tiny-random-MiniCPM-o-2_6

tiny-random-MiniCPM-o-2_6

A minimal, randomly initialized version of MiniCPM-o-2_6 designed for testing and development purposes. This model maintains the same architecture as the original MiniCPM-o-2_6 but with drastically reduced dimensions to create a lightweight test model.

Model Details

Model Description

This is a tiny, randomly initialized version of the MiniCPM-o-2_6 multimodal model. It was created by scaling down the original model's dimensions while preserving the architecture structure. The model is intended for:

• Testing and development workflows
• Integration testing with Optimum-Intel
• Quick prototyping and experimentation
• CI/CD pipelines requiring lightweight models

⚠️ Important: This model is randomly initialized and should NOT be used for production inference. It is designed solely for testing purposes.

Model Architecture

The model maintains the same architecture as MiniCPM-o-2_6 but with reduced dimensions:

Language Model (LLM):

• hidden_size: 40
• num_hidden_layers: 1
• num_attention_heads: 4
• num_key_value_heads: 2
• intermediate_size: 16
• max_position_embeddings: 128
• vocab_size: 151,700

Vision Component:

• hidden_size: 16
• num_hidden_layers: 1
• num_attention_heads: 4
• intermediate_size: 8
• patch_size: 14

Audio/TTS Components:

• Audio: Disabled (init_audio: false)
• TTS: Disabled (init_tts: false)

Model Size

• Total Parameters: ~6.17M
• Model Size: ~12.4 MB (on disk)
• Precision: bfloat16

Usage

Basic Usage

from transformers import AutoModel, AutoTokenizer, AutoProcessor
import torch
from PIL import Image

# Load model and tokenizer
model_id = "notlikejoe/tiny-random-MiniCPM-o-2_6"
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

# Prepare inputs
text = "Hello, how are you?"
image = Image.new('RGB', (224, 224), color='red')  # Dummy image

# Process inputs
inputs = processor(text=text, images=image, return_tensors="pt")

# Forward pass
model.eval()
with torch.no_grad():
    outputs = model(**inputs)

With Optimum-Intel

This model is compatible with Optimum-Intel for OpenVINO optimization:

from optimum.intel import OVModelForCausalLM
from transformers import AutoTokenizer

model_id = "notlikejoe/tiny-random-MiniCPM-o-2_6"

# Export to OpenVINO format
ov_model = OVModelForCausalLM.from_pretrained(
    model_id,
    export=True,
    trust_remote_code=True
)

# Use for inference
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

Model Validation

The model has been validated to ensure:

✅ Model loads successfully from Hugging Face
✅ Config, tokenizer, and processor load correctly
✅ Model structure matches expected architecture
✅ Compatible with Optimum-Intel export
✅ Forward pass completes without errors
✅ OpenVINO compatibility fix applied: Resampler num_heads=0 issue resolved

OpenVINO Compatibility Fix

This model includes a fix for the OpenVINO loading issue where num_heads=0 would occur with small embed_dim values. The resampler's num_heads calculation has been patched to ensure it's always at least 1:

# Original: num_heads = embed_dim // 128  # Would be 0 when embed_dim=40
# Fixed:    num_heads = 1 if embed_dim < 128 else max(1, embed_dim // 128)

The modeling_minicpmo.py file included with this model contains this fix, ensuring compatibility with Optimum-Intel OpenVINO export and loading.

Limitations

1. Random Initialization: This model is randomly initialized and will not produce meaningful outputs
2. Reduced Dimensions: The model dimensions are minimal and may not capture complex patterns
3. Testing Only: This model is intended for testing and development, not production use
4. Limited Vocabulary: The vocabulary has been reduced to 2000 entries for size optimization

Training Details

This model was not trained. It is a randomly initialized, dimensionally-reduced version of MiniCPM-o-2_6 created for testing purposes.

Training Data

N/A - Model is randomly initialized.

Evaluation

This model is not intended for evaluation on standard benchmarks as it is randomly initialized.

Citation

If you use this model, please cite the original MiniCPM-o-2_6 model:

@misc{minicpm-o-2_6,
  title={MiniCPM-o-2_6},
  author={OpenBMB},
  year={2024},
  howpublished={\url{https://huggingface.co/openbmb/MiniCPM-o-2_6}}
}

Model Card Contact

For questions or issues related to this model, please open an issue in the repository.

License

This model is licensed under the Apache 2.0 License, same as the base model.