Merge branch 'main' into hf-tokenizer

.vscode/settings.json

@@ -1,3 +1,3 @@
 {
-  "rust-analyzer.cargo.features": ["convert"]
+  "rust-analyzer.cargo.features": []
 }

binaries/generate-ggml-bindings/src/main.rs

@@ -6,7 +6,7 @@ use std::path::PathBuf;
 
 fn main() {
     let bindings = bindgen::Builder::default()
-        .header("crates/ggml/sys/bindings.h")
+        .header("crates/ggml/sys/ggml/include/ggml/ggml.h")
         // Suppress some warnings
         .raw_line("#![allow(non_upper_case_globals)]")
         .raw_line("#![allow(non_camel_case_types)]")

binaries/llm-cli/src/cli_args.rs

@@ -141,13 +141,17 @@ pub struct Perplexity {
 
 #[derive(Parser, Debug)]
 pub struct Info {
-    /// The model to inspect
+    /// The model to inspect.
     #[arg(long, short = 'm')]
     pub model_path: PathBuf,
 
-    /// Whether or not to dump the entire vocabulary
+    /// Show all of the tensors in the model, including their names, formats and shapes.
+    #[arg(long, short = 't')]
+    pub tensors: bool,
+
+    /// Show all of the tokens in the vocabulary.
     #[arg(long, short = 'v')]
-    pub dump_vocabulary: bool,
+    pub vocabulary: bool,
 }
 
 #[derive(Parser, Debug)]
@@ -371,7 +375,6 @@ impl ModelLoad {
             prefer_mmap: !self.no_mmap,
             context_size: self.num_ctx_tokens,
             lora_adapters: self.lora_paths.clone(),
-            ..Default::default()
         };
 
         let mut sp = Some(spinoff::Spinner::new(

binaries/llm-cli/src/main.rs

@@ -71,7 +71,7 @@ fn infer<M: llm::KnownModel + 'static>(
         &mut rng,
         &llm::InferenceRequest {
             prompt: prompt.as_str().into(),
-            parameters: Some(&inference_params),
+            parameters: &inference_params,
             play_back_previous_tokens: session_loaded,
             maximum_token_count: args.generate.num_predict,
         },
@@ -159,20 +159,19 @@ fn info<M: llm::KnownModel + 'static>(args: &cli_args::Info) -> Result<()> {
 
     log::info!("Container type: {:?}", loader.container_type);
     log::info!("Hyperparameters: {:?}", loader.hyperparameters);
-    log::info!(
-        "Tensors: {:?}",
-        loader
-            .tensors
-            .iter()
-            .map(|(name, tensor)| format!("{} ({:?})", name, tensor.element_type))
-            .collect::<Vec<_>>()
-    );
     log::info!("Vocabulary size: {}", loader.vocabulary.len());
 
-    if args.dump_vocabulary {
-        log::info!("Dumping vocabulary:");
+    if args.vocabulary {
+        log::info!("Vocabulary:");
         for i in 0..loader.vocabulary.len() {
-            log::info!("{}: {}", i, utf8_or_array(&loader.vocabulary.token(i)));
+            log::info!("- {}: {}", i, utf8_or_array(&loader.vocabulary.token(i)));
+        }
+    }
+
+    if args.tensors {
+        log::info!("Tensors:");
+        for (name, tensor) in &loader.tensors {
+            log::info!("- {} ({:?} {:?})", name, tensor.element_type, tensor.dims());
         }
     }
 
@@ -274,7 +273,7 @@ fn interactive<M: llm::KnownModel + 'static>(
                     &mut rng,
                     &llm::InferenceRequest {
                         prompt: "".into(),
-                        parameters: Some(&inference_params),
+                        parameters: &inference_params,
                         play_back_previous_tokens: session_loaded,
                         maximum_token_count: args.generate.num_predict,
                     },

crates/ggml/src/lib.rs

@@ -13,7 +13,6 @@ mod context;
 mod tensor;
 
 pub mod format;
-pub mod legacy;
 pub mod util;
 
 pub use context::Context;
@@ -140,7 +139,7 @@ pub enum Type {
     F32,
 
     /// Legacy: Quantized 4-bit (type 2).
-    /// This is not supported by modern `ggml` and is only here for use with [legacy].
+    /// This is not supported by modern `ggml` versions.
     LegacyQ4_2,
 }
 impl From<Type> for sys::ggml_type {

crates/ggml/sys/build.rs

@@ -9,8 +9,8 @@ fn main() {
     let mut builder = cc::Build::new();
 
     let build = builder
-        .files(["ggml/src/ggml.c", "legacy/qnt0.c"])
-        .includes(["ggml/include/ggml", "legacy"]);
+        .files(["ggml/src/ggml.c"])
+        .includes(["ggml/include/ggml"]);
 
     // This is a very basic heuristic for applying compile flags.
     // Feel free to update this to fit your operating system.