mamba-rs 0.5.1

Mamba SSM and Mamba-3 SISO in Rust with optional CUDA GPU acceleration. Inference and training (BPTT through SSM state, AdamW), CPU + GPU paths, custom CUDA kernels, CUDA Graph capture, f32 / bf16 / f16. Opt-in deterministic training (bit-identical runs, batch-invariant inference) with a tensor-core tier that beats cuBLAS on LLM-sized models.
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
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

//! Safetensors serialization for Mamba-3 SISO weights.
//!
//! Tensor naming convention:
//! - `input_proj.weight`, `input_proj.bias`
//! - `layers.{i}.norm.weight`
//! - `layers.{i}.in_proj.weight`
//! - `layers.{i}.dt_bias`
//! - `layers.{i}.b_norm.weight`, `layers.{i}.c_norm.weight`
//! - `layers.{i}.b_bias`, `layers.{i}.c_bias`
//! - `layers.{i}.D`
//! - `layers.{i}.norm_gate.weight`
//! - `layers.{i}.out_proj.weight`
//! - `norm_f.weight`

use super::config::Mamba3Config;
use super::weights::{Mamba3LayerWeights, Mamba3Weights};
use safetensors::serialize_to_file;
use safetensors::tensor::{SafeTensors, TensorView};
use std::collections::HashMap;
use std::path::Path;

fn view<'a>(data: &'a [f32], shape: &[usize]) -> TensorView<'a> {
    let bytes = unsafe { std::slice::from_raw_parts(data.as_ptr() as *const u8, data.len() * 4) };
    TensorView::new(safetensors::Dtype::F32, shape.to_vec(), bytes).unwrap()
}

/// Save Mamba-3 SISO weights to safetensors.
pub fn save_mamba3(
    path: &Path,
    weights: &Mamba3Weights,
    cfg: &Mamba3Config,
    input_dim: usize,
) -> Result<(), String> {
    let dm = cfg.d_model;
    let di = cfg.d_inner();
    let ds = cfg.d_state;
    let nh = cfg.nheads();
    let ip = cfg.in_proj_out_dim();

    macro_rules! t {
        ($name:expr, $data:expr, $shape:expr) => {
            ($name.to_string(), view($data, $shape))
        };
    }

    let mut tensors: Vec<(String, TensorView<'_>)> = vec![
        t!("input_proj.weight", &weights.input_proj_w, &[input_dim, dm]),
        t!("input_proj.bias", &weights.input_proj_b, &[dm]),
    ];

    for (i, lw) in weights.layers.iter().enumerate() {
        let p = |n: &str| format!("layers.{i}.{n}");
        tensors.extend([
            t!(p("norm.weight"), &lw.norm_weight, &[dm]),
            t!(p("in_proj.weight"), &lw.in_proj_w, &[dm, ip]),
            t!(p("dt_bias"), &lw.dt_bias, &[nh]),
            t!(p("b_norm.weight"), &lw.b_norm_weight, &[ds]),
            t!(p("c_norm.weight"), &lw.c_norm_weight, &[ds]),
            t!(p("b_bias"), &lw.b_bias, &[nh, ds]),
            t!(p("c_bias"), &lw.c_bias, &[nh, ds]),
            t!(p("D"), &lw.d_param, &[nh]),
            t!(p("norm_gate.weight"), &lw.norm_gate_weight, &[di]),
            t!(p("out_proj.weight"), &lw.out_proj_w, &[di, dm]),
        ]);
    }

    tensors.push(t!("norm_f.weight", &weights.norm_f_weight, &[dm]));

    let meta = HashMap::from([
        ("d_model".to_string(), dm.to_string()),
        ("d_state".to_string(), ds.to_string()),
        ("expand".to_string(), cfg.expand.to_string()),
        ("headdim".to_string(), cfg.headdim.to_string()),
        ("ngroups".to_string(), cfg.ngroups.to_string()),
        ("n_layers".to_string(), cfg.n_layers.to_string()),
        ("input_dim".to_string(), input_dim.to_string()),
        ("rope_fraction".to_string(), cfg.rope_fraction.to_string()),
        ("a_floor".to_string(), cfg.a_floor.to_string()),
        ("mamba_version".to_string(), "3".to_string()),
    ]);

    let refs: Vec<_> = tensors
        .iter()
        .map(|(n, t)| (n.as_str(), t.clone()))
        .collect();
    serialize_to_file(refs, Some(meta), path).map_err(|e| format!("safetensors save failed: {e:?}"))
}

/// Load Mamba-3 SISO weights from safetensors.
pub fn load_mamba3(path: &Path, cfg: &Mamba3Config) -> Result<(Mamba3Weights, usize), String> {
    let data = std::fs::read(path).map_err(|e| format!("read failed: {e}"))?;
    let (_, header) = SafeTensors::read_metadata(&data).map_err(|e| format!("metadata: {e}"))?;
    let meta_map = header.metadata().as_ref().ok_or("no metadata")?;
    let input_dim: usize = meta_map
        .get("input_dim")
        .ok_or("missing input_dim")?
        .parse()
        .map_err(|e| format!("parse input_dim: {e}"))?;

    let st = SafeTensors::deserialize(&data).map_err(|e| format!("deserialize: {e:?}"))?;

    // Byte-wise decode: the safetensors spec does not guarantee 4-byte
    // alignment of the data section, so a `*const f32` cast on the raw
    // buffer would be UB for a spec-valid file with an unpadded header.
    let get = |name: &str| -> Result<Vec<f32>, String> {
        let t = st
            .tensor(name)
            .map_err(|e| format!("tensor '{name}': {e:?}"))?;
        if t.dtype() != safetensors::Dtype::F32 {
            return Err(format!(
                "tensor '{name}': dtype {:?}, expected F32",
                t.dtype()
            ));
        }
        let bytes = t.data();
        if bytes.len() % 4 != 0 {
            return Err(format!(
                "tensor '{name}': byte length {} is not a multiple of 4",
                bytes.len()
            ));
        }
        Ok(bytes
            .chunks_exact(4)
            .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
            .collect())
    };

    let mut layers = Vec::new();
    for i in 0..cfg.n_layers {
        let p = |n: &str| format!("layers.{i}.{n}");
        layers.push(Mamba3LayerWeights {
            norm_weight: get(&p("norm.weight"))?,
            in_proj_w: get(&p("in_proj.weight"))?,
            dt_bias: get(&p("dt_bias"))?,
            b_norm_weight: get(&p("b_norm.weight"))?,
            c_norm_weight: get(&p("c_norm.weight"))?,
            b_bias: get(&p("b_bias"))?,
            c_bias: get(&p("c_bias"))?,
            d_param: get(&p("D"))?,
            norm_gate_weight: get(&p("norm_gate.weight"))?,
            out_proj_w: get(&p("out_proj.weight"))?,
        });
    }

    let w = Mamba3Weights {
        input_proj_w: get("input_proj.weight")?,
        input_proj_b: get("input_proj.bias")?,
        layers,
        norm_f_weight: get("norm_f.weight")?,
    };

    w.validate(cfg, input_dim)?;
    Ok((w, input_dim))
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_serialize_roundtrip() {
        let cfg = Mamba3Config::default();
        let w = Mamba3Weights::init(&cfg, 128, 42);
        let tmp = std::env::temp_dir().join("mamba3_test.safetensors");
        save_mamba3(&tmp, &w, &cfg, 128).unwrap();
        let (w2, input_dim) = load_mamba3(&tmp, &cfg).unwrap();
        assert_eq!(input_dim, 128);
        assert_eq!(w.layers[0].dt_bias, w2.layers[0].dt_bias);
        assert_eq!(w.layers[0].in_proj_w, w2.layers[0].in_proj_w);
        assert_eq!(w.norm_f_weight, w2.norm_f_weight);
        std::fs::remove_file(tmp).ok();
    }
}