Feat (core): quantized scale/zero_point (#1038)

src/brevitas/core/restrict_val.py

@@ -170,3 +170,30 @@ def forward(self, x: Tensor):
         x = self.float_to_int_impl(x)
         x = self.power_of_two(x)
         return x
+
+
+class QuantRestrictValue(brevitas.jit.ScriptModule):
+
+    def __init__(self, restrict_value_float_to_int_impl: Module):
+        super(QuantRestrictValue, self).__init__()
+        self.float_to_int_impl = restrict_value_float_to_int_impl
+
+    def restrict_init_float(self, x: float):
+        return Identity()
+
+    def restrict_init_tensor(self, x: torch.Tensor):
+        return Identity()
+
+    def restrict_init_module(self):
+        return Identity()
+
+    def restrict_init_inplace_module(self):
+        return Identity()
+
+    def retrocompatibility_op(self, x):
+        return Identity()
+
+    @brevitas.jit.script_method
+    def forward(self, x: torch.Tensor):
+        o, *_ = self.float_to_int_impl(x)
+        return o

src/brevitas/core/zero_point.py

@@ -60,6 +60,18 @@ def forward(self, zero_point: Tensor, scale: Tensor, bit_width: Tensor) -> Tenso
         return out
 
 
+class _ScaleShiftQuantZeroPoint(brevitas.jit.ScriptModule):
+
+    def __init__(self, zp_int_quant: Module) -> None:
+        super(_ScaleShiftQuantZeroPoint, self).__init__()
+        self.zp_int_quant = zp_int_quant
+
+    @brevitas.jit.script_method
+    def forward(self, zero_point: Tensor, scale: Tensor, bit_width: Tensor) -> Tensor:
+        quant_zp, *_ = self.zp_int_quant(zero_point)
+        return quant_zp
+
+
 class StatsFromParameterZeroPoint(brevitas.jit.ScriptModule):
 
     def __init__(
@@ -70,15 +82,21 @@ def __init__(
             zero_point_stats_input_concat_dim: int,
             zero_point_stats_impl: Module,
             zero_point_shape: Tuple[int, ...],
-            tracked_parameter_list: List[torch.nn.Parameter]) -> None:
+            tracked_parameter_list: List[torch.nn.Parameter],
+            scale_shift_zero_point_impl: Optional[Module] = None) -> None:
         super(StatsFromParameterZeroPoint, self).__init__()
         self.parameter_list_stats = _ParameterListStats(
             zero_point_stats_impl,
             zero_point_shape,
             zero_point_stats_input_view_shape_impl,
             zero_point_stats_input_concat_dim,
             tracked_parameter_list)
-        self.scale_shift_zero_point = _ScaleShiftZeroPoint(int_quant, quantize_zero_point)
+        # This is for backward compatibility. Having int_quant/quantize_zero_point required for this
+        # interface but not for the else seems a bit off and might require some clean-up.
+        if scale_shift_zero_point_impl is None:
+            self.scale_shift_zero_point = _ScaleShiftZeroPoint(int_quant, quantize_zero_point)
+        else:
+            self.scale_shift_zero_point = scale_shift_zero_point_impl
 
     @brevitas.jit.script_method
     def forward(self, x: Tensor, scale: Tensor, bit_width: Tensor) -> torch.Tensor:

tests/brevitas/core/test_scaling_quant.py

@@ -0,0 +1,131 @@
+from dependencies import this
+from dependencies import value
+import torch
+
+from brevitas.core.quant.int import RescalingIntQuant
+from brevitas.core.restrict_val import QuantRestrictValue
+from brevitas.core.stats.stats_wrapper import SCALAR_SHAPE
+from brevitas.inject.enum import ScalingPerOutputType
+import brevitas.nn as qnn
+from brevitas.proxy.groupwise_int_parameter_quant import GroupwiseWeightQuantProxyFromInjector
+from brevitas.quant.scaled_int import Int8WeightPerTensorFloat
+from brevitas.quant.shifted_scaled_int import ShiftedUint8WeightPerTensorFloat
+
+ZP_BIT_WIDTH = 6
+SCALE_BIT_WIDTH = 5
+
+
+class QuantScalingInt(Int8WeightPerTensorFloat):
+    bit_width = SCALE_BIT_WIDTH
+    module = (this << 1).module
+    tracked_parameter_list = (this << 1).tracked_parameter_list
+    upstream_scaling = (this << 1).scaling_per_output_type
+    rescaling_int_quant = RescalingIntQuant
+
+    @value
+    def scaling_shape(
+            scaling_per_output,
+            scaling_per_output_channel_shape,
+            expanded_groupwise_shape,
+            group_dim,
+            upstream_scaling):
+        if scaling_per_output == ScalingPerOutputType.TENSOR:
+            scaling = SCALAR_SHAPE
+        elif scaling_per_output == ScalingPerOutputType.CHANNEL:
+            scaling = scaling_per_output_channel_shape
+        elif scaling_per_output == ScalingPerOutputType.GROUP:
+            # Scaling shape is like expanded_groupwise_shape but has 1 in position group_dim + 1
+            assert expanded_groupwise_shape is not None, "Per Group scaling not correctly configured"
+            assert group_dim is not None, "Per Group scaling not correctly configured"
+            size = list(expanded_groupwise_shape)
+            size[group_dim + 1] = 1
+            scaling = tuple(size)
+
+        # When quantizing scale of groupwise, there will be one extra dim compared to the normal case
+        if upstream_scaling == ScalingPerOutputType.GROUP:
+            scaling = list(scaling)
+            scaling.insert(-1, 1)
+            scaling = tuple(scaling)
+        return scaling
+
+
+from brevitas.core.zero_point import _ScaleShiftQuantZeroPoint
+
+
+class QuantZPInt(Int8WeightPerTensorFloat):
+    module = (this << 1).module
+    tracked_parameter_list = (this << 1).tracked_parameter_list
+    upstream_scaling = (this << 1).scaling_per_output_type
+    rescaling_int_quant = RescalingIntQuant
+    bit_width = ZP_BIT_WIDTH
+    quantize_zero_point = True
+    scaling_per_output_type = ScalingPerOutputType.CHANNEL
+
+    @value
+    def scaling_shape(
+            scaling_per_output,
+            scaling_per_output_channel_shape,
+            expanded_groupwise_shape,
+            group_dim,
+            upstream_scaling):
+        if scaling_per_output == ScalingPerOutputType.TENSOR:
+            scaling = SCALAR_SHAPE
+        elif scaling_per_output == ScalingPerOutputType.CHANNEL:
+            scaling = scaling_per_output_channel_shape
+        elif scaling_per_output == ScalingPerOutputType.GROUP:
+            # Scaling shape is like expanded_groupwise_shape but has 1 in position group_dim + 1
+            assert expanded_groupwise_shape is not None, "Per Group scaling not correctly configured"
+            assert group_dim is not None, "Per Group scaling not correctly configured"
+            size = list(expanded_groupwise_shape)
+            size[group_dim + 1] = 1
+            scaling = tuple(size)
+
+        # When quantizing scale of groupwise, there will be one extra dim compared to the normal case
+        if upstream_scaling == ScalingPerOutputType.GROUP:
+            scaling = list(scaling)
+            scaling.insert(-1, 1)
+            scaling = tuple(scaling)
+        return scaling
+
+
+class QuantScaleQuantZPInt8WeightPerTensorFloat(ShiftedUint8WeightPerTensorFloat):
+    proxy_class = GroupwiseWeightQuantProxyFromInjector
+    scaling_int_quant = QuantScalingInt
+    zp_int = QuantZPInt
+    restrict_scaling_impl = QuantRestrictValue
+    scaling_per_output_type = ScalingPerOutputType.GROUP
+    scale_shift_zero_point_impl = _ScaleShiftQuantZeroPoint
+    group_size = 32
+
+    @value
+    def restrict_value_float_to_int_impl():
+        return this.scaling_int_quant.rescaling_int_quant
+
+    @value
+    def zp_int_quant():
+        return this.zp_int.rescaling_int_quant
+
+
+def test_quant_scale():
+
+    def hook_scale(module, inp):
+        inp = inp[0]
+        quant_scale, scale, zp, bit_width = module.float_to_int_impl(inp)
+        assert bit_width == SCALE_BIT_WIDTH
+        assert torch.allclose(quant_scale / scale, torch.round(quant_scale / scale))
+
+    def hook_zp(module, inp):
+        inp = inp[0]
+        quant_scale, scale, zp, bit_width = module.zp_int_quant(inp)
+        assert bit_width == ZP_BIT_WIDTH
+        assert torch.allclose(quant_scale / scale, torch.round(quant_scale / scale))
+
+    linear = qnn.QuantLinear(64, 768, weight_quant=QuantScaleQuantZPInt8WeightPerTensorFloat)
+    for module in linear.modules():
+        if isinstance(module, QuantRestrictValue):
+            module.register_forward_pre_hook(hook_scale)
+    for module in linear.modules():
+        if isinstance(module, _ScaleShiftQuantZeroPoint):
+            module.register_forward_pre_hook(hook_zp)
+
+    linear(torch.randn(1, 64))