Merge branch 'main' of github.com:rustformers/llama-rs into quantize

ggml-loader/src/lib.rs

@@ -37,8 +37,8 @@ pub enum LoadError<T> {
  #[error("invalid file magic number: {0}")]
  InvalidMagic(u32),
 
- #[error("invalid ggml format: version={0}")]
- InvalidFormatVersion(u32),
+ #[error("invalid ggml format: format={0:?} version={1}")]
+ InvalidFormatVersion(ContainerType, u32),
 
  #[error("{0}")]
  Io(#[from] std::io::Error),
@@ -129,7 +129,7 @@ pub fn load_model_from_reader<T, R: BufRead + Seek>(
  ContainerType::GGMF | ContainerType::GGJT => {
  let _version: u32 = match read_u32(reader)? {
  ggml::FORMAT_VERSION => ggml::FORMAT_VERSION,
- version => return Err(LoadError::InvalidFormatVersion(version)),
+ version => return Err(LoadError::InvalidFormatVersion(container_type, version)),
  };
  }
  ContainerType::GGML => {}

llama-rs/Cargo.toml

@@ -29,4 +29,4 @@ half = { version = "2.2.1", optional = true }
 
 [features]
 convert = ["dep:serde_json", "dep:protobuf", "dep:rust_tokenizers"]
-quantize = ["dep:half"]
+quantize = ["dep:half"]

llama-rs/src/file.rs

@@ -2,56 +2,18 @@ use crate::LoadError;
 pub use std::fs::File;
 pub use std::io::{BufRead, BufReader, BufWriter, Read, Write};
 
-pub fn read_bytes<const N: usize>(reader: &mut impl BufRead) -> Result<[u8; N], LoadError> {
- let mut bytes = [0u8; N];
- read(reader, &mut bytes)?;
- Ok(bytes)
-}
-
-pub fn read_bytes_with_len(reader: &mut impl BufRead, len: usize) -> Result<Vec<u8>, LoadError> {
- let mut bytes = vec![0u8; len];
- read(reader, &mut bytes)?;
- Ok(bytes)
-}
-
-pub fn read_i32(reader: &mut impl BufRead) -> Result<i32, LoadError> {
- Ok(i32::from_le_bytes(read_bytes::<4>(reader)?))
-}
-
 pub fn rw_i32(reader: &mut impl BufRead, writer: &mut impl Write) -> Result<i32, LoadError> {
  Ok(i32::from_le_bytes(rw::<4>(reader, writer)?))
 }
 
-pub fn read_u32(reader: &mut impl BufRead) -> Result<u32, LoadError> {
- Ok(u32::from_le_bytes(read_bytes::<4>(reader)?))
-}
-
 pub fn rw_u32(reader: &mut impl BufRead, writer: &mut impl Write) -> Result<u32, LoadError> {
  Ok(u32::from_le_bytes(rw::<4>(reader, writer)?))
 }
 
-pub fn read_f32(reader: &mut impl BufRead) -> Result<f32, LoadError> {
- Ok(f32::from_le_bytes(read_bytes::<4>(reader)?))
-}
-
 pub fn rw_f32(reader: &mut impl BufRead, writer: &mut impl Write) -> Result<f32, LoadError> {
  Ok(f32::from_le_bytes(rw::<4>(reader, writer)?))
 }
 
-/// Helper function. Reads a string from the buffer and returns it.
-pub fn read_string(reader: &mut BufReader<File>, len: usize) -> Result<String, LoadError> {
- Ok(String::from_utf8(read_bytes_with_len(reader, len)?)?)
-}
-
-fn read(reader: &mut impl BufRead, bytes: &mut [u8]) -> Result<(), LoadError> {
- reader
- .read_exact(bytes)
- .map_err(|e| LoadError::ReadExactFailed {
- source: e,
- bytes: bytes.len(),
- })
-}
-
 pub fn rw_bytes_with_len(
  reader: &mut impl BufRead,
  writer: &mut impl Write,
@@ -72,8 +34,7 @@ fn rw<const N: usize>(
  reader: &mut impl BufRead,
  writer: &mut impl Write,
 ) -> Result<[u8; N], LoadError> {
- let mut bytes = [0u8; N];
- read(reader, &mut bytes)?;
+ let bytes: [u8; N] = ggml_loader::util::read_bytes(reader)?;
  writer.write_all(&bytes)?;
  Ok(bytes)
 }