DeepSeek-16b-light

DeepSeek-16b-light

This is a 4-bit quantized version of the DeepSeek Coder V2 Lite Instruct model. The quantization was performed using the bitsandbytes library with 4-bit precision to reduce the model size and memory requirements while maintaining most of its capabilities.

Model Details

• Original Model: DeepSeek Coder V2 Lite Instruct
• Quantization: 4-bit quantization using bitsandbytes
• Compute Type: float16
• Double Quantization: Enabled
• Size Reduction: Approximately 75% smaller than the original model
• Use Case: Code generation, code completion, and programming assistance

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Noorhan/DeepSeek-16b-light")
model = AutoModelForCausalLM.from_pretrained("Noorhan/DeepSeek-16b-light", device_map="auto")

# Example code generation
prompt = """
Write a Python function to calculate the Fibonacci sequence up to n terms.
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    inputs.input_ids,
    max_length=500,
    temperature=0.7,
    top_p=0.95,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Performance and Limitations

This 4-bit quantized model:

• Requires significantly less memory than the original model
• Runs faster on consumer-grade hardware
• Has minimal quality degradation for most use cases
• May show some performance reduction for edge cases or complex reasoning tasks

Hardware Requirements

• Recommended: GPU with at least 8GB VRAM
• Minimum: 4GB VRAM (with potential performance limitations)

Acknowledgements

This model is a quantized version of DeepSeek Coder V2 Lite Instruct. All credits for the original model go to the DeepSeek AI team. The quantization was performed to make this powerful coding assistant more accessible for users with limited computational resources.

License

This model inherits the license of the original DeepSeek Coder V2 Lite Instruct model. Please refer to the original model's documentation for licensing details.