File size: 10,021 Bytes
ee021b1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6902421
 
 
 
 
 
 
ee021b1
6902421
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ee021b1
6902421
 
 
 
 
ee021b1
 
6902421
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ee021b1
 
 
 
 
 
 
 
 
 
 
 
 
6902421
 
 
 
 
 
 
 
 
 
ee021b1
 
 
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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
---
license: apache-2.0
language:
- vi
- en
tags:
- vision-language-model
- vlm
- qwen3
- fastvlm
- vietnamese
base_model: Qwen/Qwen3-0.6B
datasets:
- 5CD-AI/Viet-multimodal-open-r1-8k-verified
---

# Belle-VLM: Vietnamese Vision Language Model

## Model Description

Belle-VLM is a Vision Language Model trained for Vietnamese multimodal reasoning tasks.

### Architecture
- **LLM Backbone**: Qwen3-0.6B
- **Vision Encoder**: FastViTHD (MobileCLIP)
- **Projector**: MLP 2-layer (3072 -> 1024)

### Training
- **Dataset**: 5CD-AI/Viet-multimodal-open-r1-8k-verified
- **Method**: LoRA fine-tuning
- **Steps**: 2
- **Learning Rate**: 2e-05

## Usage

```python
!pip install -q transformers>=4.51.0 torch torchvision timm pillow requests datasets einops
!pip install -q open-clip-torch
!git clone https://github.com/Hert4/ml-fastvlm-v2.git
%cd ml-fastvlm-v2

import sys
import os
import torch
import requests
from PIL import Image
from io import BytesIO

# Add path to ml-fastvlm-v2
FASTVLM_PATH = "/kaggle/working/ml-fastvlm-v2"
if os.path.exists(FASTVLM_PATH):
    sys.path.insert(0, FASTVLM_PATH)
else:
    # Local path fallback
    sys.path.insert(0, ".")


MODEL_PATH = "beyoru/Belle-VLM"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32

print(f"Device: {DEVICE}")
print(f"Dtype: {DTYPE}")

from transformers import AutoTokenizer
from llava.model.language_model.llava_qwen import LlavaQwen3ForCausalLM

print(f"\nLoading model from: {MODEL_PATH}")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    MODEL_PATH,
    trust_remote_code=True,
    use_fast=False
)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Load model
model = LlavaQwen3ForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=DTYPE,
    device_map="auto",
    trust_remote_code=True,
)
model.eval()

# Setup vision tower
vision_tower = model.get_vision_tower()
if not vision_tower.is_loaded:
    vision_tower.load_model()
vision_tower = vision_tower.to(device=model.device, dtype=DTYPE)
image_processor = vision_tower.image_processor

print("Model loaded!")


def debug_model():
    """Check if model components are properly loaded."""
    print("\n" + "=" * 60)
    print("DEBUG: Checking model components")
    print("=" * 60)

    # 1. Check mm_projector
    if hasattr(model.model, 'mm_projector'):
        mm_proj = model.model.mm_projector
        print(f"[OK] mm_projector exists: {type(mm_proj)}")

        # Check weights
        if isinstance(mm_proj, torch.nn.Sequential):
            first_layer = mm_proj[0]
            w = first_layer.weight.float()  # Convert to float for stats
            print(f"     First layer shape: {w.shape}")
            print(f"     Weight mean: {w.mean().item():.6f}")
            print(f"     Weight std:  {w.std().item():.6f}")
            print(f"     Weight range: [{w.min().item():.4f}, {w.max().item():.4f}]")

            # Check if looks trained
            if w.std().item() > 0.3:
                print("     [WARNING] Std too high - might be random init!")
            else:
                print("     [OK] Weights look trained")
    else:
        print("[ERROR] mm_projector NOT FOUND!")
        return False

    # 2. Check vision tower
    print(f"\n[OK] Vision tower: {type(vision_tower).__name__}")
    print(f"     Is loaded: {vision_tower.is_loaded}")
    print(f"     Hidden size: {getattr(vision_tower, 'hidden_size', 'unknown')}")

    # 3. Check config
    print(f"\n[INFO] Config:")
    print(f"     mm_vision_tower: {getattr(model.config, 'mm_vision_tower', 'NOT SET')}")
    print(f"     mm_hidden_size: {getattr(model.config, 'mm_hidden_size', 'NOT SET')}")
    print(f"     mm_projector_type: {getattr(model.config, 'mm_projector_type', 'NOT SET')}")
    print(f"     LLM hidden_size: {model.config.hidden_size}")

    return True


def debug_image_encoding(image):
    """Test image encoding pipeline."""
    from llava.mm_utils import process_images

    print("\n" + "=" * 60)
    print("DEBUG: Testing image encoding pipeline")
    print("=" * 60)

    # Process image
    image_tensor = process_images([image], image_processor, model.config)[0]
    image_tensor = image_tensor.unsqueeze(0).to(dtype=DTYPE, device=model.device)

    print(f"[1] Image tensor:")
    print(f"    Shape: {image_tensor.shape}")
    print(f"    Dtype: {image_tensor.dtype}")
    print(f"    Range: [{image_tensor.min().item():.3f}, {image_tensor.max().item():.3f}]")

    # Get vision features
    with torch.no_grad():
        vision_features = vision_tower(image_tensor)
        print(f"\n[2] Vision features (from vision tower):")
        print(f"    Shape: {vision_features.shape}")
        print(f"    Mean: {vision_features.float().mean().item():.6f}")
        print(f"    Std:  {vision_features.float().std().item():.6f}")

        # Check for NaN/Inf
        if torch.isnan(vision_features).any():
            print("    [ERROR] Contains NaN!")
        if torch.isinf(vision_features).any():
            print("    [ERROR] Contains Inf!")

        # Project through mm_projector
        projected = model.model.mm_projector(vision_features)
        print(f"\n[3] Projected features (after mm_projector):")
        print(f"    Shape: {projected.shape}")
        print(f"    Mean: {projected.float().mean().item():.6f}")
        print(f"    Std:  {projected.float().std().item():.6f}")

        # Check for NaN/Inf
        if torch.isnan(projected).any():
            print("    [ERROR] Contains NaN!")
        if torch.isinf(projected).any():
            print("    [ERROR] Contains Inf!")

        # Compare with text embeddings scale
        sample_ids = torch.tensor([[1, 2, 3]], device=model.device)
        text_embeds = model.model.embed_tokens(sample_ids)
        print(f"\n[4] Text embeddings (for comparison):")
        print(f"    Mean: {text_embeds.float().mean().item():.6f}")
        print(f"    Std:  {text_embeds.float().std().item():.6f}")

        # Check scale match
        proj_std = projected.float().std().item()
        text_std = text_embeds.float().std().item()
        ratio = proj_std / text_std if text_std > 0 else float('inf')
        print(f"\n[5] Scale ratio (projected/text): {ratio:.2f}")
        if ratio > 10 or ratio < 0.1:
            print("    [WARNING] Scale mismatch! May cause issues.")
        else:
            print("    [OK] Scales are similar")

    return projected



from llava.conversation import conv_templates
from llava.mm_utils import process_images, tokenizer_image_token
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN


def chat(image, question, temperature=0.7, max_new_tokens=512, debug=False):
    """
    Chat with the model about an image.

    Args:
        image: PIL Image
        question: str
        temperature: float (0.0 = deterministic, higher = more random)
        max_new_tokens: int
        debug: bool - print debug info

    Returns:
        str: Model response
    """
    if debug:
        debug_image_encoding(image)

    # Process image
    image_tensor = process_images([image], image_processor, model.config)[0]

    # Build conversation
    conv = conv_templates["qwen_3"].copy()
    conv.append_message(conv.roles[0], f"{DEFAULT_IMAGE_TOKEN}\n{question}")
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()

    if debug:
        print(f"\n[DEBUG] Prompt:\n{prompt[:500]}...")

    # Tokenize
    input_ids = tokenizer_image_token(
        prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt'
    ).unsqueeze(0).to(model.device)

    if debug:
        print(f"\n[DEBUG] Input IDs shape: {input_ids.shape}")
        # Check for IMAGE_TOKEN_INDEX
        num_image_tokens = (input_ids == IMAGE_TOKEN_INDEX).sum().item()
        print(f"[DEBUG] Number of image tokens: {num_image_tokens}")

    # Generate
    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=image_tensor.unsqueeze(0).to(dtype=DTYPE, device=model.device),
            image_sizes=[image.size],
            do_sample=temperature > 0,
            temperature=temperature if temperature > 0 else None,
            max_new_tokens=max_new_tokens,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )

    # Decode
    response = tokenizer.decode(output_ids[0], skip_special_tokens=True)

    # Clean up response (remove prompt echo if present)
    if "<|im_start|>" in response:
        # Find the assistant's response
        parts = response.split("<|im_start|>assistant")
        if len(parts) > 1:
            response = parts[-1].strip()

    return response



if __name__ == "__main__":
    # Run debug checks
    debug_model()

    # Load test image
    print("\n" + "=" * 60)
    print("Loading test image...")
    print("=" * 60)

    url = "<your url here>"
    image = Image.open(BytesIO(requests.get(url).content)).convert("RGB")
    print(f"Image size: {image.size}")

    # Debug image encoding
    debug_image_encoding(image)

    # Test chat
    print("\n" + "=" * 60)
    print("Testing chat...")
    print("=" * 60)

    questions = [
        "What is in this image?",
        "Mô tả hình ảnh này",
        "Con vật trong ảnh là gì?",
    ]

    for q in questions:
        print(f"\nQ: {q}")
        response = chat(image, q, temperature=0.3, max_new_tokens=256)
        print(f"A: {response[:500]}...")
        print("-" * 40)

```

## Training Details

| Parameter | Value |
|-----------|-------|
| Base Model | Qwen/Qwen3-0.6B |
| Vision Tower | apple/MobileCLIP2-S4 |
| LoRA Rank | 4 |
| LoRA Alpha | 8 |
| Batch Size | 1 x 1 |
| Max Steps | 2 |


## Conclusion:

> Limitation of MobileClip or training might make low quality model.

## 🤝 Contribute

[![GitHub](https://img.shields.io/badge/GitHub-ml--fastvlm--v2-181717?style=for-the-badge&logo=github)](https://github.com/Hert4/ml-fastvlm-v2)


## License

Apache 2.0