File size: 11,596 Bytes
1a7c8cc
 
a7340fc
1a7c8cc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
import torch.nn as nn
import torch 
from .configuration_mamba import MambaConfig
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.modeling_utils import PreTrainedModel
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast
import math
import json
import torch
import torch.nn as nn
import torch.nn.functional as F
from dataclasses import dataclass
from einops import rearrange, repeat, einsum
from typing import Optional , Union ,Tuple

# Dear contributors of the https://github.com/johnma2006/mamba-minimal/tree/master repository, special thanks to Albert Gu and Tri Dao for their articles. (https://arxiv.org/abs/2312.00752)


class MambaRMSNorm(nn.Module):
    def __init__(self,
                 d_model: int,
                 eps: float = 1e-5):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(d_model))
    def forward(self, x):
        output = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight
        return output
    

class MambaBlock(nn.Module):
    def __init__(self, config: MambaConfig):
        """A single Mamba block, as described in Figure 3 in Section 3.4 in the Mamba paper [1]."""
        super().__init__()
        self.config = config

        self.in_proj = nn.Linear(config.d_model, config.d_inner * 2, bias=config.bias)

        self.conv1d = nn.Conv1d(
            in_channels=config.d_inner,
            out_channels=config.d_inner,
            bias=config.conv_bias,
            kernel_size=config.d_conv,
            groups=config.d_inner,
            padding=config.d_conv - 1,
        )

        # x_proj takes in `x` and outputs the input-specific Δ, B, C
        self.x_proj = nn.Linear(config.d_inner, config.dt_rank + config.d_state * 2, bias=False)
        
        # dt_proj projects Δ from dt_rank to d_in
        self.dt_proj = nn.Linear(config.dt_rank, config.d_inner, bias=True)

        A = repeat(torch.arange(1, config.d_state + 1), 'n -> d n', d=config.d_inner)
        self.A_log = nn.Parameter(torch.log(A))
        self.D = nn.Parameter(torch.ones(config.d_inner))
        self.out_proj = nn.Linear(config.d_inner, config.d_model, bias=config.bias)
        self.norm = MambaRMSNorm(config.d_model)

    def forward(self, x):
        """Mamba block forward. This looks the same as Figure 3 in Section 3.4 in the Mamba paper [1].
    
        Args:
            x: shape (b, l, d)    (See Glossary at top for definitions of b, l, d_in, n...)
    
        Returns:
            output: shape (b, l, d)
        
        Official Implementation:
            class Mamba, https://github.com/state-spaces/mamba/blob/main/mamba_ssm/modules/mamba_simple.py#L119
            mamba_inner_ref(), https://github.com/state-spaces/mamba/blob/main/mamba_ssm/ops/selective_scan_interface.py#L311
            
        """

        (b, l, d) = x.shape
        x_copy = x # There was a separate class for residual, I deleted that part and added it here.
        x = self.norm(x)
        x_and_res = self.in_proj(x)  # shape (b, l, 2 * d_in)
        (x, res) = x_and_res.split(split_size=[self.config.d_inner, self.config.d_inner], dim=-1)

        x = rearrange(x, 'b l d_in -> b d_in l')
        x = self.conv1d(x)[:, :, :l]
        x = rearrange(x, 'b d_in l -> b l d_in')
        
        x = F.silu(x)

        y = self.ssm(x)
        
        y = y * F.silu(res)
        
        output = self.out_proj(y) + x_copy

        return output

    
    def ssm(self, x):
        """Runs the SSM. See:
            - Algorithm 2 in Section 3.2 in the Mamba paper [1]
            - run_SSM(A, B, C, u) in The Annotated S4 [2]

        Args:
            x: shape (b, l, d_in)    (See Glossary at top for definitions of b, l, d_in, n...)
    
        Returns:
            output: shape (b, l, d_in)

        Official Implementation:
            mamba_inner_ref(), https://github.com/state-spaces/mamba/blob/main/mamba_ssm/ops/selective_scan_interface.py#L311
            
        """
        (d_in, n) = self.A_log.shape

        # Compute ∆ A B C D, the state space parameters.
        #     A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
        #     ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
        #                                  and is why Mamba is called **selective** state spaces)
        
        A = -torch.exp(self.A_log.float())  # shape (d_in, n)
        D = self.D.float()

        x_dbl = self.x_proj(x)  # (b, l, dt_rank + 2*n)
        
        (delta, B, C) = x_dbl.split(split_size=[self.config.dt_rank, n, n], dim=-1)  # delta: (b, l, dt_rank). B, C: (b, l, n)
        delta = F.softplus(self.dt_proj(delta))  # (b, l, d_in)
        
        y = self.selective_scan(x, delta, A, B, C, D)  # This is similar to run_SSM(A, B, C, u) in The Annotated S4 [2]
        
        return y

    
    def selective_scan(self, u, delta, A, B, C, D):
        """Does selective scan algorithm. See:
            - Section 2 State Space Models in the Mamba paper [1]
            - Algorithm 2 in Section 3.2 in the Mamba paper [1]
            - run_SSM(A, B, C, u) in The Annotated S4 [2]

        This is the classic discrete state space formula:
            x(t + 1) = Ax(t) + Bu(t)
            y(t)     = Cx(t) + Du(t)
        except B and C (and the step size delta, which is used for discretization) are dependent on the input x(t).
    
        Args:
            u: shape (b, l, d_in)    (See Glossary at top for definitions of b, l, d_in, n...)
            delta: shape (b, l, d_in)
            A: shape (d_in, n)
            B: shape (b, l, n)
            C: shape (b, l, n)
            D: shape (d_in,)
    
        Returns:
            output: shape (b, l, d_in)
    
        Official Implementation:
            selective_scan_ref(), https://github.com/state-spaces/mamba/blob/main/mamba_ssm/ops/selective_scan_interface.py#L86
            Note: I refactored some parts out of `selective_scan_ref` out, so the functionality doesn't match exactly.
            
        """
        (b, l, d_in) = u.shape
        n = A.shape[1]
        
        # Discretize continuous parameters (A, B)
        # - A is discretized using zero-order hold (ZOH) discretization (see Section 2 Equation 4 in the Mamba paper [1])
        # - B is discretized using a simplified Euler discretization instead of ZOH. From a discussion with authors:
        #   "A is the more important term and the performance doesn't change much with the simplication on B"
        deltaA = torch.exp(einsum(delta, A, 'b l d_in, d_in n -> b d_in l n'))
        deltaB_u = einsum(delta, B, u, 'b l d_in, b l n, b l d_in -> b d_in l n')
        
        # Perform selective scan (see scan_SSM() in The Annotated S4 [2])
        x = torch.zeros((b, d_in, n), device=deltaA.device)
        ys = []    
        for i in range(l):
            x = deltaA[:, :, i] * x + deltaB_u[:, :, i]
            y = einsum(x, C[:, i, :], 'b d_in n, b n -> b d_in')
            ys.append(y)
        y = torch.stack(ys, dim=1)  # shape (b, l, d_in)
        
        y = y + u * D
    
        return y
    
class MambaPreTrainedModel(PreTrainedModel):
    config_class = MambaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MambaBlock"]

    def _init_weights(self, module):
        std = 0.02
        if isinstance(module, (nn.Linear, nn.Conv1d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

class MambaModel(MambaPreTrainedModel):
    def __init__(self, config: MambaConfig):
        """Full Mamba model.
    Mamba model decoder consisting of *config.n_layer* layers. Each layer is a [`MambaBlock`]

    Args:
        config: MambaConfig
    """
        super().__init__(config)
        self.config = config
        
        self.embedding = nn.Embedding(config.vocab_size, config.d_model)
        self.layers = nn.ModuleList([MambaBlock(config) for _ in range(config.n_layer)])
        self.norm_f = MambaRMSNorm(config.d_model)

        self.gradient_checkpointing = False
        self.post_init()

    def get_input_embeddings(self):
        return self.embedding

    def set_input_embeddings(self, value):
        self.embedding = value

    def forward(self,
                input_ids: torch.LongTensor = None,
                return_dict: Optional[bool] = None,
                )-> Union[Tuple, BaseModelOutputWithPast]:
        x = self.embedding(input_ids)
        all_hidden_states = list()
        for layer in self.layers:
            x = layer(x)
            all_hidden_states.append(x)
            
        hidden_states = self.norm_f(x)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            hidden_states=all_hidden_states,
        )
class MambaForCausalLM(MambaPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = MambaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
        self.lm_head.weight = self.model.embedding.weight
        self.post_init()

    def get_input_embeddings(self):
        return self.model.embedding

    def set_input_embeddings(self, value):
        self.model.embedding = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model
    
    def forward(self,
                input_ids: torch.LongTensor = None,
                labels: Optional[torch.LongTensor] = None,
                output_attentions: Optional[bool] = None,
                output_hidden_states: Optional[bool] = None,
                return_dict: Optional[bool] = None,
                )-> Union[Tuple, CausalLMOutputWithPast]:
        outputs = self.model(
            input_ids=input_ids,
            return_dict=return_dict,
        )
        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)
        logits = logits.float()
        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
        )
    
    def prepare_inputs_for_generation(
        self, input_ids, **kwargs
    ):
        model_inputs = {"input_ids": input_ids}
        return model_inputs