emelex 1.1.1

Apple Silicon local inference toolkit powered by MLX
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

//! Gemma4 "unified" encoder-free multimodal path
//! (`vision_config.model_type == "gemma4_unified_vision"` /
//! `audio_config.model_type == "gemma4_unified_audio"`).
//!
//! Distinct from the SigLIP-style tower in [`super::vision`] and the
//! Conformer-style tower in [`super::audio`]: there is no transformer
//! encoder at all. Raw pixel patches (and, for audio, raw PCM windows) are
//! projected straight into the text embedding space by a couple of small
//! linear layers, then fused through the same `embed_vision`/`embed_audio`
//! projection (`vision::MultimodalEmbedder`) every Gemma4 checkpoint uses.
//! This keeps quantized checkpoints small (no encoder weights to ship) at
//! the cost of the encoder's spatial/temporal mixing.
//!
//! Weight key layout (matching the checkpoint verbatim, no renaming; these
//! live in a sidecar `optiq_vision.safetensors` shard for OptiQ
//! checkpoints, transparently merged into the same weight map by
//! `crate::engine::weights::load_all`):
//!   `vision_embedder.{patch_ln1,patch_dense,patch_ln2,pos_norm}.{weight,bias}`
//!   `vision_embedder.pos_embedding`
//!   `embed_vision.embedding_projection.*`,
//! `embed_audio.embedding_projection.*`

use serde_json::Value;

use super::super::config::{get_f32, get_i32};
use crate::engine::{
	array::Array,
	error::Result,
	nn::{LayerNorm, Linear, WeightMap},
	ops,
};

/// Parsed `vision_config` sub-dict of a `gemma4_unified` checkpoint.
#[derive(Debug, Clone)]
pub struct UnifiedVisionConfig {
	/// Pixel-grid patch size used by the resize math (16).
	pub patch_size: i32,
	/// Pooling kernel size (3); `model_patch_size = patch_size *
	/// pooling_kernel_size` is the raw pixel block side length fed
	/// directly into the patch embedder (no separate pooling step - the
	/// "pooling" is baked into patchifying at the larger block size).
	pub pooling_kernel_size: i32,
	/// Embedding width inside the vision embedder (== text hidden_size).
	pub mm_embed_dim: i32,
	/// Number of rows in the 2D positional-embedding table.
	pub mm_posemb_size: i32,
	/// Maximum soft tokens (patches) per image after resize.
	pub num_soft_tokens: i32,
	/// Epsilon for the embedder LayerNorms.
	pub rms_norm_eps: f32,
}

impl UnifiedVisionConfig {
	pub fn from_json(cfg: &Value) -> Result<Self> {
		Ok(UnifiedVisionConfig {
			patch_size: get_i32(cfg, "patch_size", 16)?,
			pooling_kernel_size: get_i32(cfg, "pooling_kernel_size", 3)?,
			mm_embed_dim: get_i32(cfg, "mm_embed_dim", 3840)?,
			mm_posemb_size: get_i32(cfg, "mm_posemb_size", 1120)?,
			num_soft_tokens: get_i32(cfg, "num_soft_tokens", 280)?,
			rms_norm_eps: get_f32(cfg, "rms_norm_eps", 1e-6)?,
		})
	}

	/// Raw pixel block side length fed into the patch embedder.
	pub fn model_patch_size(&self) -> i32 {
		self.patch_size * self.pooling_kernel_size
	}
}

/// Encoder-free vision embedder: `patch_ln1 -> patch_dense -> patch_ln2 ->
/// + 2D positional embedding -> pos_norm`. Patches are raw (unnormalized
/// beyond the `[0, 1]` rescale already done by
/// `crate::engine::media::image::preprocess_image_bytes`) `model_patch_size x
/// model_patch_size` pixel blocks, flattened per-patch in `(row, col,
/// channel)` order.
pub struct UnifiedVisionEmbedder {
	model_patch_size: i32,
	pooling_kernel_size: i32,
	patch_ln1: LayerNorm,
	patch_dense: Linear,
	patch_ln2: LayerNorm,
	/// `[mm_posemb_size, 2, mm_embed_dim]`.
	pos_embedding: Array,
	pos_norm: LayerNorm,
}

impl UnifiedVisionEmbedder {
	pub fn load(w: &mut WeightMap, cfg: &UnifiedVisionConfig) -> Result<Self> {
		let prefix = "vision_embedder";
		let eps = cfg.rms_norm_eps;
		Ok(UnifiedVisionEmbedder {
			model_patch_size: cfg.model_patch_size(),
			pooling_kernel_size: cfg.pooling_kernel_size,
			patch_ln1: w.layer_norm(&format!("{prefix}.patch_ln1"), eps)?,
			patch_dense: w.linear(&format!("{prefix}.patch_dense"))?,
			patch_ln2: w.layer_norm(&format!("{prefix}.patch_ln2"), eps)?,
			pos_embedding: w.take(&format!("{prefix}.pos_embedding"))?,
			pos_norm: w.layer_norm(&format!("{prefix}.pos_norm"), eps)?,
		})
	}

	/// `pixel_values`: `[1, 3, H, W]` (channel-first, `[0, 1]`-rescaled,
	/// `H`/`W` divisible by `model_patch_size` - guaranteed by
	/// `crate::engine::media::image::preprocess_image_bytes`'s alignment).
	/// `patch_h`/`patch_w` are the caller's `patch_size`-granularity grid
	/// (`H / patch_size`, `W / patch_size`); this divides them down by
	/// `pooling_kernel_size` to get the `model_patch_size`-granularity
	/// grid the raw patchify + position ids use.
	///
	/// Returns `[1, n, mm_embed_dim]`, `n = (patch_h / k) * (patch_w / k)`.
	pub fn forward(&self, pixel_values: &Array, patch_h: i32, patch_w: i32) -> Result<Array> {
		let k = self.pooling_kernel_size;
		let (grid_h, grid_w) = (patch_h / k, patch_w / k);
		let patches = self.patchify(pixel_values, grid_h, grid_w)?;
		let positions = build_positions(grid_h, grid_w)?;

		let hidden = self.patch_ln1.forward(&patches)?;
		let hidden = self.patch_dense.forward(&hidden)?;
		let hidden = self.patch_ln2.forward(&hidden)?;
		let hidden = self.add_position_embeddings(&hidden, &positions)?;
		let hidden = self.pos_norm.forward(&hidden)?;
		ops::expand_dims(&hidden, 0)
	}

	/// `[1, 3, H, W] -> [grid_h*grid_w, model_patch_size^2 * 3]`, patch
	/// order row-major over `(grid_h, grid_w)`, flatten order `(py, px,
	/// c)` within a patch - mirrors `super::vision::PatchEmbedder::patchify`
	/// but at `model_patch_size` granularity and with no `2*(x-0.5)`
	/// rescale (the raw `[0, 1]` values feed straight into `patch_ln1`,
	/// which is a learned affine LayerNorm and normalizes them itself).
	fn patchify(&self, pixel_values: &Array, grid_h: i32, grid_w: i32) -> Result<Array> {
		let shape = pixel_values.shape();
		let c = shape[1];
		let p = self.model_patch_size;

		let patches = ops::reshape(pixel_values, &[1, c, grid_h, p, grid_w, p])?;
		let patches = ops::transpose_axes(&patches, &[0, 2, 4, 3, 5, 1])?;
		let patches = ops::reshape(&patches, &[grid_h * grid_w, p * p * c])?;

		let target_dtype = self.patch_dense.weight_dtype(patches.dtype());
		ops::astype(&patches, target_dtype)
	}

	/// `hidden[i] += pos_embedding[x_i, 0] + pos_embedding[y_i, 1]`.
	fn add_position_embeddings(&self, hidden: &Array, positions: &Array) -> Result<Array> {
		let dtype = hidden.dtype();
		let n = positions.dim(0);
		let d = self.pos_embedding.dim(2);
		let pes = self.pos_embedding.dim(0);

		let x_idx = ops::reshape(&ops::slice(positions, &[0, 0], &[n, 1])?, &[n])?;
		let y_idx = ops::reshape(&ops::slice(positions, &[0, 1], &[n, 2])?, &[n])?;

		let plane_x = ops::reshape(
			&ops::slice(&self.pos_embedding, &[0, 0, 0], &[pes, 1, d])?,
			&[pes, d],
		)?;
		let plane_y = ops::reshape(
			&ops::slice(&self.pos_embedding, &[0, 1, 0], &[pes, 2, d])?,
			&[pes, d],
		)?;

		let pe_x = ops::astype(&ops::take_axis(&plane_x, &x_idx, 0)?, dtype)?;
		let pe_y = ops::astype(&ops::take_axis(&plane_y, &y_idx, 0)?, dtype)?;
		let pe = ops::add(&pe_x, &pe_y)?;
		ops::add(hidden, &pe)
	}
}

/// Host-computed `(x, y)` patch positions for a `grid_h x grid_w` grid,
/// shaped `[n, 2]` int32 (row-major, matching
/// `UnifiedVisionEmbedder::patchify`).
fn build_positions(grid_h: i32, grid_w: i32) -> Result<Array> {
	let n = (grid_h * grid_w) as usize;
	let mut data = Vec::with_capacity(n * 2);
	for row in 0..grid_h {
		for col in 0..grid_w {
			data.push(col);
			data.push(row);
		}
	}
	Array::from_slice(&data, &[grid_h * grid_w, 2])
}