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
// SPDX-License-Identifier: MIT OR Apache-2.0
//! Gemma4 multimodal embedder — lazy port.
//!
//! Projects encoder features (from the vision tower or audio tower)
//! into the language model's embedding space. Two steps:
//!
//! 1. **Pre-projection RMS norm** with NO learnable scale —
//! `x / sqrt(mean(x²) + eps)`. Just the last-dim RMS divider
//! without a gain multiplication.
//! 2. **Linear projection** (no bias) from `multimodal_hidden_size`
//! to `text_hidden_size`.
//!
//! Used by `Model` in the eager `gemma4::Model` to project both
//! the vision tower and (optionally) the audio tower outputs into
//! the Gemma4 text decoder's embedding space before concatenating
//! them with text token embeddings.
use fuel_ir::error::Result;
use fuel_tensor::{lazy::Tensor, lazy::WeightStorage};
#[cfg(test)]
use std::sync::Arc;
#[derive(Debug, Clone, PartialEq)]
pub struct Gemma4MmEmbedConfig {
pub multimodal_hidden_size: usize,
pub text_hidden_size: usize,
pub eps: f64,
}
#[derive(Debug, Clone)]
pub struct Gemma4MmEmbedWeights {
/// `[multimodal_hidden_size, text_hidden_size]` (no bias).
pub projection: WeightStorage,
}
#[derive(Debug, Clone)]
pub struct Gemma4MmEmbedder {
pub config: Gemma4MmEmbedConfig,
pub weights: Gemma4MmEmbedWeights,
}
impl Gemma4MmEmbedder {
/// Normalise then project soft encoder features into text
/// embedding space.
///
/// `soft_features` shape: `(..., multimodal_hidden_size)`.
/// Returns shape `(..., text_hidden_size)`.
pub fn forward(&self, soft_features: &Tensor) -> Result<Tensor> {
let cfg = &self.config;
// GAP-314 (a1): the `dims.last() == multimodal_hidden_size` check is a panic
// in front of two downstream typed build-time Errs — `rms_norm_last_dim`
// rejects an empty/zero last dim, and `apply_linear` rejects a trailing dim
// that is not `multimodal_hidden_size`. Both conjuncts are covered, so the
// assert only preempted a typed Err; deleting it lets the typed Err surface.
// Step 1: RMS normalize over the last dim (no learnable gain).
let normed = soft_features.rms_norm_last_dim(cfg.eps)?;
// Step 2: Linear projection (no bias).
self.weights.projection.apply_linear(
&normed,
cfg.multimodal_hidden_size,
cfg.text_hidden_size,
)
}
}
// ---- HuggingFace safetensors loader ----------------------------------------
impl Gemma4MmEmbedWeights {
/// Load Gemma4's multimodal embedder projection from HF safetensors.
/// Convention: `model.embed_vision.embedding_projection.weight`
/// `[text_hidden, multimodal_hidden]` (HF stores `[out, in]`).
pub fn load_from_mmapped(
st: &fuel_loaders::safetensors::MmapedSafetensors,
cfg: &Gemma4MmEmbedConfig,
) -> fuel_ir::error::Result<Self> {
use fuel_tensor::lazy::load_transposed_matrix_preserve_dtype as ltm;
let projection = ltm(
st,
"model.embed_vision.embedding_projection.weight",
cfg.text_hidden_size,
cfg.multimodal_hidden_size,
)?;
Ok(Self { projection })
}
}
// ---- Tests ---------------------------------------------------------------
#[cfg(test)]
mod tests {
use super::*;
use fuel_ir::Shape;
use fuel_tensor::Device;
fn rng(seed: u32) -> impl FnMut() -> f32 {
let mut s = seed;
move || {
s = s.wrapping_mul(1103515245).wrapping_add(12345);
((s >> 16) as u16 as f32 / 65535.0 - 0.5) * 0.02
}
}
#[test]
fn mm_embed_forward_shape_and_finite() {
let cfg = Gemma4MmEmbedConfig {
multimodal_hidden_size: 8,
text_hidden_size: 12,
eps: 1e-6,
};
let mut next = rng(42);
let proj: Vec<f32> = (0..cfg.multimodal_hidden_size * cfg.text_hidden_size)
.map(|_| next())
.collect();
let model = Gemma4MmEmbedder {
config: cfg.clone(),
weights: Gemma4MmEmbedWeights {
projection: WeightStorage::F32(Arc::from(proj)),
},
};
// (1, seq, multimodal_hidden) input.
let seq = 5;
let input_data: Vec<f32> = (0..1 * seq * cfg.multimodal_hidden_size)
.map(|i| ((i as f32) * 0.05) - 0.1)
.collect();
let input = Tensor::from_f32(
input_data,
Shape::from_dims(&[1, seq, cfg.multimodal_hidden_size]),
&Device::cpu(),
)
.unwrap();
let out = model.forward(&input).unwrap();
assert_eq!(out.shape().dims(), &[1, seq, cfg.text_hidden_size]);
for &v in &out.realize_f32() {
assert!(v.is_finite(), "non-finite mm_embed output: {v}");
}
}
#[test]
fn mm_embed_rms_norm_makes_input_unit_rms() {
let cfg = Gemma4MmEmbedConfig {
multimodal_hidden_size: 4,
text_hidden_size: 4,
eps: 1e-12,
};
// Identity projection so we can read the post-norm value directly.
let mut identity = vec![0.0_f32; 16];
for i in 0..4 {
identity[i * 4 + i] = 1.0;
}
let model = Gemma4MmEmbedder {
config: cfg.clone(),
weights: Gemma4MmEmbedWeights {
projection: WeightStorage::F32(Arc::from(identity)),
},
};
// Input: each row is (1, 2, 3, 4) — RMS = sqrt((1+4+9+16)/4) = sqrt(7.5).
let input = Tensor::from_f32(
vec![1.0_f32, 2.0, 3.0, 4.0],
Shape::from_dims(&[1, 1, 4]),
&Device::cpu(),
)
.unwrap();
let out = model.forward(&input).unwrap().realize_f32();
let expected = [
1.0_f32 / 7.5_f32.sqrt(),
2.0 / 7.5_f32.sqrt(),
3.0 / 7.5_f32.sqrt(),
4.0 / 7.5_f32.sqrt(),
];
for (i, (&got, &want)) in out.iter().zip(expected.iter()).enumerate() {
assert!(
(got - want).abs() < 1e-5,
"row[{i}]: got {got} expected {want}"
);
}
}
}