Sparse-Llama-3.1-8B-ultrachat_200k-2of4-FP8-dynamic

Model Overview

• Model Architecture: Llama-3.1-8B
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Sparsity: 2:4
  • Weight quantization: FP8
  • Activation quantization: FP8
• Release Date: 11/15/2024
• Version: 1.0
• License(s): llama3.1
• Model Developers: Neural Magic

This is a multi-turn conversational AI model obtained by fine-tuning the 2:4 sparse Sparse-Llama-3.1-8B-2of4 on the ultrachat_200k dataset, followed by quantization. On the AlpacaEval benchmark (version 1), it achieves a score of 62.9, compared to 62.0 for the fine-tuned dense model Llama-3.1-8B-ultrachat_200k — demonstrating a 99.4% accuracy recovery.

Model Optimizations

This model was obtained by quantizing the weights of Sparse-Llama-3.1-8B-ultrachat_200k-2of4 to FP8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per-channel scheme, where a fixed linear scaling factor is applied between FP8 and BF16 representations for each output channel dimension. Linear scaling factors are computed via by minimizing the mean squarred error (MSE). Activations are quantized with a symmetric dynamic per-token scheme, computing a linear scaling factor at runtime for each token between FP8 and BF16 representations.

Deployment with vLLM

This model can be deployed efficiently using the vLLM backend. vLLM also supports OpenAI-compatible serving. See the documentation for more details.

Evaluation

This model was evaluated on Neural Magic's fork of AlpacaEval benchmark. We adopt the same setup as in Enabling High-Sparsity Foundational Llama Models with Efficient Pretraining and Deployment, using version 1 of the benchmark and Llama-2-70b-chat as the annotator.

Accuracy

AlpacaEval Benchmark

Metric	Llama-3.1-8B-ultrachat_200k	Sparse-Llama-3.1-8B-ultrachat_200k-2of4	Sparse-Llama-3.1-8B-ultrachat_200k-2of4-FP8-dynamic
Win rate	62.0	61.1	62.9