README.md

Mixtral-8x22B-Instruct-v0.1-FP8-v2

• Weights and activations are per-tensor quantized to float8_e4m3.
• Quantization with AutoFP8 with updated activation scaling factor names.
• Calibration dataset: Ultrachat (mgoin/ultrachat_2k)
• Samples: 2048
• Sequence length: 8192

Evaluation

TBA