libmir-cuda 0.2.0

CUDA inference backend for libmir
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
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154

mod execution;
mod scratch;
#[cfg(all(test, target_os = "linux"))]
mod tests;
mod weights;

pub use execution::CudaAffineGatedDeltaExecution;
use models::layout::LinearAttentionConfig;
pub use weights::AffineGatedDeltaLayerWeights;

use super::{CudaGatedDeltaState, GatedDeltaStateConfig};
use crate::{CudaBackend, CudaTensorSet, Error, Result};

#[derive(Clone, Copy, Debug, PartialEq)]
pub struct AffineGatedDeltaLayerConfig {
    pub hidden_size: usize,
    pub key_heads: usize,
    pub value_heads: usize,
    pub key_dim: usize,
    pub value_dim: usize,
    pub convolution_kernel_size: usize,
    pub group_size: usize,
    pub bits: usize,
    pub rms_norm_epsilon: f32,
    pub norm_weight_shift: f32,
}

impl AffineGatedDeltaLayerConfig {
    pub fn from_linear_attention(
        hidden_size: usize,
        linear: &LinearAttentionConfig,
        group_size: usize,
        bits: usize,
        rms_norm_epsilon: f64,
        norm_weight_shift: f32,
    ) -> Result<Self> {
        let config = Self {
            hidden_size,
            key_heads: linear.key_heads,
            value_heads: linear.value_heads,
            key_dim: linear.key_head_dim,
            value_dim: linear.value_head_dim,
            convolution_kernel_size: linear.convolution_kernel_size,
            group_size,
            bits,
            rms_norm_epsilon: rms_norm_epsilon.to_string().parse()?,
            norm_weight_shift,
        };
        config.validate()?;
        Ok(config)
    }

    pub fn state(self) -> Result<GatedDeltaStateConfig> {
        self.validate()?;
        Ok(GatedDeltaStateConfig {
            key_heads: self.key_heads,
            value_heads: self.value_heads,
            key_dim: self.key_dim,
            value_dim: self.value_dim,
            convolution_kernel_size: self.convolution_kernel_size,
        })
    }

    pub(super) fn key_width(self) -> Result<usize> {
        checked(self.key_heads, self.key_dim)
    }

    pub(super) fn value_width(self) -> Result<usize> {
        checked(self.value_heads, self.value_dim)
    }

    pub(super) fn mixed_width(self) -> Result<usize> {
        let value_width = self.value_width()?;
        self.key_width()?
            .checked_mul(2)
            .and_then(|width| width.checked_add(value_width))
            .ok_or(Error::InvalidDecoderKernel("Gated Delta projection width overflow"))
    }

    fn validate(self) -> Result<()> {
        let state = GatedDeltaStateConfig {
            key_heads: self.key_heads,
            value_heads: self.value_heads,
            key_dim: self.key_dim,
            value_dim: self.value_dim,
            convolution_kernel_size: self.convolution_kernel_size,
        };
        if self.hidden_size == 0
            || self.group_size == 0
            || !matches!(self.bits, 4 | 8)
            || !self.hidden_size.is_multiple_of(self.group_size)
            || !self.value_width()?.is_multiple_of(self.group_size)
            || !self.rms_norm_epsilon.is_finite()
            || self.rms_norm_epsilon < 0.0
            || !self.norm_weight_shift.is_finite()
            || state.value_heads == 0
            || state.key_heads == 0
            || !state.value_heads.is_multiple_of(state.key_heads)
            || state.key_dim == 0
            || !state.key_dim.is_multiple_of(32)
            || state.value_dim == 0
            || state.convolution_kernel_size == 0
        {
            return Err(Error::InvalidDecoderKernel("invalid affine Gated Delta layer config"));
        }
        let _mixed = self.mixed_width()?;
        Ok(())
    }
}

#[derive(Clone, Debug)]
pub struct CudaAffineGatedDeltaLayer {
    backend: CudaBackend,
    config: AffineGatedDeltaLayerConfig,
    weights: AffineGatedDeltaLayerWeights,
}

impl CudaAffineGatedDeltaLayer {
    pub fn from_tensors(
        backend: &CudaBackend,
        tensors: &CudaTensorSet,
        prefix: &str,
        config: AffineGatedDeltaLayerConfig,
    ) -> Result<Self> {
        Self::new(backend, config, AffineGatedDeltaLayerWeights::load(tensors, prefix)?)
    }

    pub fn new(
        backend: &CudaBackend,
        config: AffineGatedDeltaLayerConfig,
        weights: AffineGatedDeltaLayerWeights,
    ) -> Result<Self> {
        config.validate()?;
        weights.validate(config)?;
        Ok(Self {
            backend: backend.clone(),
            config,
            weights,
        })
    }

    pub fn prepare(&self, tokens: usize) -> Result<CudaAffineGatedDeltaExecution> {
        CudaAffineGatedDeltaExecution::new(&self.backend, self.config, &self.weights, tokens)
    }

    pub fn prepare_state(&self) -> Result<CudaGatedDeltaState> {
        self.backend.prepare_gated_delta_state(self.config.state()?)
    }
}

pub(super) fn checked(left: usize, right: usize) -> Result<usize> {
    left.checked_mul(right)
        .ok_or(Error::InvalidDecoderKernel("Gated Delta shape overflow"))
}