review

Xilinx · Jan 18, 2024 · 0947d85 · 0947d85
1 parent 0fcf1e6
commit 0947d85
Show file tree

Hide file tree

Showing 3 changed files with 42 additions and 11 deletions.
diff --git a/src/brevitas/export/onnx/standard/qcdq/handler.py b/src/brevitas/export/onnx/standard/qcdq/handler.py
@@ -94,9 +94,9 @@ def int32_dtype(cls):
 
     def validate(self, module):
         super().validate(module)
-        # ONNX QuantizeLinear supports only 8b output with round to nearest even.
-        # Below 8b quantization is supported through clipping.
+        # ONNX DynamicQuantizeLinear supports only 8b output with round to nearest even.
         assert module.rounding_mode.upper() == 'ROUND', 'Only round to nearest even supported'
+        # Below 8b quantization is not supported.
         self.validate_8b_bit_width(module.bit_width(), le_then=False)
 
     def quantize_fn(self, x, dtype):

diff --git a/src/brevitas_examples/common/generative/quantize.py b/src/brevitas_examples/common/generative/quantize.py
@@ -40,6 +40,8 @@
 from brevitas_examples.common.generative.quantizers import Int8ActPerRowFloat
 from brevitas_examples.common.generative.quantizers import Int8ActPerRowFloatMSE
 from brevitas_examples.common.generative.quantizers import IntWeightSymmetricGroupQuant
+from brevitas_examples.common.generative.quantizers import ShiftedUint8ActDynamicPerGroupFloat
+from brevitas_examples.common.generative.quantizers import ShiftedUint8ActDynamicPerRowFloat
 from brevitas_examples.common.generative.quantizers import ShiftedUint8ActDynamicPerTensorFloat
 from brevitas_examples.common.generative.quantizers import ShiftedUint8ActPerRowFloat
 from brevitas_examples.common.generative.quantizers import ShiftedUint8ActPerRowFloatMSE
@@ -112,9 +114,11 @@
                         'sym': Int8ActDynamicPerTensorFloat,
                         'asym': ShiftedUint8ActDynamicPerTensorFloat},
                     'per_row': {
-                        'sym': Int8ActDynamicPerRowFloat},
+                        'sym': Int8ActDynamicPerRowFloat,
+                        'asym': ShiftedUint8ActDynamicPerRowFloat},
                     'per_group': {
-                        'sym': Int8ActDynamicPerGroupFloat},}}}},
+                        'sym': Int8ActDynamicPerGroupFloat,
+                        'asym': ShiftedUint8ActDynamicPerGroupFloat},}}}},
     'float': {
         'static': {
             'float_scale': {

diff --git a/src/brevitas_examples/common/generative/quantizers.py b/src/brevitas_examples/common/generative/quantizers.py
@@ -67,6 +67,10 @@ def reshaped_scaling_shape(module):
     block_size = None
 
 
+class ActDynamicProxyMixin(ExtendedInjector):
+    proxy_class = DynamicActQuantProxyFromInjector
+
+
 class IntWeightSymmetricGroupQuant(WeightSymmetricGroupQuantMixin, Int8WeightPerChannelFloat):
     """
     Block / group / vector signed symmetric int weight quantizer with float scales.
@@ -130,11 +134,32 @@ class Int8ActDynamicPerTensorFloat(Int8ActPerTensorFloat):
     scaling_stats_op = 'max'
 
 
-class ShiftedUint8ActDynamicPerTensorFloat(ShiftedUint8ActPerTensorFloat):
+class Int8ActDynamicPerRowFloat(Int8ActPerRowFloat, ActDynamicProxyMixin):
     """
-    Symmetric quantizer with per tensor dynamic scale.
+    Symmetric quantizer with per row dynamic scale.
+    """
+    scaling_impl = RuntimeDynamicStatsScaling
+    scaling_stats_input_view_shape_impl = OverBatchOverOutputChannelView
+    scaling_stats_op = 'max'
+
+
+class Int8ActDynamicPerGroupFloat(Int8ActPerRowFloat, ActDynamicProxyMixin):
+    """
+    Symmetric quantizer with per group scale.
+    """
+    scaling_impl = RuntimeDynamicGroupStatsScaling
+    keepdim = True
+    scaling_stats_op = 'max'
+
+    @value
+    def stats_reduce_dim(group_dim):
+        return group_dim + 1
+
+
+class ShiftedUint8ActDynamicPerTensorFloat(ShiftedUint8ActPerTensorFloat, ActDynamicProxyMixin):
+    """
+    Asymmetric quantizer with per tensor dynamic scale.
     """
-    proxy_class = DynamicActQuantProxyFromInjector
     scaling_impl = RuntimeDynamicStatsScaling
     scaling_stats_input_view_shape_impl = OverTensorView
     scaling_stats_op = 'max'
@@ -143,22 +168,24 @@ class ShiftedUint8ActDynamicPerTensorFloat(ShiftedUint8ActPerTensorFloat):
     stats_reduce_dim = 0
 
 
-class Int8ActDynamicPerRowFloat(Int8ActPerRowFloat):
+class ShiftedUint8ActDynamicPerRowFloat(ShiftedUint8ActPerRowFloat, ActDynamicProxyMixin):
     """
-    Symmetric quantizer with per row dynamic scale.
+    Asymmetric quantizer with per row dynamic scale.
     """
     scaling_impl = RuntimeDynamicStatsScaling
     scaling_stats_input_view_shape_impl = OverBatchOverOutputChannelView
     scaling_stats_op = 'max'
+    zero_point_stats_impl = NegativeMinOrZero
 
 
-class Int8ActDynamicPerGroupFloat(Int8ActPerRowFloat):
+class ShiftedUint8ActDynamicPerGroupFloat(ShiftedUint8ActPerRowFloat, ActDynamicProxyMixin):
     """
-    Symmetric quantizer with per group scale.
+    Asymmetric quantizer with per group dynamic scale.
     """
     scaling_impl = RuntimeDynamicGroupStatsScaling
     keepdim = True
     scaling_stats_op = 'max'
+    zero_point_stats_impl = NegativeMinOrZero
 
     @value
     def stats_reduce_dim(group_dim):