boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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
54
55
56
57
58
59
60
61
62
63
64
65
66

//! Mixture of Experts configuration.

use crate::error::{Error, Result};
use serde::{Deserialize, Serialize};

/// MoE (Mixture of Experts) configuration
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MoeConfig {
    /// Number of experts
    pub num_experts: usize,

    /// Number of experts per token (top-k)
    pub experts_per_tok: usize,

    /// Enable shared expert
    #[serde(default)]
    pub shared_expert: Option<bool>,

    /// Expert intermediate size
    #[serde(default)]
    pub intermediate_size: Option<usize>,

    /// Load balance loss coefficient
    #[serde(default = "default_load_balance_alpha")]
    pub load_balance_alpha: f64,

    /// Router z-loss coefficient
    #[serde(default = "default_z_loss_alpha")]
    pub z_loss_alpha: f64,
}

fn default_load_balance_alpha() -> f64 {
    0.01
}

fn default_z_loss_alpha() -> f64 {
    1e-3
}

impl MoeConfig {
    pub fn validate(&self) -> Result<()> {
        if self.num_experts == 0 {
            return Err(Error::ModelError {
                reason: "moe.num_experts must be > 0".into(),
            });
        }
        if self.experts_per_tok == 0 {
            return Err(Error::ModelError {
                reason: "moe.experts_per_tok must be > 0".into(),
            });
        }
        if self.experts_per_tok > self.num_experts {
            return Err(Error::ModelError {
                reason: format!(
                    "moe.experts_per_tok ({}) cannot exceed moe.num_experts ({})",
                    self.experts_per_tok, self.num_experts
                ),
            });
        }
        Ok(())
    }

    pub fn has_shared_expert(&self) -> bool {
        self.shared_expert.unwrap_or(false)
    }
}