hf2q 0.1.22

Pure Rust CLI for converting HuggingFace models to hardware-optimized formats and serving them over an OpenAI-compatible API on Apple Silicon
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

//! Stored-format embedding gather for Gemma input activations.

use crate::serve::forward_mlx_shared::MlxQWeight;
use anyhow::{bail, Result};
use mlx_native::{GgmlType, MlxBuffer, MlxDevice};

/// Gather stored embedding rows and apply Gemma's sqrt(hidden) scale.
#[allow(clippy::too_many_arguments)]
pub fn encode_embedding_rows(
    session: &mut mlx_native::GraphSession<'_>,
    registry: &mut mlx_native::KernelRegistry,
    device: &MlxDevice,
    weight: &MlxQWeight,
    token_ids: &MlxBuffer,
    output: &MlxBuffer,
    n_tokens: usize,
) -> Result<()> {
    let params = (weight.info.rows, weight.info.cols, n_tokens);
    match weight.info.ggml_dtype {
        GgmlType::F32 | GgmlType::F16 | GgmlType::BF16 => {
            mlx_native::ops::embedding_dense::embedding_gather_dense(
                session.encoder_mut(),
                registry,
                device,
                &weight.buffer,
                token_ids,
                output,
                &mlx_native::ops::embedding_dense::EmbeddingDenseParams {
                    vocab_size: params.0,
                    embed_dim: params.1,
                    n_tokens: params.2,
                },
            )?
        }
        GgmlType::Q4_0 => session.embedding_gather_q4_0(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::ops::embedding_q4_0::EmbeddingQ4_0Params {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        GgmlType::Q5_0 => session.embedding_gather_q5_0(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::EmbeddingQ5_0Params {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        GgmlType::Q2_K => session.embedding_gather_q2_k(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::ops::embedding_q2_k::EmbeddingQ2KParams {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        GgmlType::Q4_K => session.embedding_gather_q4_k(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::ops::embedding_q4_k::EmbeddingQ4KParams {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        GgmlType::Q5_K => session.embedding_gather_q5_k(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::ops::embedding_kquant::EmbeddingQ5KParams {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        GgmlType::Q6_K => session.embedding_gather_q6_k(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::ops::embedding_kquant::EmbeddingQ6KParams {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        GgmlType::Q8_0 => session.embedding_gather_q8_0(
            registry,
            device,
            &weight.buffer,
            token_ids,
            output,
            &mlx_native::ops::embedding_q8_0::EmbeddingQ8_0Params {
                vocab_size: params.0,
                embed_dim: params.1,
                n_tokens: params.2,
            },
        )?,
        other => bail!("Gemma embedding dispatch reached unadmitted type {other:?}"),
    }
    session.track_dispatch(&[&weight.buffer, token_ids], &[output]);
    Ok(())
}

/// Gather stored embedding rows and apply Gemma's model-input scale.
#[allow(clippy::too_many_arguments)]
pub fn encode_embedding(
    session: &mut mlx_native::GraphSession<'_>,
    registry: &mut mlx_native::KernelRegistry,
    device: &MlxDevice,
    weight: &MlxQWeight,
    token_ids: &MlxBuffer,
    output: &MlxBuffer,
    n_tokens: usize,
) -> Result<()> {
    encode_embedding_rows(
        session, registry, device, weight, token_ids, output, n_tokens,
    )?;
    session.barrier_between(&[output], &[output]);
    mlx_native::ops::elementwise::scalar_mul_f32(
        session.encoder_mut(),
        registry,
        device.metal_device(),
        output,
        output,
        n_tokens * weight.info.cols,
        (weight.info.cols as f32).sqrt(),
    )?;
    session.track_dispatch(&[output], &[output]);
    Ok(())
}