Switch SDPA quantization to FP8 FNUZ

Xilinx · Sep 10, 2024 · fcfedfa · fcfedfa
1 parent ea8cb36
commit fcfedfa
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/src/brevitas_examples/stable_diffusion/main.py b/src/brevitas_examples/stable_diffusion/main.py
@@ -364,7 +364,7 @@ def input_zp_stats_type():
                 dtype=dtype,
                 device=args.device,
                 weight_bit_width=weight_bit_width,
-                weight_quant_format='float_ocp_e4m3',
+                weight_quant_format='float_fnuz_e4m3',
                 weight_quant_type='sym',
                 weight_param_method=args.weight_param_method,
                 weight_scale_precision=args.weight_scale_precision,
@@ -373,7 +373,7 @@ def input_zp_stats_type():
                 quantize_weight_zero_point=args.quantize_weight_zero_point,
                 quantize_input_zero_point=args.quantize_input_zero_point,
                 input_bit_width=args.linear_output_bit_width,
-                input_quant_format='float_ocp_e4m3',
+                input_quant_format='float_fnuz_e4m3',
                 input_scale_type=args.input_scale_type,
                 input_scale_precision=args.input_scale_precision,
                 input_param_method=args.input_param_method,