File size: 5,566 Bytes
bb01096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
OpticalCompressor: learned Conv2D compression for Qwen3-VL visual tokens.

Inserted AFTER model.visual() (ViT + Merger), BEFORE the LLM.
Reduces ~6700 merged visual tokens → target_tokens (e.g. 256).

The key idea from DeepSeek-OCR: compress at the encoding stage via
learned spatial downsampling, rather than pruning at inference time.
"""

import math
import torch
import torch.nn as nn


class OpticalCompressor(nn.Module):
    def __init__(self, hidden_dim=3584, target_tokens=256, num_refine_layers=1):
        """
        Args:
            hidden_dim: LLM hidden dimension (3584 for Qwen3-VL-8B after merger)
            target_tokens: number of output tokens (default 256 = 16x16 grid)
            num_refine_layers: number of Transformer refinement layers
        """
        super().__init__()
        self.hidden_dim = hidden_dim
        self.target_tokens = target_tokens
        self.target_h = int(math.sqrt(target_tokens))
        self.target_w = self.target_h
        assert self.target_h * self.target_w == target_tokens

        # Stage 1: Spatial Conv compression
        # Input: [B, D, H, W] → progressive downsampling
        # Using depthwise-separable conv to keep params manageable
        # Note: GroupNorm(1, D) == LayerNorm over channels, works on [B,C,H,W]
        num_groups = 32 if hidden_dim % 32 == 0 else 1
        self.compress = nn.Sequential(
            # Block 1: reduce spatial by 2x
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=2,
                      padding=1, groups=hidden_dim, bias=False),
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=1, bias=False),
            nn.GroupNorm(num_groups, hidden_dim),
            nn.GELU(),
            # Block 2: reduce spatial by 2x again
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=2,
                      padding=1, groups=hidden_dim, bias=False),
            nn.Conv2d(hidden_dim, hidden_dim, kernel_size=1, bias=False),
            nn.GroupNorm(num_groups, hidden_dim),
            nn.GELU(),
        )

        # Stage 2: Adaptive pool to fixed grid
        self.pool = nn.AdaptiveAvgPool2d((self.target_h, self.target_w))

        # Stage 3: Transformer refinement (recover inter-token relationships)
        self.refine = nn.ModuleList([
            nn.TransformerEncoderLayer(
                d_model=hidden_dim, nhead=8,
                dim_feedforward=hidden_dim * 2,
                dropout=0.1, activation="gelu",
                batch_first=True, norm_first=True,
            )
            for _ in range(num_refine_layers)
        ])

        # Learnable position embedding for output tokens
        self.pos_embed = nn.Parameter(
            torch.randn(1, target_tokens, hidden_dim) * 0.02
        )

        self._init_weights()

    def _init_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out")
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
            elif isinstance(m, nn.Linear):
                nn.init.trunc_normal_(m.weight, std=0.02)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    def forward(self, visual_embeds, grid_thw):
        """
        Args:
            visual_embeds: [total_tokens, hidden_dim] — flat output from model.visual()
                           Contains tokens for ALL images in the batch concatenated.
            grid_thw: [num_images, 3] — (temporal, height, width) for each image
                      after the merger's spatial merge.

        Returns:
            compressed: [total_compressed, hidden_dim] — flat compressed tokens
            new_grid_thw: [num_images, 3] — new grid dimensions (1, target_h, target_w)
        """
        device = visual_embeds.device
        compressed_list = []
        offset = 0

        for i in range(grid_thw.shape[0]):
            t, h, w = grid_thw[i].tolist()
            t, h, w = int(t), int(h), int(w)
            n_tokens = t * h * w
            img_tokens = visual_embeds[offset:offset + n_tokens]  # [n, D]
            offset += n_tokens

            # Reshape to spatial: take first temporal frame for 2D conv
            # For images (not video), t=1 always
            x = img_tokens.reshape(t * h, w, self.hidden_dim)
            # Merge temporal into height for simplicity
            x = x.reshape(1, t * h, w, self.hidden_dim)  # [1, H, W, D]
            x = x.permute(0, 3, 1, 2)  # [1, D, H, W]

            # Stage 1: Conv compression
            x = self.compress(x)  # [1, D, H', W']

            # Stage 2: Pool to fixed size
            x = self.pool(x)  # [1, D, target_h, target_w]

            x = x.flatten(2).transpose(1, 2)  # [1, target_tokens, D]

            # Stage 3: Add position embedding + refine
            x = x + self.pos_embed.to(x.dtype)
            for layer in self.refine:
                x = layer(x)

            compressed_list.append(x.squeeze(0))  # [target_tokens, D]

        compressed = torch.cat(compressed_list, dim=0)  # [total_compressed, D]
        new_grid_thw = torch.tensor(
            [[1, self.target_h, self.target_w]] * grid_thw.shape[0],
            device=device, dtype=grid_thw.dtype,
        )
        return compressed, new_grid_thw

    def count_parameters(self):
        total = sum(p.numel() for p in self.parameters())
        trainable = sum(p.numel() for p in self.parameters() if p.requires_grad)
        return {"total": total, "trainable": trainable}