!("Loading embeddings from {}...", model_data_dir);
     let emb = Embedding::from_unpickled(&result, model_data_dir.clone())?;
 
-    let max_seq_len = 512;
+    let max_seq_len = match cli.max_seq_len {
+        Some(max_seq_len) => max_seq_len,
+        None => 1024,
+    };
 
     pln!("Loading transformer weights from {}...", model_data_dir);
     let tr = Transformer::from_unpickled(
         &result,
         emb,
-        4096,
-        32,
-        32,
+        params.dim,
+        params.n_layers,
+        params.n_heads,
         max_seq_len,
-        1e-6,
-        32,
-        128,
+        params.norm_eps,
         model_data_dir,
     )?;
     pln!("All is loaded. Starting inference.");
@@ -116,6 +143,14 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
     }
 
     pln!("---");
+    pln!(" dim: {}", params.dim);
+    pln!(" multiple_of: {}", params.multiple_of);
+    pln!(" n_heads: {}", params.n_heads);
+    pln!(" n_layers: {}", params.n_layers);
+    pln!(" norm_eps: {}", params.norm_eps);
+    pln!(" vocab_size: {}", params.vocab_size);
+    pln!("---");
+    pln!("Max sequence length: {}", max_seq_len);
     pln!("Temperature: {}", token_sampler.get_temperature());
     pln!("Top P: {}", token_sampler.get_top_p());
     pln!("Top K: {}", token_sampler.get_top_k());
@@ -126,12 +161,7 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
     );
     pln!(
         "{}",
-        "  This is the color of the generated text while full context is available"
-            .truecolor(128, 255, 128)
-    );
-    pln!(
-        "{}",
-        "  Remaining text is in this color".truecolor(255, 128, 128)
+        "  This is the color of the generated text".truecolor(128, 255, 128)
     );
     pln!("---");
     print!("{}", prompt.as_str().truecolor(128, 128, 255));
@@ -139,17 +169,8 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     let mut caches = tr.make_caches();
     let mut first: bool = true;
-    let mut shifts: usize = 0;
-    loop {
-        if toks_id.len() >= max_seq_len {
-            toks_id = toks_id[1..].to_vec();
-            prev_pos -= 1;
-            caches.shift_left(1);
-            shifts += 1;
-            // TODO: it seems that text beyond context is just broken.
-            // Maybe I cannot just go and shift it.
-        }
-        let preds = tr.forward(&toks_id[prev_pos..], prev_pos, &mut caches, shifts);
+    while toks_id.len() < max_seq_len {
+        let preds = tr.forward(&toks_id[prev_pos..], prev_pos, &mut caches);
         let highest_pred_idx = token_sampler.sample(&preds);
         toks_id.push(highest_pred_idx as TokenId);
 
@@ -166,14 +187,14 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
             }
             if first && tok_idx < toks_id.len() - 1 {
                 // intentionally left empty
-            } else if shifts == 0 {
-                print!("{}", tok_str.truecolor(128, 255, 128));
             } else {
-                print!("{}", tok_str.truecolor(255, 128, 128));
+                print!("{}", tok_str.truecolor(128, 255, 128));
             }
         }
         let _ = std::io::stdout().flush();
         prev_pos = toks_id.len() - 1;
         first = false;
     }
+    println!("");
+    Ok(())
 }

src/transformer.rs

@@ -121,10 +121,14 @@ impl Transformer {
         n_heads: usize,
         max_seq_len: usize,
         eps: f64,
-        n_local_heads: usize,
-        head_dim: usize,
         data_dir: P,
     ) -> Result<Transformer, UnpicklingError> {
+        assert_eq!(dim % n_heads, 0);
+        let head_dim = dim / n_heads;
+        let n_local_heads = n_heads; // I think the local heads is an artifact of the original
+                                     // implementation that used multi-GPU in the Facebook repo.
+                                     // Should delete it later.
+
         let data_dir: &Path = data_dir.as_ref();
 
         let progress_bar = ProgressBar::new(n_layers as u64);
@@ -149,9 +153,7 @@ impl Transformer {
         let output = Tensor::from_unpickled(unpickled, "output.weight", data_dir)?.to_f32();
 
         Ok(Transformer {
-            // TODO: maybe rotary embedding can be computed on the fly if the sequence gets really long. I just
-            // slapped * 20 on max seq len here.
-            freqs_cis: compute_freqs_cis(dim / n_heads, max_seq_len * 20, 10000.0),
+            freqs_cis: compute_freqs_cis(dim / n_heads, max_seq_len, 10000.0),
             emb,
             dim,
             n_layers,
@@ -186,7 +188,6 @@ impl Transformer {
         tokens: &[TokenId],
         start_pos: usize,
         caches: &mut TransformerCaches,
-        shifts: usize,
     ) -> Tensor {
         assert!(caches.layer_caches.len() == self.n_layers);
         let mask: Option<Tensor> = if tokens.len() > 1 {
@@ -215,7 +216,6 @@ impl Transformer {
                 &self.freqs_cis,
                 &mask,
                 &mut caches.layer_caches[idx],
-                shifts,
             );
         }
         let out = self.norm.forward(&emb_tensor);
@@ -265,17 +265,11 @@ impl TransformerBlock {
         freqs_cis: &FreqsCis,
         mask: &Option<Tensor>,
         attention_cache: &mut AttentionCache,
-        shifts: usize,
     ) -> Tensor {
         let attnorm_out = self.attention_norm.forward(x);
-        let att_out = self.attn.forward(
-            &attnorm_out,
-            start_pos,
-            freqs_cis,
-            mask,
-            attention_cache,
-            shifts,
-        );
+        let att_out = self
+            .attn
+            .forward(&attnorm_out, start_pos, freqs_cis, mask, attention_cache);
         let h = x.add(&att_out);
         let att_out = self.ffn_norm.forward(&h);
         let att_out = self.feed_forward.forward(&att_out).transpose();
@@ -404,7 +398,6 @@ impl Attention {
         freqs_cis: &FreqsCis,
         mask: &Option<Tensor>,
         attention_cache: &mut AttentionCache,
-        shifts: usize,
     ) -> Tensor {
         let seq_len = x.rows();
         let (xq_out, (xk_out, xv_out)) = rayon::join(
@@ -432,13 +425,8 @@ impl Attention {
                 .row(idx)
                 .view(self.n_local_heads as i64, self.head_dim as i64);
 
-            let (xq_row, xk_row) = apply_rotary_emb(
-                &xq_row,
-                &xk_row,
-                freqs_cis,
-                idx as usize,
-                start_pos + shifts,
-            );
+            let (xq_row, xk_row) =
+                apply_rotary_emb(&xq_row, &xk_row, freqs_cis, idx as usize, start_pos);
 
             xq_views.push(xq_row);
             xk_views.push(xk_row);