README.md

Llama-3-70B-Instruct-FP8-v2

• Weights and activations are per-tensor quantized to float8_e4m3.
• Quantization with AutoFP8 with the updated activation scaling factor names.
• Calibration dataset: Ultrachat (mgoin/ultrachat_2k)
• Samples: 1024
• Sequence length: 4096

Evaluation

TBA