README.md

---
library_name: transformers
base_model:
- openbmb/MiniCPM-V-4.6
---

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from [openbmb/MiniCPM-V-4.6](https://huggingface.co/openbmb/MiniCPM-V-4.6).

| File path | Size |
|------|------|
| model.safetensors | 7.2MB |


### Example usage:

```python
import torch
from transformers import MiniCPMV4_6ForConditionalGeneration, AutoProcessor

model_id = "tiny-random/minicpm-v-4.6"
model = MiniCPMV4_6ForConditionalGeneration.from_pretrained(
    model_id,
    trust_remote_code=True,
    attn_implementation='sdpa',
    dtype=torch.bfloat16,
)
model = model.eval().cuda()
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "video", "url": "https://huggingface.co/datasets/openbmb/DemoCase/resolve/main/football.mp4"},
            {"type": "text", "text": "Describe this video in detail. Follow the timeline and focus on on-screen text, interface changes, main actions, and scene changes."},
        ],
    },
    {
        "role": "assistant",
        "content": [{"type": "text", "text": "Dummy response for video"}],
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/openbmb/DemoCase/resolve/main/refract.png"},
            {"type": "text", "text": "What causes this phenomenon?"},
        ],
    },
]

downsample_mode = "16x"  # Using `downsample_mode="4x"` for Finer Detail
inputs = processor.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_dict=True, return_tensors="pt",
    downsample_mode=downsample_mode,
    max_num_frames=128,
    stack_frames=1,
    max_slice_nums=1,
    use_image_id=False,
).to(model.device)

generated_ids = model.generate(**inputs, downsample_mode=downsample_mode, max_new_tokens=32)[0]
output_text = processor.decode(generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False)
print(output_text.replace('<|video_pad|>', 'V').replace('<|image_pad|>', 'I'))
```

### Codes to create this repo:

<details>
<summary>Click to expand</summary>

```python
import json
from pathlib import Path

import torch
from huggingface_hub import hf_hub_download
from transformers import (
    AutoConfig,
    AutoModel,
    AutoProcessor,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "openbmb/MiniCPM-V-4.6"
save_folder = "/tmp/tiny-random/minicpm-v-46"

processor = AutoProcessor.from_pretrained(source_model_id, trust_remote_code=True)
processor.save_pretrained(save_folder)

with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model'), 'r', encoding='utf-8') as f:
    config_json: dict = json.load(f)

text_config = config_json["text_config"]
text_config.update({
    "head_dim": 32,
    "hidden_size": 8,
    "intermediate_size": 64,
    "layer_types": [
        "linear_attention",
        "linear_attention",
        "linear_attention",
        "full_attention",
    ],
    "linear_key_head_dim": 32,
    "linear_num_key_heads": 4,
    "linear_num_value_heads": 4,
    "linear_value_head_dim": 32,
    "num_attention_heads": 8,
    "num_hidden_layers": 4,
    "num_key_value_heads": 2,
})
vision_config = config_json["vision_config"]
vision_config.update({
    "hidden_size": 128,
    "intermediate_size": 128,
    "num_attention_heads": 4,
    "num_hidden_layers": 2,
})

with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(
    save_folder,
    trust_remote_code=True,
)
print(config)
torch.set_default_dtype(torch.bfloat16)
model = AutoModel.from_config(config, trust_remote_code=True)
torch.set_default_dtype(torch.float32)
model.generation_config = GenerationConfig.from_pretrained(
    source_model_id, trust_remote_code=True,
)
set_seed(42)
num_params = sum(p.numel() for p in model.parameters())
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape, p.dtype, p.device, f'{p.numel() / num_params * 100: .2f}%')
model.save_pretrained(save_folder)
for f in Path(save_folder).glob('*.py'):
    f.unlink()
```

</details>

### Printing the model:

<details><summary>Click to expand</summary>

```text
MiniCPMV4_6Model(
  (vision_tower): MiniCPMV4_6VisionModel(
    (embeddings): MiniCPMV4_6VisionEmbeddings(
      (patch_embedding): Conv2d(3, 128, kernel_size=(14, 14), stride=(14, 14), padding=valid)
      (position_embedding): Embedding(4900, 128)
    )
    (encoder): MiniCPMV4_6VisionEncoder(
      (layers): ModuleList(
        (0-1): 2 x MiniCPMV4_6VisionEncoderLayer(
          (layer_norm1): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
          (self_attn): MiniCPMV4_6VisionAttention(
            (k_proj): Linear(in_features=128, out_features=128, bias=True)
            (v_proj): Linear(in_features=128, out_features=128, bias=True)
            (q_proj): Linear(in_features=128, out_features=128, bias=True)
            (out_proj): Linear(in_features=128, out_features=128, bias=True)
          )
          (layer_norm2): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
          (mlp): MiniCPMV4_6VisionMLP(
            (activation_fn): GELUTanh()
            (fc1): Linear(in_features=128, out_features=128, bias=True)
            (fc2): Linear(in_features=128, out_features=128, bias=True)
          )
        )
      )
    )
    (post_layernorm): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
    (vit_merger): MiniCPMV4_6ViTWindowAttentionMerger(
      (self_attn): MiniCPMV4_6VisionAttention(
        (k_proj): Linear(in_features=128, out_features=128, bias=True)
        (v_proj): Linear(in_features=128, out_features=128, bias=True)
        (q_proj): Linear(in_features=128, out_features=128, bias=True)
        (out_proj): Linear(in_features=128, out_features=128, bias=True)
      )
      (layer_norm1): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
      (pre_norm): LayerNorm((512,), eps=1e-06, elementwise_affine=True)
      (linear_1): Linear(in_features=512, out_features=512, bias=True)
      (act): GELUTanh()
      (linear_2): Linear(in_features=512, out_features=128, bias=True)
    )
  )
  (language_model): Qwen3_5TextModel(
    (embed_tokens): Embedding(248094, 8)
    (layers): ModuleList(
      (0-2): 3 x Qwen3_5DecoderLayer(
        (linear_attn): Qwen3_5GatedDeltaNet(
          (act): SiLUActivation()
          (conv1d): Conv1d(384, 384, kernel_size=(4,), stride=(1,), padding=(3,), groups=384, bias=False)
          (norm): Qwen3_5RMSNormGated()
          (out_proj): Linear(in_features=128, out_features=8, bias=False)
          (in_proj_qkv): Linear(in_features=8, out_features=384, bias=False)
          (in_proj_z): Linear(in_features=8, out_features=128, bias=False)
          (in_proj_b): Linear(in_features=8, out_features=4, bias=False)
          (in_proj_a): Linear(in_features=8, out_features=4, bias=False)
        )
        (mlp): Qwen3_5MLP(
          (gate_proj): Linear(in_features=8, out_features=64, bias=False)
          (up_proj): Linear(in_features=8, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=8, bias=False)
          (act_fn): SiLUActivation()
        )
        (input_layernorm): Qwen3_5RMSNorm((8,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5RMSNorm((8,), eps=1e-06)
      )
      (3): Qwen3_5DecoderLayer(
        (self_attn): Qwen3_5Attention(
          (q_proj): Linear(in_features=8, out_features=512, bias=False)
          (k_proj): Linear(in_features=8, out_features=64, bias=False)
          (v_proj): Linear(in_features=8, out_features=64, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
          (q_norm): Qwen3_5RMSNorm((32,), eps=1e-06)
          (k_norm): Qwen3_5RMSNorm((32,), eps=1e-06)
        )
        (mlp): Qwen3_5MLP(
          (gate_proj): Linear(in_features=8, out_features=64, bias=False)
          (up_proj): Linear(in_features=8, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=8, bias=False)
          (act_fn): SiLUActivation()
        )
        (input_layernorm): Qwen3_5RMSNorm((8,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5RMSNorm((8,), eps=1e-06)
      )
    )
    (norm): Qwen3_5RMSNorm((8,), eps=1e-06)
    (rotary_emb): Qwen3_5TextRotaryEmbedding()
  )
  (merger): MiniCPMV4_6Merger(
    (mlp): ModuleList(
      (0): MiniCPMV4_6DownsampleMLP(
        (pre_norm): LayerNorm((512,), eps=1e-06, elementwise_affine=True)
        (linear_1): Linear(in_features=512, out_features=512, bias=True)
        (act): GELU(approximate='none')
        (linear_2): Linear(in_features=512, out_features=8, bias=True)
      )
    )
  )
)
```

</details>

### Test environment:

- torch: 2.10.0+cu128
- transformers: 5.9.0