mistralrs-core 0.8.1

Fast, flexible LLM inference.
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
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101

use std::collections::HashMap;

use mistralrs_quant::{QuantizedConfig, StaticLoraConfig};
use serde::{Deserialize, Serialize};

use crate::{
    layers::{Activation, Phi4MMRopeScalingConfig},
    vision_models::conformer::config::ConformerEncoderConfig,
};

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct Phi4MMImageEmbedConfig {
    pub n_embd: Option<usize>,
    pub crop_size: Option<usize>,
    pub embedding_cls: String,
    pub enable_gradient_checkpointing: bool,
    pub hd_transform_order: Option<String>,
    pub image_token_compression_cls: Option<String>,
    pub projection_cls: Option<String>,
    pub use_hd_transform: Option<bool>,
    pub with_learnable_separator: Option<bool>,
}

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct Phi4MMAudioEmbedConfig {
    pub n_embd: Option<usize>,
    pub compression_rate: usize,
    pub downsample_rate: usize,
    pub embedding_cls: String,
    pub projection_cls: String,
}

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct Phi4MMEmbdLayerConfig {
    pub image_embd_layer: Option<Phi4MMImageEmbedConfig>,
    pub audio_embd_layer: Option<Phi4MMAudioEmbedConfig>,
}

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct Phi4MMImgProcessorConfig {
    pub layer_idx: Option<isize>,
    pub type_feature: Option<String>,
}

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct Phi4MMAudioConfig {
    pub config: ConformerEncoderConfig,
    pub name: String,
}

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct Phi4MMConfig {
    pub vocab_size: usize,
    pub hidden_size: usize,
    pub intermediate_size: usize,
    pub num_hidden_layers: usize,
    pub num_attention_heads: usize,
    pub num_key_value_heads: Option<usize>,
    pub resid_pdrop: f64,
    pub embd_pdrop: f64,
    pub attention_dropout: f64,
    pub hidden_act: Activation,
    pub max_position_embeddings: usize,
    pub original_max_position_embeddings: usize,
    pub initializer_range: f64,
    pub rms_norm_eps: f64,
    pub use_cache: bool,
    pub tie_word_embeddings: bool,
    pub rope_theta: f64,
    pub rope_scaling: Option<Phi4MMRopeScalingConfig>,
    pub partial_rotary_factor: f64,
    pub bos_token_id: usize,
    pub eos_token_id: usize,
    pub pad_token_id: usize,
    pub image_input_id: Option<f64>,
    pub sliding_window: Option<usize>,
    pub embd_layer: Phi4MMEmbdLayerConfig,
    pub img_processor: Option<Phi4MMImgProcessorConfig>,
    pub audio_processor: Option<Phi4MMAudioConfig>,
    pub vision_lora: StaticLoraConfig,
    pub speech_lora: StaticLoraConfig,
    pub quantization_config: Option<QuantizedConfig>,
}

impl Phi4MMConfig {
    pub fn num_key_value_heads(&self) -> usize {
        self.num_key_value_heads.unwrap_or(self.num_attention_heads)
    }

    pub fn head_dim(&self) -> usize {
        self.hidden_size / self.num_attention_heads
    }

    pub fn loras(&self) -> HashMap<String, StaticLoraConfig> {
        let mut accum = HashMap::new();
        // Add all the loras
        // accum.insert("speech".to_string(), self.speech_lora.clone());
        accum.insert("vision".to_string(), self.vision_lora.clone());
        accum
    }
}