diff --git a/src/brevitas_examples/common/generative/quantize.py b/src/brevitas_examples/common/generative/quantize.py index 7ec94f507..f11c42efd 100644 --- a/src/brevitas_examples/common/generative/quantize.py +++ b/src/brevitas_examples/common/generative/quantize.py @@ -58,7 +58,7 @@ from brevitas_examples.common.generative.nn import LoRACompatibleQuantConv2d from brevitas_examples.common.generative.nn import LoRACompatibleQuantLinear from brevitas_examples.common.generative.quantizers import Fp8e4m3DynamicActPerGroupFloat -from brevitas_examples.common.generative.quantizers import FP8e4m3OCPDynamicActPerRowFloat +from brevitas_examples.common.generative.quantizers import FP8e4m3OCPDynamicActPerRowFixedPoint from brevitas_examples.common.generative.quantizers import Fp8e4m3WeightSymmetricGroupQuant from brevitas_examples.common.generative.quantizers import Int8DynamicActPerGroupFloat from brevitas_examples.common.generative.quantizers import Int8DynamicActPerRowFixedPoint @@ -198,10 +198,10 @@ 'dynamic': { 'po2_scale': { 'stats': { - 'per_group': { - 'sym': MXFloat8e4m3Act}, 'per_row': { - 'sym': FP8e4m3OCPDynamicActPerRowFloat}}}}}, + 'sym': FP8e4m3OCPDynamicActPerRowFixedPoint}, + 'per_group': { + 'sym': MXFloat8e4m3Act}}}}}, 'float_fnuz': { 'static': { 'float_scale': { diff --git a/src/brevitas_examples/common/generative/quantizers.py b/src/brevitas_examples/common/generative/quantizers.py index adf2e5bd9..8c1ae119c 100644 --- a/src/brevitas_examples/common/generative/quantizers.py +++ b/src/brevitas_examples/common/generative/quantizers.py @@ -130,7 +130,7 @@ class Fp8e4m3DynamicActPerGroupFloat(DynamicActProxyMixin, Fp8e4m3ActPerTensorFl scaling_stats_op = 'min_max' -class FP8e4m3OCPDynamicActPerRowFloat(Fp8e4m3ActPerTensorFloat): +class FP8e4m3OCPDynamicActPerRowFixedPoint(Fp8e4m3ActPerTensorFloat): """ Symmetric quantizer with per row dynamic scale. """