entrenar 0.7.11

Training & Optimization library with autograd, LoRA, quantization, and model merging
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
43
44
45
46
47
48
49
50
51
52
53

//! LoRA Configuration
//!
//! Contains Low-Rank Adaptation configuration types for training manifests.

use serde::{Deserialize, Serialize};

/// LoRA (Low-Rank Adaptation) configuration
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct LoraConfig {
    /// Enable LoRA
    pub enabled: bool,

    /// Rank of low-rank matrices (r)
    pub rank: usize,

    /// Scaling factor (alpha)
    pub alpha: f64,

    /// LoRA dropout
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub dropout: Option<f64>,

    /// Target modules for LoRA
    pub target_modules: Vec<String>,

    /// Target modules pattern (regex)
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub target_modules_pattern: Option<String>,

    /// Bias training (none, all, lora_only)
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub bias: Option<String>,

    /// Weight initialization (gaussian, xavier, kaiming)
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub init_weights: Option<String>,

    /// QLoRA: Quantize base model
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub quantize_base: Option<bool>,

    /// QLoRA: Quantization bits
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub quantize_bits: Option<u8>,

    /// QLoRA: Double quantization
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub double_quantize: Option<bool>,

    /// QLoRA: Quantization type (nf4, fp4)
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub quant_type: Option<String>,
}