Qwen3-8B-FP8-KV

Introduction

This model was built by applying Quark with calibration samples from Pile dataset to Qwen/Qwen3-8B.

Quantization Strategy

• Quantized Layers: All linear layers excluding "lm_head", ".mlp.experts."
• Weight: FP8 symmetric per-tensor
• Activation: FP8 symmetric per-tensor
• KV Cache: FP8 symmetric per-tensor

Deployment

Quark has its own export format and allows FP8 quantized models to be efficiently deployed using the vLLM backend (vLLM-compatible).

Evaluation

Quark currently uses perplexity (PPL) as the evaluation metric for accuracy loss before and after quantization. The specific PPL algorithm can be referenced in the quantize_quark.py. The quantization evaluation results are conducted in pseudo-quantization mode, which may slightly differ from the actual quantized inference accuracy. These results are provided for reference only.

Evaluation scores

Benchmark	Qwen3-8B	Qwen3-8B-FP8-KV (this model)
Perplexity-wikitext2	9.531	9.708

Performance Summary

• Accuracy Retention: 98.15% (only 1.85% perplexity increase)
• Model Size: ~42% reduction vs FP16
• Memory Efficiency: FP8 KV-cache for extended context
• Hardware Optimization: AMD ROCm/HIP optimized

License

Based on Qwen3-8B licensing terms.