rig-candle 0.41.0

Local Candle Llama, SmolLM2, and Qwen3 completion models for Rig
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

//! Caller-supplied model artifact buffers and inexpensive buffer validation.

use crate::CandleError;

/// Owned model artifacts for exactly one unsharded checkpoint.
#[derive(Debug)]
pub struct ModelData {
    /// Contents of `config.json`.
    pub config: Vec<u8>,
    /// Contents of `tokenizer.json`.
    pub tokenizer: Vec<u8>,
    /// Contents of one safetensors or GGUF checkpoint, as identified by [`ModelArtifacts`].
    pub weights: Vec<u8>,
}

/// Borrowed GGUF artifacts for zero-copy loading from embedded/static bytes.
#[derive(Debug, Clone, Copy)]
pub struct GgufModelData<'a> {
    /// Contents of `config.json`.
    pub config: &'a [u8],
    /// Contents of `tokenizer.json`.
    pub tokenizer: &'a [u8],
    /// Contents of one GGUF checkpoint.
    pub weights: &'a [u8],
}

/// Byte-backed checkpoint format supplied to [`crate::CandleModel`].
#[derive(Debug)]
pub enum ModelArtifacts {
    /// One unsharded Hugging Face safetensors checkpoint.
    Safetensors(ModelData),
    /// A validated SmolLM2 or Qwen3 Q4_K_M GGUF checkpoint.
    Gguf(ModelData),
}

pub(crate) fn require_nonempty(bytes: &[u8], artifact: &'static str) -> Result<(), CandleError> {
    if bytes.is_empty() {
        Err(CandleError::EmptyBuffer { artifact })
    } else {
        Ok(())
    }
}