Update quant params structure
#2
by
nickfraser
- opened
- math_model.py +14 -10
- test_quant_conv2d.py +3 -1
- test_quant_linear.py +3 -1
math_model.py
CHANGED
|
@@ -21,8 +21,10 @@ class QuantLinear(nn.Module):
|
|
| 21 |
self.linear = nn.Linear(in_ch, out_ch)
|
| 22 |
weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
|
| 23 |
weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
|
|
|
|
| 24 |
input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
|
| 25 |
input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
|
|
|
|
| 26 |
self.register_buffer('weight_scale', weight_scale)
|
| 27 |
self.register_buffer('weight_zp', weight_zp)
|
| 28 |
self.register_buffer('input_scale', input_scale)
|
|
@@ -31,9 +33,10 @@ class QuantLinear(nn.Module):
|
|
| 31 |
# I.e., "fake quantization"
|
| 32 |
def qdq_forward(self, x):
|
| 33 |
scaled_x = x * self.mul_factor
|
| 34 |
-
|
|
|
|
| 35 |
quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
|
| 36 |
-
dequantized_weight = dequantize(quant_weight, self.weight_scale,
|
| 37 |
dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
|
| 38 |
out = torch.nn.functional.linear(dequantized_input, dequantized_weight, self.linear.bias)
|
| 39 |
return out
|
|
@@ -47,12 +50,11 @@ class QuantLinear(nn.Module):
|
|
| 47 |
# - multiply this sum with every weight zero-point (e.g., `torch.sum(quant_input, dim=-1) * self.weight_zp`
|
| 48 |
# - Subtract from previous output (e.g., `quant_output -= torch.sum(quant_input, dim=-1) * self.weight_zp`)
|
| 49 |
# - All other code is just to make sure the broadcasting semantics work correctly
|
| 50 |
-
|
| 51 |
-
quant_weight = quantize(self.linear.weight, self.weight_scale, weight_zp_int8, is_asym=False).to(torch.int8)
|
| 52 |
fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
|
| 53 |
quant_input = quantize(x, fused_input_scale, self.input_zp, is_asym=False).to(torch.int8)
|
| 54 |
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
|
| 55 |
-
correction = torch.sum(quant_input, dim=-1, keepdim=True).to(torch.int32) *
|
| 56 |
quant_output = quant_output - correction
|
| 57 |
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)
|
| 58 |
output += self.linear.bias
|
|
@@ -72,8 +74,10 @@ class QuantConv2d(nn.Module):
|
|
| 72 |
self.conv2d = nn.Conv2d(in_ch, out_ch, kernel_size)
|
| 73 |
weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
|
| 74 |
weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
|
|
|
|
| 75 |
input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
|
| 76 |
input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
|
|
|
|
| 77 |
self.register_buffer('weight_scale', weight_scale)
|
| 78 |
self.register_buffer('weight_zp', weight_zp)
|
| 79 |
self.register_buffer('input_scale', input_scale)
|
|
@@ -82,9 +86,10 @@ class QuantConv2d(nn.Module):
|
|
| 82 |
# I.e., "fake quantization"
|
| 83 |
def qdq_forward(self, x):
|
| 84 |
scaled_x = x * self.mul_factor
|
| 85 |
-
|
|
|
|
| 86 |
quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
|
| 87 |
-
dequantized_weight = dequantize(quant_weight, self.weight_scale,
|
| 88 |
dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
|
| 89 |
out = torch.nn.functional.conv2d(dequantized_input, dequantized_weight, self.conv2d.bias)
|
| 90 |
return out
|
|
@@ -104,8 +109,7 @@ class QuantConv2d(nn.Module):
|
|
| 104 |
# - multiply this sum with every weight zero-point (e.g., `sum * self.weight_zp`
|
| 105 |
# - Subtract from previous output (e.g., `quant_output -= sum * self.weight_zp`)
|
| 106 |
# - All other code is just to make sure the broadcasting semantics work correctly
|
| 107 |
-
|
| 108 |
-
quant_weight = quantize(self.conv2d.weight, self.weight_scale, weight_zp_int8, is_asym=False).to(torch.int8)
|
| 109 |
b_shape = list(quant_weight.shape) # Used for weight zero-point correction
|
| 110 |
b_shape[0] = 1 # Used for weight zero-point correction
|
| 111 |
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):
|
|
| 113 |
fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
|
| 114 |
quant_input = quantize(x, fused_input_scale, self.input_zp, is_asym=False).to(torch.int8)
|
| 115 |
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
|
| 116 |
-
correction = quant_output[:,-1,:,:] *
|
| 117 |
quant_output = quant_output[:,:-1,:,:] - correction
|
| 118 |
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)
|
| 119 |
output += self.conv2d.bias.view([1, self.conv2d.bias.nelement()] + [1]*(quant_output.ndim-2))
|
|
|
|
| 21 |
self.linear = nn.Linear(in_ch, out_ch)
|
| 22 |
weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
|
| 23 |
weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
|
| 24 |
+
assert quant_param['weight_zp_dtype'] == 'torch.int8', f"Weight Zero-Point dtype should be 'torch.int8', found: {quant_param['weight_zp_dype']}"
|
| 25 |
input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
|
| 26 |
input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
|
| 27 |
+
assert quant_param['input_zp_dtype'] == 'torch.int8', f"Input Zero-Point dtype should be 'torch.int8', found: {quant_param['input_zp_dype']}"
|
| 28 |
self.register_buffer('weight_scale', weight_scale)
|
| 29 |
self.register_buffer('weight_zp', weight_zp)
|
| 30 |
self.register_buffer('input_scale', input_scale)
|
|
|
|
| 33 |
# I.e., "fake quantization"
|
| 34 |
def qdq_forward(self, x):
|
| 35 |
scaled_x = x * self.mul_factor
|
| 36 |
+
weight_zp_uint8 = (self.weight_zp + 128).to(torch.uint8).to(torch.float32)
|
| 37 |
+
quant_weight = quantize(self.linear.weight, self.weight_scale, weight_zp_uint8, is_asym=True)
|
| 38 |
quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
|
| 39 |
+
dequantized_weight = dequantize(quant_weight, self.weight_scale, weight_zp_uint8)
|
| 40 |
dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
|
| 41 |
out = torch.nn.functional.linear(dequantized_input, dequantized_weight, self.linear.bias)
|
| 42 |
return out
|
|
|
|
| 50 |
# - multiply this sum with every weight zero-point (e.g., `torch.sum(quant_input, dim=-1) * self.weight_zp`
|
| 51 |
# - Subtract from previous output (e.g., `quant_output -= torch.sum(quant_input, dim=-1) * self.weight_zp`)
|
| 52 |
# - All other code is just to make sure the broadcasting semantics work correctly
|
| 53 |
+
quant_weight = quantize(self.linear.weight, self.weight_scale, self.weight_zp, is_asym=False).to(torch.int8)
|
|
|
|
| 54 |
fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
|
| 55 |
quant_input = quantize(x, fused_input_scale, self.input_zp, is_asym=False).to(torch.int8)
|
| 56 |
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
|
| 57 |
+
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
|
| 58 |
quant_output = quant_output - correction
|
| 59 |
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)
|
| 60 |
output += self.linear.bias
|
|
|
|
| 74 |
self.conv2d = nn.Conv2d(in_ch, out_ch, kernel_size)
|
| 75 |
weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
|
| 76 |
weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
|
| 77 |
+
assert quant_param['weight_zp_dtype'] == 'torch.int8', f"Weight Zero-Point dtype should be 'torch.int8', found: {quant_param['weight_zp_dype']}"
|
| 78 |
input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
|
| 79 |
input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
|
| 80 |
+
assert quant_param['input_zp_dtype'] == 'torch.int8', f"Input Zero-Point dtype should be 'torch.int8', found: {quant_param['input_zp_dype']}"
|
| 81 |
self.register_buffer('weight_scale', weight_scale)
|
| 82 |
self.register_buffer('weight_zp', weight_zp)
|
| 83 |
self.register_buffer('input_scale', input_scale)
|
|
|
|
| 86 |
# I.e., "fake quantization"
|
| 87 |
def qdq_forward(self, x):
|
| 88 |
scaled_x = x * self.mul_factor
|
| 89 |
+
weight_zp_uint8 = (self.weight_zp + 128).to(torch.uint8).to(torch.float32)
|
| 90 |
+
quant_weight = quantize(self.conv2d.weight, self.weight_scale, weight_zp_uint8, is_asym=True)
|
| 91 |
quant_input = quantize(scaled_x, self.input_scale, self.input_zp, is_asym=False)
|
| 92 |
+
dequantized_weight = dequantize(quant_weight, self.weight_scale, weight_zp_uint8)
|
| 93 |
dequantized_input = dequantize(quant_input, self.input_scale, self.input_zp)
|
| 94 |
out = torch.nn.functional.conv2d(dequantized_input, dequantized_weight, self.conv2d.bias)
|
| 95 |
return out
|
|
|
|
| 109 |
# - multiply this sum with every weight zero-point (e.g., `sum * self.weight_zp`
|
| 110 |
# - Subtract from previous output (e.g., `quant_output -= sum * self.weight_zp`)
|
| 111 |
# - All other code is just to make sure the broadcasting semantics work correctly
|
| 112 |
+
quant_weight = quantize(self.conv2d.weight, self.weight_scale, self.weight_zp, is_asym=False).to(torch.int8)
|
|
|
|
| 113 |
b_shape = list(quant_weight.shape) # Used for weight zero-point correction
|
| 114 |
b_shape[0] = 1 # Used for weight zero-point correction
|
| 115 |
weight_cat = torch.ones((1,1,1,1)).broadcast_to(b_shape).to(torch.int8) # Used for weight zero-point correction
|
|
|
|
| 117 |
fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
|
| 118 |
quant_input = quantize(x, fused_input_scale, self.input_zp, is_asym=False).to(torch.int8)
|
| 119 |
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
|
| 120 |
+
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
|
| 121 |
quant_output = quant_output[:,:-1,:,:] - correction
|
| 122 |
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)
|
| 123 |
output += self.conv2d.bias.view([1, self.conv2d.bias.nelement()] + [1]*(quant_output.ndim-2))
|
test_quant_conv2d.py
CHANGED
|
@@ -20,12 +20,14 @@ quant_params = {
|
|
| 20 |
'weight_scale': torch.rand((out_ch,)),
|
| 21 |
'weight_scale': torch.max(torch.abs(torch.flatten(l.weight, start_dim=1)), dim=1).values / 128.,
|
| 22 |
'weight_scale_shape': (out_ch,1,1,1),
|
| 23 |
-
'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))
|
| 24 |
'weight_zp_shape': (out_ch,1,1,1),
|
|
|
|
| 25 |
'input_scale': torch.max(torch.abs(i)) / 128.,
|
| 26 |
'input_scale_shape': tuple(),
|
| 27 |
'input_zp': torch.zeros((1,)),
|
| 28 |
'input_zp_shape': tuple(),
|
|
|
|
| 29 |
}
|
| 30 |
|
| 31 |
print(quant_params)
|
|
|
|
| 20 |
'weight_scale': torch.rand((out_ch,)),
|
| 21 |
'weight_scale': torch.max(torch.abs(torch.flatten(l.weight, start_dim=1)), dim=1).values / 128.,
|
| 22 |
'weight_scale_shape': (out_ch,1,1,1),
|
| 23 |
+
'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),
|
| 24 |
'weight_zp_shape': (out_ch,1,1,1),
|
| 25 |
+
'weight_zp_dtype': 'torch.int8',
|
| 26 |
'input_scale': torch.max(torch.abs(i)) / 128.,
|
| 27 |
'input_scale_shape': tuple(),
|
| 28 |
'input_zp': torch.zeros((1,)),
|
| 29 |
'input_zp_shape': tuple(),
|
| 30 |
+
'input_zp_dtype': 'torch.int8',
|
| 31 |
}
|
| 32 |
|
| 33 |
print(quant_params)
|
test_quant_linear.py
CHANGED
|
@@ -16,12 +16,14 @@ quant_params = {
|
|
| 16 |
'smoothquant_mul_shape': (1,in_ch),
|
| 17 |
'weight_scale': torch.max(torch.abs(l.weight), dim=1).values / 128.,
|
| 18 |
'weight_scale_shape': (out_ch,1),
|
| 19 |
-
'weight_zp': torch.clamp(torch.round((torch.mean((l.weight), dim=1)) * (128 / torch.max(torch.abs(l.weight), dim=1).values))
|
| 20 |
'weight_zp_shape': (out_ch,1),
|
|
|
|
| 21 |
'input_scale': torch.max(torch.abs(i)) / 128.,
|
| 22 |
'input_scale_shape': tuple(),
|
| 23 |
'input_zp': torch.zeros((1,)),
|
| 24 |
'input_zp_shape': tuple(),
|
|
|
|
| 25 |
}
|
| 26 |
|
| 27 |
print(quant_params)
|
|
|
|
| 16 |
'smoothquant_mul_shape': (1,in_ch),
|
| 17 |
'weight_scale': torch.max(torch.abs(l.weight), dim=1).values / 128.,
|
| 18 |
'weight_scale_shape': (out_ch,1),
|
| 19 |
+
'weight_zp': torch.clamp(torch.round((torch.mean((l.weight), dim=1)) * (128 / torch.max(torch.abs(l.weight), dim=1).values)), -128, 127),
|
| 20 |
'weight_zp_shape': (out_ch,1),
|
| 21 |
+
'weight_zp_dtype': 'torch.int8',
|
| 22 |
'input_scale': torch.max(torch.abs(i)) / 128.,
|
| 23 |
'input_scale_shape': tuple(),
|
| 24 |
'input_zp': torch.zeros((1,)),
|
| 25 |
'input_zp_shape': tuple(),
|
| 26 |
+
'input_zp_dtype': 'torch.int8',
|
| 27 |
}
|
| 28 |
|
| 29 |
print(quant_params)
|