tiny-minicpm-o2_6-test

Tiny MiniCPM-o-2_6 Model (≈6.3 MB)

This folder contains a compressed MiniCPM-o-2_6-compatible model designed for testing and Optimum-Intel integration.

Model Details

• Original Size: ~160 MB
• Compressed Size: ~6.3 MB (full folder)
• Vocab Size: 4000
• Hidden Size: 96
• Intermediate Size: 384
• Num Layers: 6
• Attention Heads: 4
• Total Parameters: 1,653,984
• Model Type: minicpmo (MiniCPMOConfig / MiniCPMOForCausalLM)

The model exposes a minimal vpm attribute for multimodal compatibility and a simple generate() method for basic generation tests.

Files

• config.json – Hugging Face config (auto_map, architectures, hyperparameters)
• pytorch_model.bin – model weights (~6.3 MB)
• configuration_minicpm.py – MiniCPMOConfig implementation
• modeling_minicpmo.py – MiniCPMOForCausalLM implementation (with vpm and generate)
• tokenizer.json, tokenizer_config.json – minimal tokenizer definition
• preprocessor_config.json, processor_config.json – minimal image/processor config

Usage

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained(
    "./tiny-minicpm",
    trust_remote_code=True,
)

processor = AutoProcessor.from_pretrained("./tiny-minicpm")

# Dummy input ids for a quick forward pass
import torch
input_ids = torch.randint(0, model.config.vocab_size, (1, 8))
outputs = model(input_ids=input_ids)
print(outputs.logits.shape)  # (1, 8, vocab_size)

Architecture

The compressed model keeps the MiniCPM-o-2_6 causal LM structure, but with much smaller dimensions:

• 6 transformer decoder layers
• 4 attention heads (hidden size 96)
• 96-dimensional token embeddings
• 4000-token vocabulary

It is intended for correctness and integration testing, not for high-quality text generation.