| Property | Value |
|---|---|
| Parameter Count | 70.6B |
| Model Type | Language Model (Instruct) |
| License | Llama3 |
| Quantization | FP8 |
| OpenLLM Score | 79.16 |
What is Meta-Llama-3-70B-Instruct-FP8?
Meta-Llama-3-70B-Instruct-FP8 is an optimized version of Meta's Llama-3 70B model, specifically designed for efficient deployment while maintaining near-original performance. This model implements FP8 quantization for both weights and activations, effectively reducing the model's memory footprint by approximately 50% compared to the original 16-bit version.
Implementation Details
The model employs sophisticated quantization techniques using AutoFP8, focusing on the linear operators within transformer blocks. It achieves remarkable efficiency while maintaining 99.55% of the original model's performance on benchmark tasks.
- Weight and activation quantization using FP8 data type
- Symmetric per-tensor quantization implementation
- Compatible with vLLM >= 0.5.0 for inference
- Calibrated using 512 sequences from UltraChat
Core Capabilities
- Benchmark Performance: 80.06% on MMLU (5-shot)
- Strong reasoning capabilities with 91.12% on GSM-8K
- Excellent performance on Hellaswag (85.41%) and Winogrande (83.03%)
- Optimized for English language tasks
- Suitable for commercial and research applications
Frequently Asked Questions
Q: What makes this model unique?
This model stands out for its optimal balance between performance and efficiency, using FP8 quantization to reduce resource requirements while maintaining 99.55% of the original model's accuracy. It's specifically optimized for deployment with vLLM, making it ideal for production environments.
Q: What are the recommended use cases?
The model is best suited for English language tasks, particularly in commercial and research applications requiring assistant-like chat capabilities. It's optimized for deployment scenarios where resource efficiency is crucial while maintaining high performance standards.