File size: 16,661 Bytes
eafbe80
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385

from __future__ import annotations

import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Optional

import torch
from torch import nn
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import ModelOutput, logging
from transformers.utils.deprecation import deprecate_kwarg

from fla.layers.attn import Attention
from fla.layers.mamba import Mamba
from fla.models.mamba.modeling_mamba import MambaCache
from fla.models.samba.configuration_samba import SambaConfig
from fla.models.utils import FLAGenerationMixin
from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, RMSNorm
from fla.modules import GatedMLP as SambaMLP
from fla.modules.l2warp import l2_warp

if TYPE_CHECKING:
    from transformers.processing_utils import Unpack


try:
    from transformers.modeling_layers import GradientCheckpointingLayer
except ImportError:
    from fla.models.modeling_layers import GradientCheckpointingLayer

logger = logging.get_logger(__name__)


class SambaBlock(GradientCheckpointingLayer):

    def __init__(self, config, layer_idx):
        super().__init__()

        self.config = config
        self.layer_idx = layer_idx

        self.mixer_norm = RMSNorm(hidden_size=config.hidden_size, eps=config.norm_eps)
        if config.attn is not None and layer_idx in config.attn['layers']:
            self.mixer = Attention(
                hidden_size=config.hidden_size,
                num_heads=config.attn['num_heads'],
                num_kv_heads=config.attn['num_kv_heads'],
                qkv_bias=config.attn['qkv_bias'],
                window_size=config.attn['window_size'],
                rope_theta=config.attn['rope_theta'],
                max_position_embeddings=config.max_position_embeddings,
                layer_idx=layer_idx,
            )
        else:
            self.mixer = Mamba(
                hidden_size=config.hidden_size,
                state_size=config.state_size,
                conv_kernel=config.conv_kernel,
                intermediate_size=config.intermediate_size,
                time_step_rank=config.time_step_rank,
                use_bias=config.use_bias,
                layer_idx=layer_idx,
            )
        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
        self.mlp = SambaMLP(
            hidden_size=config.hidden_size,
            hidden_ratio=config.hidden_ratio,
            hidden_act=config.hidden_act,
            fuse_swiglu=config.fuse_swiglu,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: tuple[torch.Tensor] | None = None,
        **kwargs: Unpack[dict],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:

        residual = hidden_states
        hidden_states = self.mixer_norm(hidden_states)
        if isinstance(self.mixer, Mamba):
            hidden_states = self.mixer(hidden_states, cache_params=cache_params, **kwargs)
        else:
            hidden_states, _, cache_params = self.mixer(hidden_states=hidden_states, past_key_values=cache_params, **kwargs)
        if self.config.fuse_norm:
            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
        else:
            hidden_states = residual + hidden_states
            residual = hidden_states
            hidden_states = self.mlp_norm(hidden_states)
        hidden_states = self.mlp(hidden_states, **kwargs)
        hidden_states = residual + hidden_states
        return hidden_states


class SambaPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = SambaConfig
    base_model_prefix = "backbone"
    _no_split_modules = ["SambaBlock"]
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                if not getattr(module.bias, "_no_reinit", False):
                    nn.init.zeros_(module.bias)
        elif isinstance(module, Mamba):
            module.A_log._no_weight_decay = True
            module.D._no_weight_decay = True

            dt_init_std = self.config.time_step_rank**-0.5 * self.config.time_step_scale
            if self.config.time_step_init_scheme == "constant":
                nn.init.constant_(module.dt_proj.weight, dt_init_std)
            elif self.config.time_step_init_scheme == "random":
                nn.init.uniform_(module.dt_proj.weight, -dt_init_std, dt_init_std)

            dt = torch.exp(
                torch.rand(self.config.intermediate_size)
                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
                + math.log(self.config.time_step_min),
            ).clamp(min=self.config.time_step_floor)
            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
            inv_dt = dt + torch.log(-torch.expm1(-dt))
            with torch.no_grad():
                module.dt_proj.bias.data = nn.Parameter(inv_dt.to(module.dt_proj.bias.device))
            module.dt_proj.bias._no_reinit = True
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, std=self.config.initializer_range)
        elif hasattr(module, 'reset_parameters'):
            module.reset_parameters()

        if self.config.rescale_prenorm_residual:
            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
            #
            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
            for name, p in module.named_parameters():
                if name in ["out_proj.weight"]:
                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
                    # We need to reinit p since this code could be called multiple times
                    # Having just p *= scale would repeatedly scale it down
                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
                    with torch.no_grad():
                        p /= math.sqrt(self.config.num_layers)


@dataclass
class SambaOutput(ModelOutput):
    """
    Class for the Samba model outputs.

    Args:
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        cache_params (`MambaCache`):
            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
            avoid providing the old `input_ids`.

            Includes both the State space model state matrices after the selective scan, and the Convolutional states
        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
    """

    last_hidden_state: torch.FloatTensor | None = None
    cache_params: MambaCache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None


@dataclass
class SambaCausalLMOutput(ModelOutput):
    """
    Base class for causal language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        cache_params (`MambaCache`):
            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
            avoid providing the old `input_ids`.

            Includes both the State space model state matrices after the selective scan, and the Convolutional states
        hidden_states (`tuple(torch.FloatTensor)`, *optional*,
            returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    cache_params: MambaCache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None


class SambaModel(SambaPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([SambaBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])

        self.gradient_checkpointing = False
        self.norm_f = RMSNorm(config.hidden_size, eps=config.norm_eps)
        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings

    def set_input_embeddings(self, new_embeddings):
        self.embeddings = new_embeddings

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.LongTensor | None = None,
        cache_params: MambaCache | None = None,
        use_cache: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs: Unpack[dict],
    ) -> tuple | SambaOutput:
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):  # ^ is python for xor
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one",
            )

        if inputs_embeds is None:
            inputs_embeds = self.embeddings(input_ids)

        if cache_params is None and use_cache:
            cache_params = MambaCache(
                self.config, inputs_embeds.size(0), device=inputs_embeds.device, dtype=inputs_embeds.dtype,
            )

        hidden_states = inputs_embeds
        all_hidden_states = () if output_hidden_states else None
        for mixer_block in self.layers:
            hidden_states = mixer_block(
                hidden_states,
                cache_params=cache_params,
                **kwargs,
            )

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

        if use_cache:
            cache_params.seqlen_offset += inputs_embeds.shape[1]

        hidden_states = self.norm_f(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)

        return SambaOutput(
            last_hidden_state=hidden_states,
            cache_params=cache_params if use_cache else None,
            hidden_states=all_hidden_states,
        )


class SambaForCausalLM(SambaPreTrainedModel, FLAGenerationMixin):

    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.backbone = SambaModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.criterion = None

        # Initialize weights and apply final processing
        self.post_init()

    def get_output_embeddings(self):
        return self.lm_head

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

    def get_input_embeddings(self):
        return self.backbone.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.backbone.set_input_embeddings(new_embeddings)

    def _update_model_kwargs_for_generation(
        self, outputs: ModelOutput, model_kwargs: dict[str, Any], **kwargs,
    ) -> dict[str, Any]:
        model_kwargs["cache_params"] = outputs.get("cache_params", None)
        return model_kwargs

    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: Optional[torch.Tensor] = None,  # noqa
        inputs_embeds: torch.FloatTensor | None = None,
        cache_params: MambaCache | None = None,
        labels: torch.LongTensor | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | None = 0,
        **kwargs: Unpack[dict],
    ) -> tuple | SambaCausalLMOutput:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.backbone(
            input_ids,
            cache_params=cache_params,
            inputs_embeds=inputs_embeds,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = outputs[0]

        loss, logits = None, None
        if not self.config.fuse_linear_cross_entropy or labels is None:
            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
        if labels is not None:
            if getattr(self, 'criterion', None) is None:
                if self.config.fuse_linear_cross_entropy:
                    criterion = FusedLinearCrossEntropyLoss(use_l2warp=self.config.use_l2warp)
                elif self.config.fuse_cross_entropy:
                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
                else:
                    criterion = nn.CrossEntropyLoss()
            else:
                criterion = self.criterion
            labels = labels.to(hidden_states.device)
            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
            if self.config.fuse_linear_cross_entropy:
                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
            else:
                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
                loss = l2_warp(loss, logits) if self.config.use_l2warp else loss

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

        return SambaCausalLMOutput(
            loss=loss,
            logits=logits,
            cache_params=outputs.cache_params,
            hidden_states=outputs.hidden_states,
        )