mlex 0.1.0

Safe, idiomatic Rust runtime for Apple MLX: quantized LLM inference (Qwen, Gemma4, NemotronH, DharaAR, ...) with multi-modal support
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213

//! `GatedDeltaNet`: the linear-attention layer used by Qwen3.5/3.6 in place
//! of quadratic self-attention on most decoder layers. Ports mlx-lm's
//! `gated_delta.py` + `Qwen3NextGatedDeltaNet`/`GatedDeltaNet` ops-based
//! reference path (no fused Metal kernel): a short causal depthwise
//! convolution feeds a per-timestep delta-rule recurrence.
//!
//! Batch size is assumed to be 1 (single-sequence generation), so the
//! `mask`-based padding handling present in the Python reference (used for
//! batched training) is omitted.

use crate::array::{Array, Dtype};
use crate::error::Result;
use crate::nn::{Linear, WeightMap};
use crate::ops;

use super::cache::GatedDeltaCache;

pub struct GatedDeltaNet {
    conv1d_weight: Array,
    in_proj_qkv: Linear,
    in_proj_z: Linear,
    in_proj_b: Linear,
    in_proj_a: Linear,
    dt_bias: Array,
    a_log: Array,
    norm_weight: Array,
    norm_eps: f32,
    out_proj: Linear,

    num_v_heads: i32,
    num_k_heads: i32,
    head_k_dim: i32,
    head_v_dim: i32,
    key_dim: i32,
    value_dim: i32,
    conv_dim: i32,
    conv_kernel_size: i32,
}

pub struct GatedDeltaConfig {
    pub num_v_heads: i32,
    pub num_k_heads: i32,
    pub head_k_dim: i32,
    pub head_v_dim: i32,
    pub conv_kernel_size: i32,
    pub rms_norm_eps: f32,
}

impl GatedDeltaNet {
    pub fn load(w: &mut WeightMap, prefix: &str, cfg: &GatedDeltaConfig) -> Result<Self> {
        let key_dim = cfg.head_k_dim * cfg.num_k_heads;
        let value_dim = cfg.head_v_dim * cfg.num_v_heads;
        let conv_dim = key_dim * 2 + value_dim;

        Ok(GatedDeltaNet {
            conv1d_weight: w.take(&format!("{prefix}.conv1d.weight"))?,
            in_proj_qkv: w.linear(&format!("{prefix}.in_proj_qkv"))?,
            in_proj_z: w.linear(&format!("{prefix}.in_proj_z"))?,
            in_proj_b: w.linear(&format!("{prefix}.in_proj_b"))?,
            in_proj_a: w.linear(&format!("{prefix}.in_proj_a"))?,
            dt_bias: w.take(&format!("{prefix}.dt_bias"))?,
            a_log: w.take(&format!("{prefix}.A_log"))?,
            norm_weight: w.take(&format!("{prefix}.norm.weight"))?,
            norm_eps: cfg.rms_norm_eps,
            out_proj: w.linear(&format!("{prefix}.out_proj"))?,
            num_v_heads: cfg.num_v_heads,
            num_k_heads: cfg.num_k_heads,
            head_k_dim: cfg.head_k_dim,
            head_v_dim: cfg.head_v_dim,
            key_dim,
            value_dim,
            conv_dim,
            conv_kernel_size: cfg.conv_kernel_size,
        })
    }

    pub fn forward(&self, inputs: &Array, cache: &mut GatedDeltaCache) -> Result<Array> {
        let shape = inputs.shape();
        let (b, s) = (shape[0], shape[1]);

        let qkv = self.in_proj_qkv.forward(inputs)?;
        let z = self.in_proj_z.forward(inputs)?;
        let z = ops::reshape(&z, &[b, s, self.num_v_heads, self.head_v_dim])?;
        let bt = self.in_proj_b.forward(inputs)?;
        let at = self.in_proj_a.forward(inputs)?;

        let n_keep = self.conv_kernel_size - 1;
        let conv_state = cache
            .conv_state
            .take()
            .unwrap_or(ops::zeros(&[b, n_keep, self.conv_dim], inputs.dtype())?);
        let conv_input = ops::concatenate(&[&conv_state, &qkv], 1)?;
        let total_len = conv_input.dim(1);
        cache.conv_state = Some(ops::contiguous(&ops::slice(
            &conv_input,
            &[0, total_len - n_keep, 0],
            &[b, total_len, self.conv_dim],
        )?)?);

        let conv_out = ops::silu(&ops::conv1d(
            &conv_input,
            &self.conv1d_weight,
            1,
            0,
            1,
            self.conv_dim,
        )?)?;

        let parts = ops::split_sections(&conv_out, &[self.key_dim, 2 * self.key_dim], -1)?;
        let q = ops::reshape(&parts[0], &[b, s, self.num_k_heads, self.head_k_dim])?;
        let k = ops::reshape(&parts[1], &[b, s, self.num_k_heads, self.head_k_dim])?;
        let v = ops::reshape(&parts[2], &[b, s, self.num_v_heads, self.head_v_dim])?;

        let inv_scale = (self.head_k_dim as f32).powf(-0.5);
        let q = ops::scale_by(&ops::rms_norm(&q, None, 1e-6)?, inv_scale * inv_scale)?;
        let k = ops::scale_by(&ops::rms_norm(&k, None, 1e-6)?, inv_scale)?;

        let repeat_factor = self.num_v_heads / self.num_k_heads;
        let (q, k) = if repeat_factor > 1 {
            (
                ops::repeat_axis(&q, repeat_factor, 2)?,
                ops::repeat_axis(&k, repeat_factor, 2)?,
            )
        } else {
            (q, k)
        };

        // Recurrence runs in f32 for numerical stability, matching the
        // Python reference (`state` is created as float32).
        let q = ops::astype(&q, Dtype::Float32)?;
        let k = ops::astype(&k, Dtype::Float32)?;
        let v = ops::astype(&v, Dtype::Float32)?;
        let beta = ops::sigmoid(&ops::astype(&bt, Dtype::Float32)?)?;
        let a32 = ops::astype(&at, Dtype::Float32)?;
        let a_log32 = ops::astype(&self.a_log, Dtype::Float32)?;
        let dt_bias32 = ops::astype(&self.dt_bias, Dtype::Float32)?;
        // g = exp(-exp(A_log) * softplus(a + dt_bias)), shape [B, S, Hv].
        let g = {
            let inner = ops::softplus(&ops::add(&a32, &dt_bias32)?)?;
            let decay_rate = ops::exp(&a_log32)?;
            ops::exp(&ops::negative(&ops::multiply(&decay_rate, &inner)?)?)?
        };

        let mut state = cache.recur_state.take().unwrap_or(ops::zeros(
            &[b, self.num_v_heads, self.head_v_dim, self.head_k_dim],
            Dtype::Float32,
        )?);

        let mut ys = Vec::with_capacity(s as usize);
        for t in 0..s {
            let qt = ops::reshape(
                &ops::slice(
                    &q,
                    &[0, t, 0, 0],
                    &[b, t + 1, self.num_v_heads, self.head_k_dim],
                )?,
                &[b, self.num_v_heads, self.head_k_dim],
            )?;
            let kt = ops::reshape(
                &ops::slice(
                    &k,
                    &[0, t, 0, 0],
                    &[b, t + 1, self.num_v_heads, self.head_k_dim],
                )?,
                &[b, self.num_v_heads, self.head_k_dim],
            )?;
            let vt = ops::reshape(
                &ops::slice(
                    &v,
                    &[0, t, 0, 0],
                    &[b, t + 1, self.num_v_heads, self.head_v_dim],
                )?,
                &[b, self.num_v_heads, self.head_v_dim],
            )?;
            let gt = ops::reshape(
                &ops::slice(&g, &[0, t, 0], &[b, t + 1, self.num_v_heads])?,
                &[b, self.num_v_heads],
            )?;
            let betat = ops::reshape(
                &ops::slice(&beta, &[0, t, 0], &[b, t + 1, self.num_v_heads])?,
                &[b, self.num_v_heads],
            )?;

            let decay = ops::reshape(&gt, &[b, self.num_v_heads, 1, 1])?;
            state = ops::multiply(&state, &decay)?;
            let kt_row = ops::expand_dims(&kt, 2)?; // [B, Hv, 1, Dk]
            let kv_mem = ops::sum_axes(&ops::multiply(&state, &kt_row)?, &[-1], false)?; // [B, Hv, Dv]
            let delta = ops::multiply(
                &ops::subtract(&vt, &kv_mem)?,
                &ops::expand_dims(&betat, -1)?,
            )?; // [B,Hv,Dv]
            let update = ops::multiply(&kt_row, &ops::expand_dims(&delta, -1)?)?; // [B,Hv,Dv,Dk]
            state = ops::add(&state, &update)?;
            let qt_row = ops::expand_dims(&qt, 2)?; // [B, Hv, 1, Dk]
            let yt = ops::sum_axes(&ops::multiply(&state, &qt_row)?, &[-1], false)?; // [B, Hv, Dv]
            ys.push(ops::astype(&yt, inputs.dtype())?);
        }
        cache.recur_state = Some(state);

        let y_refs: Vec<&Array> = ys.iter().collect();
        let y = ops::stack_axis(&y_refs, 1)?; // [B, S, Hv, Dv]

        // RMSNormGated: normalize then gate with silu(z), all done in f32.
        let y32 = ops::astype(&y, Dtype::Float32)?;
        let normed = ops::rms_norm(&y32, None, self.norm_eps)?;
        let normed = ops::multiply(&normed, &ops::astype(&self.norm_weight, Dtype::Float32)?)?;
        let gated = ops::multiply(&ops::silu(&ops::astype(&z, Dtype::Float32)?)?, &normed)?;
        let out = ops::astype(&gated, inputs.dtype())?;
        let out = ops::reshape(&out, &[b, s, self.value_dim])?;

        self.out_proj.forward(&out)
    }
}