File size: 7,052 Bytes
1bb5aa8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9d1613c
1bb5aa8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import List, Optional, Tuple, Union
import torch
from transformers import LlamaModel, LlamaPreTrainedModel, LlamaForCausalLM, AutoModel
from transformers.models.llama.modeling_llama import LlamaDecoderLayer, LlamaRMSNorm, LlamaRotaryEmbedding, LlamaConfig, LlamaMLP, LlamaAttention, LlamaFlashAttention2, LlamaSdpaAttention
from transformers.utils import logging
from torch import nn
from torch.nn import CrossEntropyLoss
import torch.nn.functional as F
from transformers.modeling_outputs import BaseModelOutputWithPast, MaskedLMOutput, CausalLMOutputWithPast, TokenClassifierOutput
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.cache_utils import Cache, DynamicCache, StaticCache

logger = logging.get_logger(__name__)

class ModifiedLlamaAttention(LlamaAttention):

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.is_causal = False


class ModifiedLlamaFlashAttention2(LlamaFlashAttention2):

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.is_causal = False


class ModifiedLlamaSdpaAttention(LlamaSdpaAttention):
    
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.is_causal = False


LLAMA_ATTENTION_CLASSES = {
    "eager": ModifiedLlamaAttention,
    "flash_attention_2": ModifiedLlamaFlashAttention2,
    "sdpa": ModifiedLlamaSdpaAttention,
}


class ModifiedLlamaDecoderLayer(LlamaDecoderLayer):
    def __init__(self, config: LlamaConfig, layer_idx: int):
        nn.Module.__init__(self)
        self.hidden_size = config.hidden_size

        self.self_attn = LLAMA_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)

        self.mlp = LlamaMLP(config)
        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)


class LlamaEncoderModel(LlamaModel):
    def __init__(self, config):
        LlamaPreTrainedModel.__init__(self, config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [ModifiedLlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = LlamaRotaryEmbedding(config=config)

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


    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114

        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        # if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
        #     if AttentionMaskConverter._ignore_causal_mask_sdpa(
        #         attention_mask,
        #         inputs_embeds=input_tensor,
        #         past_key_values_length=past_seen_tokens,
        #         is_training=self.training,
        #     ):
        #         return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_length()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        if attention_mask is not None and attention_mask.dim() == 4:
            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
            if attention_mask.max() != 0:
                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
            causal_mask = attention_mask
        else:
            causal_mask = torch.zeros(
                (sequence_length, target_length), dtype=dtype, device=device
            )
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )
        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask