Update quant params structure

math_model.py

@@ -21,8 +21,10 @@ class QuantLinear(nn.Module):
         self.linear = nn.Linear(in_ch, out_ch)
         weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
         weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
+        assert quant_param['weight_zp_dtype'] == 'torch.int8', f"Weight Zero-Point dtype should be 'torch.int8', found: {quant_param['weight_zp_dype']}"
         input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
         input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
+        assert quant_param['input_zp_dtype'] == 'torch.int8', f"Input Zero-Point dtype should be 'torch.int8', found: {quant_param['input_zp_dype']}"
         self.register_buffer('weight_scale', weight_scale)
         self.register_buffer('weight_zp', weight_zp)
         self.register_buffer('input_scale', input_scale)
@@ -31,9 +33,10 @@ class QuantLinear(nn.Module):
     # I.e., "fake quantization"
     def qdq_forward(self, x):
         scaled_x = x * self.mul_factor
-        quant_weight = quantize(self.linear.weight, self.weight_scale, self.weight_zp, is_asym=True)
+        weight_zp_uint8 = (self.weight_zp + 128).to(torch.uint8).to(torch.float32)
+        quant_weight = quantize(self.linear.weight, self.weight_scale, weight_zp_uint8, is_asym=True)
         quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
-        dequantized_weight = dequantize(quant_weight, self.weight_scale, self.weight_zp)
+        dequantized_weight = dequantize(quant_weight, self.weight_scale, weight_zp_uint8)
         dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
         out = torch.nn.functional.linear(dequantized_input, dequantized_weight, self.linear.bias)
         return out
@@ -47,12 +50,11 @@ class QuantLinear(nn.Module):
         #  - multiply this sum with every weight zero-point (e.g., `torch.sum(quant_input, dim=-1) * self.weight_zp`
         #  - Subtract from previous output (e.g., `quant_output -= torch.sum(quant_input, dim=-1) * self.weight_zp`)
         #  - All other code is just to make sure the broadcasting semantics work correctly
-        weight_zp_int8 = (self.weight_zp - 128).to(torch.int8).to(torch.float32) # Conversion from uint8 -> int8, can be computed offline
-        quant_weight = quantize(self.linear.weight, self.weight_scale, weight_zp_int8, is_asym=False).to(torch.int8)
+        quant_weight = quantize(self.linear.weight, self.weight_scale, self.weight_zp, is_asym=False).to(torch.int8)
         fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
         quant_input = quantize(x, fused_input_scale, self.input_zp, is_asym=False).to(torch.int8)
         quant_output = torch.nn.functional.linear(quant_input.to(torch.float32), quant_weight.to(torch.float32), None).to(torch.int32) # Convert inputs to FP32 to avoid F.linear quantizing the output to int8
-        correction = torch.sum(quant_input, dim=-1, keepdim=True).to(torch.int32) * weight_zp_int8.to(torch.int8).view([1]*(quant_input.ndim-1) + [self.weight_zp.nelement()]) # Correct for weight zero-point
+        correction = torch.sum(quant_input, dim=-1, keepdim=True).to(torch.int32) * self.weight_zp.to(torch.int8).view([1]*(quant_input.ndim-1) + [self.weight_zp.nelement()]) # Correct for weight zero-point
         quant_output = quant_output - correction
         output = dequantize(quant_output, (self.weight_scale * self.input_scale).view([1]*(quant_output.ndim-1) + [(self.weight_scale * self.input_scale).nelement()]), 0.0)
         output += self.linear.bias
@@ -72,8 +74,10 @@ class QuantConv2d(nn.Module):
         self.conv2d = nn.Conv2d(in_ch, out_ch, kernel_size)
         weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
         weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
+        assert quant_param['weight_zp_dtype'] == 'torch.int8', f"Weight Zero-Point dtype should be 'torch.int8', found: {quant_param['weight_zp_dype']}"
         input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
         input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
+        assert quant_param['input_zp_dtype'] == 'torch.int8', f"Input Zero-Point dtype should be 'torch.int8', found: {quant_param['input_zp_dype']}"
         self.register_buffer('weight_scale', weight_scale)
         self.register_buffer('weight_zp', weight_zp)
         self.register_buffer('input_scale', input_scale)
@@ -82,9 +86,10 @@ class QuantConv2d(nn.Module):
     # I.e., "fake quantization"
     def qdq_forward(self, x):
         scaled_x = x * self.mul_factor
-        quant_weight = quantize(self.conv2d.weight, self.weight_scale, self.weight_zp, is_asym=True)
+        weight_zp_uint8 = (self.weight_zp + 128).to(torch.uint8).to(torch.float32)
+        quant_weight = quantize(self.conv2d.weight, self.weight_scale, weight_zp_uint8, is_asym=True)
         quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
-        dequantized_weight = dequantize(quant_weight, self.weight_scale, self.weight_zp)
+        dequantized_weight = dequantize(quant_weight, self.weight_scale, weight_zp_uint8)
         dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
         out = torch.nn.functional.conv2d(dequantized_input, dequantized_weight, self.conv2d.bias)
         return out
@@ -104,8 +109,7 @@ class QuantConv2d(nn.Module):
         #  - multiply this sum with every weight zero-point (e.g., `sum * self.weight_zp`
         #  - Subtract from previous output (e.g., `quant_output -= sum * self.weight_zp`)
         #  - All other code is just to make sure the broadcasting semantics work correctly
-        weight_zp_int8 = (self.weight_zp - 128).to(torch.int8).to(torch.float32) # Conversion from uint8 -> int8, can be computed offline
-        quant_weight = quantize(self.conv2d.weight, self.weight_scale, weight_zp_int8, is_asym=False).to(torch.int8)
+        quant_weight = quantize(self.conv2d.weight, self.weight_scale, self.weight_zp, is_asym=False).to(torch.int8)
         b_shape = list(quant_weight.shape) # Used for weight zero-point correction
         b_shape[0] = 1 # Used for weight zero-point correction
         weight_cat = torch.ones((1,1,1,1)).broadcast_to(b_shape).to(torch.int8) # Used for weight zero-point correction
@@ -113,7 +117,7 @@ class QuantConv2d(nn.Module):
         fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
         quant_input = quantize(x, fused_input_scale, self.input_zp, is_asym=False).to(torch.int8)
         quant_output = torch.nn.functional.conv2d(quant_input.to(torch.float32), quant_weight.to(torch.float32), None).to(torch.int32) # Convert inputs to FP32 to avoid F.conv2d quantizing the output to int8
-        correction = quant_output[:,-1,:,:] * weight_zp_int8.to(torch.int8).view([1, self.weight_zp.nelement()] +  [1]*(quant_output.ndim-2)) # Correct zero-point for weight
+        correction = quant_output[:,-1,:,:] * self.weight_zp.to(torch.int8).view([1, self.weight_zp.nelement()] +  [1]*(quant_output.ndim-2)) # Correct zero-point for weight
         quant_output = quant_output[:,:-1,:,:] - correction
         output = dequantize(quant_output, (self.weight_scale * self.input_scale).view([1, (self.weight_scale * self.input_scale).nelement()] + [1]*(quant_output.ndim-2)), 0.0)
         output += self.conv2d.bias.view([1, self.conv2d.bias.nelement()] + [1]*(quant_output.ndim-2))

test_quant_conv2d.py

@@ -20,12 +20,14 @@ quant_params = {
     'weight_scale': torch.rand((out_ch,)),
     'weight_scale': torch.max(torch.abs(torch.flatten(l.weight, start_dim=1)), dim=1).values / 128.,
     'weight_scale_shape': (out_ch,1,1,1),
-    'weight_zp': torch.clamp(torch.round((torch.mean((l.weight), dim=(1,2,3))) * (128 / torch.max(torch.abs(torch.flatten(l.weight, start_dim=1)), dim=1).values)) + 128, 0, 255),
+    'weight_zp': torch.clamp(torch.round((torch.mean((l.weight), dim=(1,2,3))) * (128 / torch.max(torch.abs(torch.flatten(l.weight, start_dim=1)), dim=1).values)), -128, 127),
     'weight_zp_shape': (out_ch,1,1,1),
+    'weight_zp_dtype': 'torch.int8',
     'input_scale': torch.max(torch.abs(i)) / 128.,
     'input_scale_shape': tuple(),
     'input_zp': torch.zeros((1,)),
     'input_zp_shape': tuple(),
+    'input_zp_dtype': 'torch.int8',
 }
 
 print(quant_params)

test_quant_linear.py

@@ -16,12 +16,14 @@ quant_params = {
     'smoothquant_mul_shape': (1,in_ch),
     'weight_scale': torch.max(torch.abs(l.weight), dim=1).values / 128.,
     'weight_scale_shape': (out_ch,1),
-    'weight_zp': torch.clamp(torch.round((torch.mean((l.weight), dim=1)) * (128 / torch.max(torch.abs(l.weight), dim=1).values)) + 128, 0, 255),
+    'weight_zp': torch.clamp(torch.round((torch.mean((l.weight), dim=1)) * (128 / torch.max(torch.abs(l.weight), dim=1).values)), -128, 127),
     'weight_zp_shape': (out_ch,1),
+    'weight_zp_dtype': 'torch.int8',
     'input_scale': torch.max(torch.abs(i)) / 128.,
     'input_scale_shape': tuple(),
     'input_zp': torch.zeros((1,)),
     'input_zp_shape': tuple(),
+    'input_zp_dtype': 'torch.int8',
 }
 
 print(quant_params)