Replace with per-group INT8 quantization (group_size=128, lossless vs FP16)

Upload PyTorch INT8 quantized FastVLM-1.5B (weights + config + eval)

Upload evaluation results

initial commit