Create math_model.py
Browse files- math_model.py +143 -0
math_model.py
ADDED
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch
|
3 |
+
|
4 |
+
def quantize_fp8(tensor: torch.Tensor, scale: torch.Tensor):
|
5 |
+
dtype = tensor.dtype
|
6 |
+
clamp_min, clamp_max = torch.tensor(-240., dtype=dtype), torch.tensor(240., dtype=dtype)
|
7 |
+
quant_tensor = torch.clamp((tensor/scale), clamp_min, clamp_max).to(torch.float8_e4m3fnuz).to(dtype)
|
8 |
+
return quant_tensor
|
9 |
+
|
10 |
+
def dequantize_fp8(tensor: torch.Tensor, scale: torch.Tensor):
|
11 |
+
return tensor * scale
|
12 |
+
|
13 |
+
|
14 |
+
class QuantLinear(nn.Module):
|
15 |
+
def __init__(self, in_ch, out_ch, quant_param):
|
16 |
+
super().__init__()
|
17 |
+
mul_factor = torch.tensor(quant_param['smoothquant_mul']).view(quant_param['smoothquant_mul_shape'])
|
18 |
+
self.register_buffer('mul_factor', mul_factor)
|
19 |
+
self.linear = nn.Linear(in_ch, out_ch)
|
20 |
+
weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
|
21 |
+
# weight_zp = torch.tensor(quant_param['weight_zp']).view(quant_param['weight_zp_shape'])
|
22 |
+
# assert quant_param['weight_zp_dtype'] == 'torch.float8_e4m3fnuz', f"Weight Zero-Point dtype should be 'torch.float8_e4m3fnuz', found: {quant_param['weight_zp_dype']}"
|
23 |
+
input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
|
24 |
+
input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
|
25 |
+
assert quant_param['input_zp_dtype'] == 'torch.float8_e4m3fnuz', f"Input Zero-Point dtype should be 'torch.float8_e4m3fnuz', found: {quant_param['input_zp_dype']}"
|
26 |
+
self.register_buffer('weight_scale', weight_scale)
|
27 |
+
# self.register_buffer('weight_zp', weight_zp)
|
28 |
+
self.register_buffer('input_scale', input_scale)
|
29 |
+
self.register_buffer('input_zp', input_zp)
|
30 |
+
|
31 |
+
# I.e., "fake quantization"
|
32 |
+
def qdq_forward(self, x):
|
33 |
+
print(self.mul_factor.shape)
|
34 |
+
scaled_x = x * self.mul_factor
|
35 |
+
quant_weight = quantize_fp8(self.linear.weight, self.weight_scale)
|
36 |
+
quant_input = quantize_fp8(scaled_x, self.input_scale)
|
37 |
+
dequantized_weight = dequantize_fp8(quant_weight, self.weight_scale)
|
38 |
+
dequantized_input = dequantize_fp8(quant_input, self.input_scale)
|
39 |
+
out = torch.nn.functional.linear(dequantized_input, dequantized_weight, self.linear.bias)
|
40 |
+
return out
|
41 |
+
|
42 |
+
# Accelerated version
|
43 |
+
def qop_forward(self, x):
|
44 |
+
quant_weight = quantize_fp8(self.linear.weight, self.weight_scale).to(torch.float8_e4m3fnuz)
|
45 |
+
fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
|
46 |
+
quant_input = quantize_fp8(x, fused_input_scale).to(torch.float8_e4m3fnuz)
|
47 |
+
quant_output = torch.nn.functional.linear(quant_input.to(torch.float32), quant_weight.to(torch.float32), None).to(torch.float32) # Convert inputs to FP32 to avoid F.linear quantizing the output to int8
|
48 |
+
output = dequantize_fp8(quant_output, (self.weight_scale * self.input_scale).view([1]*(quant_output.ndim-1) + [(self.weight_scale * self.input_scale).nelement()]))
|
49 |
+
output += self.linear.bias
|
50 |
+
return output
|
51 |
+
|
52 |
+
def forward(self, x, qop=False):
|
53 |
+
if qop:
|
54 |
+
return self.qop_forward(x)
|
55 |
+
else:
|
56 |
+
return self.qdq_forward(x)
|
57 |
+
|
58 |
+
class QuantConv2d(nn.Module):
|
59 |
+
def __init__(self, in_ch, out_ch, kernel_size, quant_param):
|
60 |
+
super().__init__()
|
61 |
+
mul_factor = torch.tensor(quant_param['smoothquant_mul']).view(quant_param['smoothquant_mul_shape'])
|
62 |
+
self.register_buffer('mul_factor', mul_factor)
|
63 |
+
self.conv2d = nn.Conv2d(in_ch, out_ch, kernel_size)
|
64 |
+
weight_scale = torch.tensor(quant_param['weight_scale']).view(quant_param['weight_scale_shape'])
|
65 |
+
|
66 |
+
input_scale = torch.tensor(quant_param['input_scale']).view(quant_param['input_scale_shape'])
|
67 |
+
input_zp = torch.tensor(quant_param['input_zp']).view(quant_param['input_zp_shape'])
|
68 |
+
assert quant_param['input_zp_dtype'] == 'torch.float8_e4m3fnuz', f"Input Zero-Point dtype should be 'torch.float8_e4m3fnuz', found: {quant_param['input_zp_dype']}"
|
69 |
+
self.register_buffer('weight_scale', weight_scale)
|
70 |
+
self.register_buffer('input_scale', input_scale)
|
71 |
+
self.register_buffer('input_zp', input_zp)
|
72 |
+
|
73 |
+
# I.e., "fake quantization"
|
74 |
+
def qdq_forward(self, x):
|
75 |
+
scaled_x = x * self.mul_factor
|
76 |
+
quant_weight = quantize_fp8(self.conv2d.weight, self.weight_scale)
|
77 |
+
quant_input = quantize_fp8(scaled_x, self.input_scale)
|
78 |
+
dequantized_weight = dequantize_fp8(quant_weight, self.weight_scale)
|
79 |
+
dequantized_input = dequantize_fp8(quant_input, self.input_scale)
|
80 |
+
out = torch.nn.functional.conv2d(dequantized_input, dequantized_weight, self.conv2d.bias)
|
81 |
+
return out
|
82 |
+
|
83 |
+
# Accelerated version
|
84 |
+
def qop_forward(self, x):
|
85 |
+
quant_weight = quantize_fp8(self.conv2d.weight, self.weight_scale).to(torch.float8_e4m3fnuz)
|
86 |
+
fused_input_scale = self.input_scale / self.mul_factor # Fuse SmoothQuant and input scales, can be computed offline
|
87 |
+
quant_input = quantize_fp8(x, fused_input_scale).to(torch.float8_e4m3fnuz)
|
88 |
+
quant_output = torch.nn.functional.conv2d(quant_input.to(torch.float32), quant_weight.to(torch.float32), None).to(torch.float32) # Convert inputs to FP32 to avoid F.conv2d quantizing the output to int8
|
89 |
+
output = dequantize_fp8(quant_output, (self.weight_scale * self.input_scale).view([1, (self.weight_scale * self.input_scale).nelement()] + [1]*(quant_output.ndim-2)))
|
90 |
+
output += self.conv2d.bias.view([1, self.conv2d.bias.nelement()] + [1]*(quant_output.ndim-2))
|
91 |
+
return output
|
92 |
+
|
93 |
+
def forward(self, x, qop=False):
|
94 |
+
if qop:
|
95 |
+
return self.qop_forward(x)
|
96 |
+
else:
|
97 |
+
return self.qdq_forward(x)
|
98 |
+
|
99 |
+
|
100 |
+
torch.manual_seed(0)
|
101 |
+
|
102 |
+
batch_size = 1
|
103 |
+
seq_len = 11
|
104 |
+
hidden_size = 21
|
105 |
+
output_size = 36
|
106 |
+
shape = 5
|
107 |
+
query = 2.*torch.rand((batch_size, seq_len, hidden_size)) - 1.
|
108 |
+
conv_input = 2.*torch.rand((batch_size, hidden_size, shape, shape)) - 1.
|
109 |
+
|
110 |
+
quant_params = {
|
111 |
+
"quant_linear": {
|
112 |
+
"smoothquant_mul": torch.randn(hidden_size).abs(),
|
113 |
+
"smoothquant_mul_shape": [1, 1, hidden_size],
|
114 |
+
"input_scale": torch.max(torch.abs(query)) / 240.,
|
115 |
+
"input_scale_shape": [],
|
116 |
+
"input_zp": 0.0,
|
117 |
+
"input_zp_shape": [],
|
118 |
+
"input_zp_dtype": "torch.float8_e4m3fnuz",
|
119 |
+
"weight_scale":torch.randn(output_size).abs(),
|
120 |
+
"weight_scale_shape": [output_size, 1]
|
121 |
+
},
|
122 |
+
"quant_conv": {
|
123 |
+
"smoothquant_mul": torch.randn(hidden_size).abs(),
|
124 |
+
"smoothquant_mul_shape": [1, hidden_size, 1, 1],
|
125 |
+
"input_scale": torch.max(torch.abs(query)) / 240.,
|
126 |
+
"input_scale_shape": [],
|
127 |
+
"input_zp": 0.0,
|
128 |
+
"input_zp_shape": [],
|
129 |
+
"input_zp_dtype": "torch.float8_e4m3fnuz",
|
130 |
+
"weight_scale":torch.randn(output_size).abs(),
|
131 |
+
"weight_scale_shape": [output_size, 1, 1, 1]
|
132 |
+
|
133 |
+
}
|
134 |
+
}
|
135 |
+
|
136 |
+
qlinear = QuantLinear(hidden_size, output_size, quant_params['quant_linear'])
|
137 |
+
o = qlinear(query)
|
138 |
+
q_o = qlinear(query, qop=True)
|
139 |
+
assert torch.allclose(o, q_o)
|
140 |
+
qconv = QuantConv2d(hidden_size, output_size, shape, quant_params['quant_conv'])
|
141 |
+
o = qconv(conv_input)
|
142 |
+
q_o = qconv(conv_input, qop=True)
|
143 |
+
assert torch.allclose(o, q_o, atol=1e-6)
|