Update custom_bitnet.py
Browse files- custom_bitnet.py +440 -46
custom_bitnet.py
CHANGED
@@ -1,76 +1,470 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
2 |
import torch
|
3 |
-
|
|
|
|
|
|
|
|
|
4 |
|
|
|
5 |
class BitNetConfig(PretrainedConfig):
|
6 |
model_type = "bitnet"
|
|
|
|
|
7 |
def __init__(
|
8 |
self,
|
9 |
-
vocab_size=
|
10 |
-
hidden_size=
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
|
|
16 |
initializer_range=0.02,
|
17 |
-
|
18 |
-
|
19 |
-
pad_token_id=
|
20 |
-
bos_token_id=
|
21 |
-
eos_token_id=
|
22 |
-
|
|
|
|
|
|
|
|
|
23 |
):
|
24 |
self.vocab_size = vocab_size
|
|
|
25 |
self.hidden_size = hidden_size
|
|
|
26 |
self.num_hidden_layers = num_hidden_layers
|
27 |
self.num_attention_heads = num_attention_heads
|
28 |
-
self.
|
29 |
self.hidden_act = hidden_act
|
30 |
-
self.max_position_embeddings = max_position_embeddings
|
31 |
self.initializer_range = initializer_range
|
32 |
-
self.
|
33 |
-
self.
|
|
|
|
|
|
|
34 |
super().__init__(
|
35 |
pad_token_id=pad_token_id,
|
36 |
bos_token_id=bos_token_id,
|
37 |
eos_token_id=eos_token_id,
|
38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
|
|
|
41 |
class BitNetForCausalLM(PreTrainedModel):
|
42 |
config_class = BitNetConfig
|
|
|
|
|
43 |
def __init__(self, config):
|
44 |
super().__init__(config)
|
45 |
-
self.
|
46 |
-
self.
|
47 |
-
nn.TransformerEncoderLayer(
|
48 |
-
d_model=config.hidden_size,
|
49 |
-
nhead=config.num_attention_heads,
|
50 |
-
dim_feedforward=config.intermediate_size,
|
51 |
-
dropout=config.dropout
|
52 |
-
) for _ in range(config.num_hidden_layers)
|
53 |
-
])
|
54 |
-
self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
55 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
56 |
-
self.
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
def
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
logits = self.lm_head(hidden_states)
|
70 |
loss = None
|
71 |
if labels is not None:
|
72 |
loss_fct = nn.CrossEntropyLoss()
|
73 |
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
|
74 |
-
return
|
75 |
-
|
76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2025 Microsoft, EleutherAI, and the HuggingFace Inc. team. All rights reserved.
|
3 |
+
# Licensed under the Apache License, Version 2.0.
|
4 |
+
|
5 |
+
from typing import Optional, Tuple, Union
|
6 |
import torch
|
7 |
+
from torch import nn
|
8 |
+
from transformers import PreTrainedModel, PretrainedConfig
|
9 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
|
10 |
+
from transformers.cache_utils import DynamicCache
|
11 |
+
from transformers.activations import ACT2FN
|
12 |
|
13 |
+
# BitNetConfig
|
14 |
class BitNetConfig(PretrainedConfig):
|
15 |
model_type = "bitnet"
|
16 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
17 |
+
|
18 |
def __init__(
|
19 |
self,
|
20 |
+
vocab_size=128256,
|
21 |
+
hidden_size=2560,
|
22 |
+
intermediate_size=6912,
|
23 |
+
num_hidden_layers=30,
|
24 |
+
num_attention_heads=20,
|
25 |
+
num_key_value_heads=5,
|
26 |
+
hidden_act="relu2",
|
27 |
+
max_position_embeddings=2048,
|
28 |
initializer_range=0.02,
|
29 |
+
rms_norm_eps=1e-5,
|
30 |
+
use_cache=True,
|
31 |
+
pad_token_id=None,
|
32 |
+
bos_token_id=128000,
|
33 |
+
eos_token_id=128001,
|
34 |
+
tie_word_embeddings=False,
|
35 |
+
rope_theta=500000.0,
|
36 |
+
attention_bias=False,
|
37 |
+
attention_dropout=0.0,
|
38 |
+
**kwargs,
|
39 |
):
|
40 |
self.vocab_size = vocab_size
|
41 |
+
self.max_position_embeddings = max_position_embeddings
|
42 |
self.hidden_size = hidden_size
|
43 |
+
self.intermediate_size = intermediate_size
|
44 |
self.num_hidden_layers = num_hidden_layers
|
45 |
self.num_attention_heads = num_attention_heads
|
46 |
+
self.num_key_value_heads = num_key_value_heads or num_attention_heads
|
47 |
self.hidden_act = hidden_act
|
|
|
48 |
self.initializer_range = initializer_range
|
49 |
+
self.rms_norm_eps = rms_norm_eps
|
50 |
+
self.use_cache = use_cache
|
51 |
+
self.rope_theta = rope_theta
|
52 |
+
self.attention_bias = attention_bias
|
53 |
+
self.attention_dropout = attention_dropout
|
54 |
super().__init__(
|
55 |
pad_token_id=pad_token_id,
|
56 |
bos_token_id=bos_token_id,
|
57 |
eos_token_id=eos_token_id,
|
58 |
+
tie_word_embeddings=tie_word_embeddings,
|
59 |
+
**kwargs,
|
60 |
+
)
|
61 |
+
|
62 |
+
# BitNetRMSNorm
|
63 |
+
class BitNetRMSNorm(nn.Module):
|
64 |
+
def __init__(self, hidden_size, eps=1e-6):
|
65 |
+
super().__init__()
|
66 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
67 |
+
self.variance_epsilon = eps
|
68 |
+
|
69 |
+
def forward(self, hidden_states):
|
70 |
+
input_dtype = hidden_states.dtype
|
71 |
+
hidden_states = hidden_states.to(torch.float32)
|
72 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
73 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
74 |
+
return self.weight * hidden_states.to(input_dtype)
|
75 |
+
|
76 |
+
# BitNetMLP
|
77 |
+
class BitNetMLP(nn.Module):
|
78 |
+
def __init__(self, config: BitNetConfig):
|
79 |
+
super().__init__()
|
80 |
+
self.config = config
|
81 |
+
self.hidden_size = config.hidden_size
|
82 |
+
self.intermediate_size = config.intermediate_size
|
83 |
+
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
84 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
85 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
86 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
87 |
+
self.ffn_sub_norm = BitNetRMSNorm(config.intermediate_size, eps=config.rms_norm_eps)
|
88 |
+
|
89 |
+
def forward(self, x):
|
90 |
+
down_proj = self.down_proj(self.ffn_sub_norm(self.act_fn(self.gate_proj(x)) * self.up_proj(x)))
|
91 |
+
return down_proj
|
92 |
+
|
93 |
+
# Utility Functions
|
94 |
+
def rotate_half(x):
|
95 |
+
x1 = x[..., : x.shape[-1] // 2]
|
96 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
97 |
+
return torch.cat((-x2, x1), dim=-1)
|
98 |
+
|
99 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
100 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
101 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
102 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
103 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
104 |
+
return q_embed, k_embed
|
105 |
+
|
106 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
107 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
108 |
+
if n_rep == 1:
|
109 |
+
return hidden_states
|
110 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
111 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
112 |
+
|
113 |
+
def eager_attention_forward(
|
114 |
+
module: nn.Module,
|
115 |
+
query: torch.Tensor,
|
116 |
+
key: torch.Tensor,
|
117 |
+
value: torch.Tensor,
|
118 |
+
attention_mask: Optional[torch.Tensor],
|
119 |
+
scaling: float,
|
120 |
+
dropout: float = 0.0,
|
121 |
+
**kwargs,
|
122 |
+
):
|
123 |
+
key_states = repeat_kv(key, module.num_key_value_groups)
|
124 |
+
value_states = repeat_kv(value, module.num_key_value_groups)
|
125 |
+
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
126 |
+
if attention_mask is not None:
|
127 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
128 |
+
attn_weights = attn_weights + causal_mask
|
129 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
130 |
+
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
131 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
132 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
133 |
+
return attn_output, attn_weights
|
134 |
+
|
135 |
+
# BitNetAttention
|
136 |
+
class BitNetAttention(nn.Module):
|
137 |
+
def __init__(self, config: BitNetConfig, layer_idx: int):
|
138 |
+
super().__init__()
|
139 |
+
self.config = config
|
140 |
+
self.layer_idx = layer_idx
|
141 |
+
self.head_dim = config.hidden_size // config.num_attention_heads
|
142 |
+
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
|
143 |
+
self.scaling = self.head_dim**-0.5
|
144 |
+
self.attention_dropout = config.attention_dropout
|
145 |
+
self.is_causal = True
|
146 |
+
self.q_proj = nn.Linear(
|
147 |
+
config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
|
148 |
+
)
|
149 |
+
self.k_proj = nn.Linear(
|
150 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
151 |
+
)
|
152 |
+
self.v_proj = nn.Linear(
|
153 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
154 |
+
)
|
155 |
+
self.o_proj = nn.Linear(
|
156 |
+
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
157 |
+
)
|
158 |
+
self.attn_sub_norm = BitNetRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
159 |
+
|
160 |
+
def forward(
|
161 |
+
self,
|
162 |
+
hidden_states: torch.Tensor,
|
163 |
+
position_embeddings: Tuple[torch.Tensor, torch.Tensor],
|
164 |
+
attention_mask: Optional[torch.Tensor],
|
165 |
+
past_key_value: Optional[DynamicCache] = None,
|
166 |
+
cache_position: Optional[torch.LongTensor] = None,
|
167 |
+
**kwargs,
|
168 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
169 |
+
input_shape = hidden_states.shape[:-1]
|
170 |
+
hidden_shape = (*input_shape, -1, self.head_dim)
|
171 |
+
query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
172 |
+
key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
173 |
+
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
174 |
+
cos, sin = position_embeddings
|
175 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
176 |
+
if past_key_value is not None:
|
177 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
178 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
179 |
+
attn_output, attn_weights = eager_attention_forward(
|
180 |
+
self,
|
181 |
+
query_states,
|
182 |
+
key_states,
|
183 |
+
value_states,
|
184 |
+
attention_mask,
|
185 |
+
dropout=0.0 if not self.training else self.attention_dropout,
|
186 |
+
scaling=self.scaling,
|
187 |
+
)
|
188 |
+
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
189 |
+
attn_output = self.attn_sub_norm(attn_output)
|
190 |
+
attn_output = self.o_proj(attn_output)
|
191 |
+
return attn_output, attn_weights
|
192 |
+
|
193 |
+
# BitNetDecoderLayer
|
194 |
+
class BitNetDecoderLayer(nn.Module):
|
195 |
+
def __init__(self, config: BitNetConfig, layer_idx: int):
|
196 |
+
super().__init__()
|
197 |
+
self.hidden_size = config.hidden_size
|
198 |
+
self.self_attn = BitNetAttention(config=config, layer_idx=layer_idx)
|
199 |
+
self.mlp = BitNetMLP(config)
|
200 |
+
self.input_layernorm = BitNetRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
201 |
+
self.post_attention_layernorm = BitNetRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
202 |
+
|
203 |
+
def forward(
|
204 |
+
self,
|
205 |
+
hidden_states: torch.Tensor,
|
206 |
+
attention_mask: Optional[torch.Tensor] = None,
|
207 |
+
position_ids: Optional[torch.LongTensor] = None,
|
208 |
+
past_key_value: Optional[DynamicCache] = None,
|
209 |
+
output_attentions: Optional[bool] = False,
|
210 |
+
use_cache: Optional[bool] = False,
|
211 |
+
cache_position: Optional[torch.LongTensor] = None,
|
212 |
+
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
213 |
+
**kwargs,
|
214 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
215 |
+
residual = hidden_states
|
216 |
+
hidden_states = self.input_layernorm(hidden_states)
|
217 |
+
hidden_states, self_attn_weights = self.self_attn(
|
218 |
+
hidden_states=hidden_states,
|
219 |
+
attention_mask=attention_mask,
|
220 |
+
position_ids=position_ids,
|
221 |
+
past_key_value=past_key_value,
|
222 |
+
output_attentions=output_attentions,
|
223 |
+
use_cache=use_cache,
|
224 |
+
cache_position=cache_position,
|
225 |
+
position_embeddings=position_embeddings,
|
226 |
+
**kwargs,
|
227 |
+
)
|
228 |
+
hidden_states = residual + hidden_states
|
229 |
+
residual = hidden_states
|
230 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
231 |
+
hidden_states = self.mlp(hidden_states)
|
232 |
+
hidden_states = residual + hidden_states
|
233 |
+
outputs = (hidden_states,)
|
234 |
+
if output_attentions:
|
235 |
+
outputs += (self_attn_weights,)
|
236 |
+
return outputs
|
237 |
+
|
238 |
+
# BitNetRotaryEmbedding
|
239 |
+
class BitNetRotaryEmbedding(nn.Module):
|
240 |
+
def __init__(self, config: BitNetConfig, device=None):
|
241 |
+
super().__init__()
|
242 |
+
self.rope_type = "default"
|
243 |
+
self.max_seq_len_cached = config.max_position_embeddings
|
244 |
+
self.config = config
|
245 |
+
dim = config.hidden_size // config.num_attention_heads
|
246 |
+
inv_freq = 1.0 / (config.rope_theta ** (torch.arange(0, dim, 2, dtype=torch.float, device=device) / dim))
|
247 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
248 |
+
self.original_inv_freq = self.inv_freq
|
249 |
+
|
250 |
+
def forward(self, x, position_ids):
|
251 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
252 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
253 |
+
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
254 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
255 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
256 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
257 |
+
cos = emb.cos()
|
258 |
+
sin = emb.sin()
|
259 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
260 |
+
|
261 |
+
# BitNetModel
|
262 |
+
class BitNetModel(PreTrainedModel):
|
263 |
+
config_class = BitNetConfig
|
264 |
+
supports_gradient_checkpointing = True
|
265 |
+
_no_split_modules = ["BitNetDecoderLayer"]
|
266 |
+
|
267 |
+
def __init__(self, config: BitNetConfig):
|
268 |
+
super().__init__(config)
|
269 |
+
self.padding_idx = config.pad_token_id
|
270 |
+
self.vocab_size = config.vocab_size
|
271 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
272 |
+
self.layers = nn.ModuleList(
|
273 |
+
[BitNetDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
274 |
+
)
|
275 |
+
self.norm = BitNetRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
276 |
+
self.rotary_emb = BitNetRotaryEmbedding(config=config)
|
277 |
+
self.gradient_checkpointing = False
|
278 |
+
self.post_init()
|
279 |
+
|
280 |
+
def get_input_embeddings(self):
|
281 |
+
return self.embed_tokens
|
282 |
+
|
283 |
+
def set_input_embeddings(self, value):
|
284 |
+
self.embed_tokens = value
|
285 |
+
|
286 |
+
def forward(
|
287 |
+
self,
|
288 |
+
input_ids: Optional[torch.LongTensor] = None,
|
289 |
+
attention_mask: Optional[torch.Tensor] = None,
|
290 |
+
position_ids: Optional[torch.LongTensor] = None,
|
291 |
+
past_key_values: Optional[DynamicCache] = None,
|
292 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
293 |
+
use_cache: Optional[bool] = None,
|
294 |
+
output_attentions: Optional[bool] = None,
|
295 |
+
output_hidden_states: Optional[bool] = None,
|
296 |
+
cache_position: Optional[torch.LongTensor] = None,
|
297 |
+
**kwargs,
|
298 |
+
) -> BaseModelOutputWithPast:
|
299 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
300 |
+
output_hidden_states = (
|
301 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
302 |
+
)
|
303 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
304 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
305 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
306 |
+
if inputs_embeds is None:
|
307 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
308 |
+
if use_cache and past_key_values is None:
|
309 |
+
past_key_values = DynamicCache()
|
310 |
+
if cache_position is None:
|
311 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
312 |
+
cache_position = torch.arange(
|
313 |
+
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
314 |
+
)
|
315 |
+
if position_ids is None:
|
316 |
+
position_ids = cache_position.unsqueeze(0)
|
317 |
+
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
|
318 |
+
hidden_states k= inputs_embeds
|
319 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
320 |
+
all_hidden_states = () if output_hidden_states else None
|
321 |
+
all_self_attns = () if output_attentions else None
|
322 |
+
for decoder_layer in self.layers:
|
323 |
+
if output_hidden_states:
|
324 |
+
all_hidden_states += (hidden_states,)
|
325 |
+
layer_outputs = decoder_layer(
|
326 |
+
hidden_states,
|
327 |
+
attention_mask=causal_mask,
|
328 |
+
position_ids=position_ids,
|
329 |
+
past_key_value=past_key_values,
|
330 |
+
output_attentions=output_attentions,
|
331 |
+
use_cache=use_cache,
|
332 |
+
cache_position=cache_position,
|
333 |
+
position_embeddings=position_embeddings,
|
334 |
+
)
|
335 |
+
hidden_states = layer_outputs[0]
|
336 |
+
if output_attentions:
|
337 |
+
all_self_attns += (layer_outputs[1],)
|
338 |
+
hidden_states = self.norm(hidden_states)
|
339 |
+
if output_hidden_states:
|
340 |
+
all_hidden_states += (hidden_states,)
|
341 |
+
return BaseModelOutputWithPast(
|
342 |
+
last_hidden_state=hidden_states,
|
343 |
+
past_key_values=past_key_values if use_cache else None,
|
344 |
+
hidden_states=all_hidden_states,
|
345 |
+
attentions=all_self_attns,
|
346 |
+
)
|
347 |
+
|
348 |
+
def _update_causal_mask(
|
349 |
+
self,
|
350 |
+
attention_mask: Optional[torch.Tensor],
|
351 |
+
input_tensor: torch.Tensor,
|
352 |
+
cache_position: torch.Tensor,
|
353 |
+
past_key_values: Optional[DynamicCache],
|
354 |
+
):
|
355 |
+
dtype, device = input_tensor.dtype, input_tensor.device
|
356 |
+
sequence_length = input_tensor.shape[1]
|
357 |
+
target_length = past_key_values.get_seq_length() + sequence_length + 1 if past_key_values else sequence_length + 1
|
358 |
+
min_dtype = torch.finfo(dtype).min
|
359 |
+
causal_mask = torch.full(
|
360 |
+
(sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
|
361 |
)
|
362 |
+
if sequence_length != 1:
|
363 |
+
causal_mask = torch.triu(causal_mask, diagonal=1)
|
364 |
+
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
365 |
+
causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
|
366 |
+
if attention_mask is not None:
|
367 |
+
causal_mask = causal_mask.clone()
|
368 |
+
mask_length = attention_mask.shape[-1]
|
369 |
+
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(causal_mask.device)
|
370 |
+
padding_mask = padding_mask == 0
|
371 |
+
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(padding_mask, min_dtype)
|
372 |
+
return causal_mask
|
373 |
|
374 |
+
# BitNetForCausalLM
|
375 |
class BitNetForCausalLM(PreTrainedModel):
|
376 |
config_class = BitNetConfig
|
377 |
+
_tied_weights_keys = ["lm_head.weight"]
|
378 |
+
|
379 |
def __init__(self, config):
|
380 |
super().__init__(config)
|
381 |
+
self.model = BitNetModel(config)
|
382 |
+
self.vocab_size = config.vocab_size
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
383 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
384 |
+
self.post_init()
|
385 |
+
|
386 |
+
def get_input_embeddings(self):
|
387 |
+
return self.model.embed_tokens
|
388 |
+
|
389 |
+
def set_input_embeddings(self, value):
|
390 |
+
self.model.embed_tokens = value
|
391 |
+
|
392 |
+
def get_output_embeddings(self):
|
393 |
+
return self.lm_head
|
394 |
+
|
395 |
+
def set_output_embeddings(self, new_embeddings):
|
396 |
+
self.lm_head = new_embeddings
|
397 |
+
|
398 |
+
def set_decoder(self, decoder):
|
399 |
+
self.model = decoder
|
400 |
+
|
401 |
+
def get_decoder(self):
|
402 |
+
return self.model
|
403 |
+
|
404 |
+
def forward(
|
405 |
+
self,
|
406 |
+
input_ids: Optional[torch.LongTensor] = None,
|
407 |
+
attention_mask: Optional[torch.Tensor] = None,
|
408 |
+
position_ids: Optional[torch.LongTensor] = None,
|
409 |
+
past_key_values: Optional[DynamicCache] = None,
|
410 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
411 |
+
labels: Optional[torch.LongTensor] = None,
|
412 |
+
use_cache: Optional[bool] = None,
|
413 |
+
output_attentions: Optional[bool] = None,
|
414 |
+
output_hidden_states: Optional[bool] = None,
|
415 |
+
cache_position: Optional[torch.LongTensor] = None,
|
416 |
+
**kwargs,
|
417 |
+
) -> CausalLMOutputWithPast:
|
418 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
419 |
+
output_hidden_states = (
|
420 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
421 |
+
)
|
422 |
+
outputs = self.model(
|
423 |
+
input_ids=input_ids,
|
424 |
+
attention_mask=attention_mask,
|
425 |
+
position_ids=position_ids,
|
426 |
+
past_key_values=past_key_values,
|
427 |
+
inputs_embeds=inputs_embeds,
|
428 |
+
use_cache=use_cache,
|
429 |
+
output_attentions=output_attentions,
|
430 |
+
output_hidden_states=output_hidden_states,
|
431 |
+
cache_position=cache_position,
|
432 |
+
**kwargs,
|
433 |
+
)
|
434 |
+
hidden_states = outputs.last_hidden_state
|
435 |
logits = self.lm_head(hidden_states)
|
436 |
loss = None
|
437 |
if labels is not None:
|
438 |
loss_fct = nn.CrossEntropyLoss()
|
439 |
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
|
440 |
+
return CausalLMOutputWithPast(
|
441 |
+
loss=loss,
|
442 |
+
logits=logits,
|
443 |
+
past_key_values=outputs.past_key_values,
|
444 |
+
hidden_states=outputs.hidden_states,
|
445 |
+
attentions=outputs.attentions,
|
446 |
+
)
|
447 |
+
|
448 |
+
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, **kwargs):
|
449 |
+
if past_key_values is None:
|
450 |
+
past_key_values = DynamicCache()
|
451 |
+
cache_position = kwargs.get("cache_position", None)
|
452 |
+
if cache_position is None:
|
453 |
+
past_length = past_key_values.get_seq_length() if past_key_values is not None else 0
|
454 |
+
cache_position = torch.arange(past_length, past_length + input_ids.shape[-1], device=input_ids.device)
|
455 |
+
position_ids = cache_position.unsqueeze(0)
|
456 |
+
if attention_mask is not None and attention_mask.shape[1] != input_ids.shape[1]:
|
457 |
+
attention_mask = self._update_causal_mask(
|
458 |
+
attention_mask,
|
459 |
+
input_ids,
|
460 |
+
cache_position,
|
461 |
+
past_key_values
|
462 |
+
)
|
463 |
+
return {
|
464 |
+
"input_ids": input_ids,
|
465 |
+
"position_ids": position_ids,
|
466 |
+
"attention_mask": attention_mask,
|
467 |
+
"past_key_values": past_key_values,
|
468 |
+
"cache_position": cache_position,
|
469 |
+
"use_cache": kwargs.get("use_cache", True),
|
470 |
+
}
|