Llama prefix

catgrad-llm/examples/siglip/main.rs

@@ -575,10 +575,7 @@ pub fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     let interp = Interpreter::new(backend, env, parameters);
 
-    let results = interp.eval(
-        interp.environment.to_core(typed_term.term),
-        vec![input_tensor, image_tensor],
-    )?;
+    let results = interp.run(typed_term.term, vec![input_tensor, image_tensor])?;
     let result_tensor = match &results[1] {
         interpreter::Value::Tensor(t) => t,
         _ => panic!("Expected tensor output"),

catgrad-llm/src/models/gemma3.rs

@@ -11,6 +11,7 @@ pub enum GemmaConfig {
     VLM {
         text_config: GemmaTextConfig,
         image_token_index: usize,
+        #[serde(default)]
         mm_tokens_per_image: usize,
     },
     #[serde(untagged)]

catgrad-llm/src/models/llama.rs

@@ -4,6 +4,7 @@ use catgrad::prelude::ops::*;
 use catgrad::prelude::*;
 use nn::*;
 pub struct LlamaModel {
+    pub root: String,
     pub config: Config,
     pub max_sequence_length: usize,
 }
@@ -65,8 +66,8 @@ impl LlamaModel {
 
         let sh = shape!(builder, b, s, num_kv_heads, head_dim);
         let k = reshape(builder, sh.clone(), k);
-
         let v = reshape(builder, sh, v);
+
         let sh = shape!(builder, b, s, num_heads, head_dim);
         let q = reshape(builder, sh, q);
 
@@ -157,7 +158,12 @@ impl Module<1, 1> for LlamaModel {
     }
 
     fn def(&self, builder: &Builder, [x]: [Var; 1]) -> [Var; 1] {
-        let root = self.path();
+        let mut root = self.path();
+        if !self.root.is_empty() {
+            root = root
+                .extend(self.root.split('.').collect::<Vec<&str>>())
+                .unwrap();
+        }
 
         let mut cache = Cache::init(builder, &self.config, self.max_sequence_length);
 

catgrad-llm/src/utils.rs

@@ -243,6 +243,7 @@ pub fn get_model(
             ))
         }
         "MistralForCausalLM" | "LlamaForCausalLM" => Box::new(llama::LlamaModel {
+            root: "".to_string(),
             config: config.clone(),
             max_sequence_length,
         }),