Apertus-8B-Instruct-2509-LOGEQ-FP8_dynamic

This model applies Logarithmic Equalization (LogEQ) followed by
full FP8 dynamic quantization using the LLM Compressor framework.

What this means in practice

• LogEQ smooths and rescales activations using logarithmic statistics,
  reducing activation outliers before quantization.
• Both weights and activations are quantized to FP8 dynamic format
  (E5M2 or E4M3 depending on layer type).
• The quantized model is compatible with vLLM FP8 execution. data-free activation tracing.

Quantization Details

• Quantization Scheme: FP8_dynamic
• Preprocessing: LogarithmicEqualizationModifier
• Method: Data-free FP8 dynamic quantization
• Targets: All Linear layers
• Ignored Layers: lm_head (kept in FP16 for stability)
• Tool: llm-compressor