File size: 9,112 Bytes
e75cdce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
import bitsandbytes as bnb
from accelerate import init_empty_weights
from bitsandbytes.nn.modules import Params4bit, Int8Params
import torch 

def Params4bitCuda(self, device):
    self.data = self.data.cuda(device)
    self.quant_state[0] = self.quant_state[0].cuda(device)
    self.quant_state[4][0] = self.quant_state[4][0].cuda(device)
    self.quant_state[4][1][0] = self.quant_state[4][1][0].cuda(device)
    self.quant_state[4][1][1] = self.quant_state[4][1][1].cuda(device)

    self.quant_state[6] = self.quant_state[6].cuda(device)
    return self

class Linear4bitOnline(torch.nn.Module):
    def __init__(self, weight, bias, quant_type):
        super().__init__()
        self.weight = Params4bit(
            weight.data, requires_grad=False, compress_statistics=True, quant_type=quant_type
        )
        self.compute_dtype = None
        #self.weight.cuda(weight.device)
        self.bias = bias

    def forward(self, x: torch.Tensor):
        # weights are cast automatically as Int8Params, but the bias has to be cast manually
        if self.bias is not None and self.bias.dtype != x.dtype:
            self.bias.data = self.bias.data.to(x.dtype)

        if getattr(self.weight, "quant_state", None) is None:
            print(
                "FP4 quantization state not initialized. Please call .cuda() or .to(device) on the LinearFP4 layer first."
            )
        inp_dtype = x.dtype
        if self.compute_dtype is not None:
            x = x.to(self.compute_dtype)

        bias = None if self.bias is None else self.bias.to(self.compute_dtype)
        out = bnb.matmul_4bit(
            x, self.weight.t(), bias=bias, quant_state=self.weight.quant_state
        )

        out = out.to(inp_dtype)

        return out
    
class Linear8bitLtOnline(torch.nn.Module):
    def __init__(
        self,
        weight,
        bias,
        has_fp16_weights=True,
        memory_efficient_backward=False,
        threshold=0.0,
        index=None,
    ):
        super().__init__()
        assert (
            not memory_efficient_backward
        ), "memory_efficient_backward is no longer required and the argument is deprecated in 0.37.0 and will be removed in 0.39.0"
        self.state = bnb.MatmulLtState()
        self.index = index

        # Necessary for stacked layers
        self.state.threshold = threshold
        self.state.has_fp16_weights = has_fp16_weights
        self.state.memory_efficient_backward = memory_efficient_backward
        if threshold > 0.0 and not has_fp16_weights:
            self.state.use_pool = True

        self.weight = Int8Params(
            weight.data,
            has_fp16_weights=has_fp16_weights,
            requires_grad=has_fp16_weights,
        )
        self.bias = bias

    def init_8bit_state(self):
        self.state.CB = self.weight.CB
        self.state.SCB = self.weight.SCB
        self.weight.CB = None
        self.weight.SCB = None

    def forward(self, x: torch.Tensor):
        self.state.is_training = self.training
        if self.weight.CB is not None:
            self.init_8bit_state()

        # weights are cast automatically as Int8Params, but the bias has to be cast manually
        if self.bias is not None and self.bias.dtype != x.dtype:
            self.bias.data = self.bias.data.to(x.dtype)
        
        out = bnb.matmul(x, self.weight, bias=self.bias, state=self.state)

        if not self.state.has_fp16_weights:
            if self.state.CB is not None and self.state.CxB is not None:
                # we converted 8-bit row major to turing/ampere format in the first inference pass
                # we no longer need the row-major weight
                del self.state.CB
                self.weight.data = self.state.CxB
        return out
    
def quantize_offline(model, bits: int):
    assert (bits == 4), f'bits: {bits} is not supported'
    
    for i, layer in enumerate(model.model.layers):
        layer.self_attn.W_pack = bnb.nn.Linear4bit(
                            layer.self_attn.W_pack.weight.shape[1],
                            layer.self_attn.W_pack.weight.shape[0],
                            False,
                            torch.float16,
                            compress_statistics=True,
                            quant_type="nf4",
                        )
        layer.self_attn.o_proj = bnb.nn.Linear4bit(
                            layer.self_attn.o_proj.weight.shape[1],
                            layer.self_attn.o_proj.weight.shape[0],
                            False,
                            torch.float16,
                            compress_statistics=True,
                            quant_type="nf4",
                        )

        layer.mlp.gate_proj = bnb.nn.Linear4bit(
                            layer.mlp.gate_proj.weight.shape[1],
                            layer.mlp.gate_proj.weight.shape[0],
                            False,
                            torch.float16,
                            compress_statistics=True,
                            quant_type="nf4",
                        )
        layer.mlp.down_proj = bnb.nn.Linear4bit(
                            layer.mlp.down_proj.weight.shape[1],
                            layer.mlp.down_proj.weight.shape[0],
                            False,
                            torch.float16,
                            compress_statistics=True,
                            quant_type="nf4",
                        )
        layer.mlp.up_proj = bnb.nn.Linear4bit(
                            layer.mlp.up_proj.weight.shape[1],
                            layer.mlp.up_proj.weight.shape[0],
                            False,
                            torch.float16,
                            compress_statistics=True,
                            quant_type="nf4",
                        )
    return model

def quantize_online(model, bits: int):
    def quant(weight, bias=None):
        if bits == 8:
            linear = Linear8bitLtOnline(
                weight,
                bias,
                has_fp16_weights=False,
                threshold=6.0,
            )
            if bias is not None:
                linear.bias = torch.nn.Parameter(bias)
        elif bits == 4:
            linear = Linear4bitOnline(
                weight,
                bias,
                quant_type="nf4", #fp4/nf4
            )
        else:
            raise ValueError("quantize only support 4/8 bit")
        return linear

    for i, layer in enumerate(model.model.layers):
        layer.self_attn.W_pack = quant(layer.self_attn.W_pack.weight)
        layer.self_attn.o_proj = quant(layer.self_attn.o_proj.weight)
        layer.mlp.gate_proj = quant(layer.mlp.gate_proj.weight)
        layer.mlp.down_proj = quant(layer.mlp.down_proj.weight)
        layer.mlp.up_proj = quant(layer.mlp.up_proj.weight)
    return model

def init_model_weight_int4(config, model, state_dict):
    #replace Params4bit.cuda with Params4bitCuda
    Params4bit.cuda = Params4bitCuda

    for i in range(config.num_hidden_layers):
        weight_data = state_dict[f'model.layers.{i}.self_attn.W_pack.weight.data']
        weight_quant_state = state_dict[f'model.layers.{i}.self_attn.W_pack.weight.quant_state']
        model.model.layers[i].self_attn.W_pack.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
        
        weight_data = state_dict[f'model.layers.{i}.self_attn.o_proj.weight.data']
        weight_quant_state = state_dict[f'model.layers.{i}.self_attn.o_proj.weight.quant_state']
        model.model.layers[i].self_attn.o_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
        
        weight_data = state_dict[f'model.layers.{i}.mlp.gate_proj.weight.data']
        weight_quant_state = state_dict[f'model.layers.{i}.mlp.gate_proj.weight.quant_state']
        model.model.layers[i].mlp.gate_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
        
        weight_data = state_dict[f'model.layers.{i}.mlp.up_proj.weight.data']
        weight_quant_state = state_dict[f'model.layers.{i}.mlp.up_proj.weight.quant_state']
        model.model.layers[i].mlp.up_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
        
        weight_data = state_dict[f'model.layers.{i}.mlp.down_proj.weight.data']
        weight_quant_state = state_dict[f'model.layers.{i}.mlp.down_proj.weight.quant_state']
        model.model.layers[i].mlp.down_proj.weight = Params4bit(weight_data, requires_grad=False, quant_state=weight_quant_state)
        
        model.model.layers[i].input_layernorm.weight = state_dict[f'model.layers.{i}.input_layernorm.weight']
        model.model.layers[i].post_attention_layernorm.weight = state_dict[f'model.layers.{i}.post_attention_layernorm.weight']
    
    model.model.embed_tokens.weight = state_dict['model.embed_tokens.weight']
    model.model.norm.weight = state_dict['model.norm.weight']
    model.lm_head.weight = state_dict['lm_head.weight'] 
    return model