mlex 0.1.3

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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231

//! Mamba2 SSM mixer used by NemotronH's hybrid layers. Ports mlx-lm's
//! `NemotronHMamba2Mixer` + `ssm.py` reference math as a straightforward
//! per-timestep recurrence (mathematically equivalent to the chunked
//! "SSD" formulation `ssm_attn` uses for training-time parallelism, and to
//! the fused Metal `ssm_kernel` used for single-token decode) rather than
//! porting the chunked matrix algorithm — same pattern as `gated_delta.rs`.
//!
//! Batch size is assumed to be 1 (single-sequence generation).

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

use super::cache::GatedDeltaCache;

pub struct Mamba2Config {
    pub num_heads: i32,
    pub head_dim: i32,
    pub n_groups: i32,
    pub state_size: i32,
    pub conv_kernel: i32,
    pub proj_bias: bool,
    pub conv_bias: bool,
    pub norm_eps: f32,
    pub time_step_min: f32,
    pub time_step_max: f32,
}

pub struct Mamba2Mixer {
    conv1d_weight: Array,
    conv1d_bias: Option<Array>,
    in_proj: Linear,
    dt_bias: Array,
    a_log: Array,
    d: Array,
    norm_weight: Array,
    norm_eps: f32,
    out_proj: Linear,

    num_heads: i32,
    head_dim: i32,
    n_groups: i32,
    state_size: i32,
    conv_kernel: i32,
    intermediate_size: i32,
    conv_dim: i32,
    time_step_min: f32,
    time_step_max: f32,
}

impl Mamba2Mixer {
    pub fn load(w: &mut WeightMap, prefix: &str, cfg: &Mamba2Config) -> Result<Self> {
        let intermediate_size = cfg.num_heads * cfg.head_dim;
        let conv_dim = intermediate_size + 2 * cfg.n_groups * cfg.state_size;

        let conv1d_bias = if cfg.conv_bias {
            Some(w.take(&format!("{prefix}.conv1d.bias"))?)
        } else {
            None
        };

        Ok(Mamba2Mixer {
            conv1d_weight: w.take(&format!("{prefix}.conv1d.weight"))?,
            conv1d_bias,
            in_proj: w.linear(&format!("{prefix}.in_proj"))?,
            dt_bias: w.take(&format!("{prefix}.dt_bias"))?,
            a_log: w.take(&format!("{prefix}.A_log"))?,
            d: w.take(&format!("{prefix}.D"))?,
            norm_weight: w.take(&format!("{prefix}.norm.weight"))?,
            norm_eps: cfg.norm_eps,
            out_proj: w.linear(&format!("{prefix}.out_proj"))?,
            num_heads: cfg.num_heads,
            head_dim: cfg.head_dim,
            n_groups: cfg.n_groups,
            state_size: cfg.state_size,
            conv_kernel: cfg.conv_kernel,
            intermediate_size,
            conv_dim,
            time_step_min: cfg.time_step_min,
            time_step_max: cfg.time_step_max,
        })
    }

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

        let projected = self.in_proj.forward(hidden_states)?;
        let parts = ops::split_sections(
            &projected,
            &[
                self.intermediate_size,
                self.intermediate_size + self.conv_dim,
            ],
            -1,
        )?;
        let (gate, conv_input, dt_raw) = (&parts[0], &parts[1], &parts[2]);

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

        let mut conv_out = ops::conv1d(&padded, &self.conv1d_weight, 1, 0, 1, self.conv_dim)?;
        if let Some(bias) = &self.conv1d_bias {
            conv_out = ops::add(&conv_out, bias)?;
        }
        let conv_out = ops::silu(&conv_out)?;

        let gs = self.n_groups * self.state_size;
        let parts = ops::split_sections(
            &conv_out,
            &[self.intermediate_size, self.intermediate_size + gs],
            -1,
        )?;
        let (x, bmat, cmat) = (&parts[0], &parts[1], &parts[2]);

        let x = ops::reshape(x, &[b, s, self.num_heads, self.head_dim])?;
        let bmat = ops::reshape(bmat, &[b, s, self.n_groups, self.state_size])?;
        let cmat = ops::reshape(cmat, &[b, s, self.n_groups, self.state_size])?;

        // f32 recurrence for numerical stability, matching the Python
        // reference's fused-kernel accumulation precision.
        let x = ops::astype(&x, Dtype::Float32)?;
        let bmat = ops::astype(&bmat, Dtype::Float32)?;
        let cmat = ops::astype(&cmat, Dtype::Float32)?;
        let dt_raw = ops::astype(dt_raw, Dtype::Float32)?;
        let dt_bias32 = ops::astype(&self.dt_bias, Dtype::Float32)?;
        let a_log32 = ops::astype(&self.a_log, Dtype::Float32)?;
        let d32 = ops::astype(&self.d, Dtype::Float32)?;

        // dt = clip(softplus(dt_raw + dt_bias), min, max), shape [B, S, H].
        let dt = {
            let sp = ops::softplus(&ops::add(&dt_raw, &dt_bias32)?)?;
            let lo = ops::maximum(&sp, &Array::scalar_f32(self.time_step_min))?;
            ops::minimum(&lo, &Array::scalar_f32(self.time_step_max))?
        };
        let neg_a = ops::negative(&ops::exp(&a_log32)?)?; // [H]

        let repeat_factor = self.num_heads / self.n_groups;
        let (bmat, cmat) = if repeat_factor > 1 {
            (
                ops::repeat_axis(&bmat, repeat_factor, 2)?,
                ops::repeat_axis(&cmat, repeat_factor, 2)?,
            )
        } else {
            (bmat, cmat)
        };

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

        let mut ys = Vec::with_capacity(s as usize);
        for t in 0..s {
            let xt = ops::reshape(
                &ops::slice(
                    &x,
                    &[0, t, 0, 0],
                    &[b, t + 1, self.num_heads, self.head_dim],
                )?,
                &[b, self.num_heads, self.head_dim],
            )?;
            let bt = ops::reshape(
                &ops::slice(
                    &bmat,
                    &[0, t, 0, 0],
                    &[b, t + 1, self.num_heads, self.state_size],
                )?,
                &[b, self.num_heads, self.state_size],
            )?;
            let ct = ops::reshape(
                &ops::slice(
                    &cmat,
                    &[0, t, 0, 0],
                    &[b, t + 1, self.num_heads, self.state_size],
                )?,
                &[b, self.num_heads, self.state_size],
            )?;
            let dtt = ops::reshape(
                &ops::slice(&dt, &[0, t, 0], &[b, t + 1, self.num_heads])?,
                &[b, self.num_heads],
            )?;

            let da = ops::exp(&ops::multiply(&dtt, &neg_a)?)?; // [B, H]
            let da = ops::reshape(&da, &[b, self.num_heads, 1, 1])?;

            let dbx = {
                let x_col = ops::expand_dims(&xt, -1)?; // [B,H,Dh,1]
                let dt_col = ops::reshape(&dtt, &[b, self.num_heads, 1, 1])?;
                let b_row = ops::expand_dims(&bt, 2)?; // [B,H,1,Ds]
                ops::multiply(&ops::multiply(&x_col, &dt_col)?, &b_row)?
            };
            state = ops::add(&ops::multiply(&state, &da)?, &dbx)?;

            let c_row = ops::expand_dims(&ct, 2)?; // [B,H,1,Ds]
            let y_ssm = ops::sum_axes(&ops::multiply(&state, &c_row)?, &[-1], false)?; // [B,H,Dh]
            let d_row = ops::reshape(&d32, &[1, self.num_heads, 1])?;
            let yt = ops::add(&y_ssm, &ops::multiply(&xt, &d_row)?)?;
            ys.push(yt);
        }
        cache.recur_state = Some(state);

        let y_refs: Vec<&Array> = ys.iter().collect();
        let y = ops::stack_axis(&y_refs, 1)?; // [B, S, H, Dh]
        let y = ops::reshape(&y, &[b, s, self.intermediate_size])?;

        // MambaRMSNormGated: swiglu(gate, y) then grouped RMSNorm.
        let gate32 = ops::astype(gate, Dtype::Float32)?;
        let gated = ops::multiply(&ops::silu(&gate32)?, &y)?;
        let group_size = self.intermediate_size / self.n_groups;
        let n_groups_flat = self.intermediate_size / group_size;
        let grouped = ops::reshape(&gated, &[b, s, n_groups_flat, group_size])?;
        let normed = ops::rms_norm(&grouped, None, self.norm_eps)?;
        let normed = ops::reshape(&normed, &[b, s, self.intermediate_size])?;
        let normed = ops::multiply(&normed, &ops::astype(&self.norm_weight, Dtype::Float32)?)?;
        let out = ops::astype(&normed, hidden_states.dtype())?;

        self.out_proj.forward(&out)
    }
}