| {"model": "stabilityai/stable-diffusion-xl-base-1.0", "device": "cuda:0", "batch_size": 2, "prompt": 5000, "calibration_prompt": 500, "calibration_prompt_path": "captions.tsv", "checkpoint_name": "UNet_500by20steps_fnuz.ckpt", "load_checkpoint": null, "path_to_latents": "/proj/xlabs_t3/users/gfranco/fulltime/datasets/coco/latents/latents.pt", "path_to_coco": "/proj/xlabs_t3/users/gfranco/fulltime/datasets/coco", "resolution": 1024, "guidance_scale": 8.0, "calibration_steps": 20, "output_path": "/proj/xlabs_t3/users/gfranco/fulltime/stable_diffusion/checkpoints/fp8_v2", "quantize": true, "activation_equalization": false, "gptq": false, "bias_correction": false, "dtype": "float16", "attention_slicing": false, "export_target": "params_only", "export_weight_q_node": false, "conv_weight_bit_width": 8, "linear_weight_bit_width": 8, "conv_input_bit_width": 8, "act_eq_alpha": 0.9, "linear_input_bit_width": 8, "weight_param_method": "stats", "input_param_method": "stats", "input_scale_stats_op": "minmax", "input_zp_stats_op": "minmax", "weight_scale_precision": "float_scale", "input_scale_precision": "float_scale", "weight_quant_type": "sym", "input_quant_type": "sym", "weight_quant_format": "float_fnuz_e4m3", "input_quant_format": "float_fnuz_e4m3", "weight_quant_granularity": "per_tensor", "input_quant_granularity": "per_tensor", "input_scale_type": "static", "weight_group_size": 16, "sdpa_bit_width": 8, "sdpa_param_method": "stats", "sdpa_scale_stats_op": "minmax", "sdpa_zp_stats_op": "minmax", "sdpa_scale_precision": "float_scale", "sdpa_quant_type": "sym", "sdpa_quant_format": "float_fnuz_e4m3", "sdpa_quant_granularity": "per_tensor", "sdpa_scale_type": "static", "quant_blacklist": ["time_emb_proj", "time_embedding", "conv_in", "conv_out", "conv_shortcut", "add_embedding", "up_blocks.2.resnets.2.conv1", "up_blocks.2.resnets.2.conv2"], "quantize_weight_zero_point": true, "exclude_blacklist_act_eq": false, "quantize_input_zero_point": false, "quantize_sdpa_zero_point": false, "export_cpu_float32": false, "use_mlperf_inference": true, "use_negative_prompts": true, "dry_run": false, "override_conv_quant_config": false, "vae_fp16_fix": false, "share_qkv_quant": true} |