Ubuntu
commited on
Commit
·
6b3e945
1
Parent(s):
6a5e92d
new
Browse files- config.json +1 -7
- modeling_glm.py +0 -1304
config.json
CHANGED
|
@@ -2,13 +2,7 @@
|
|
| 2 |
"architectures": [
|
| 3 |
"GlmForCausalLM"
|
| 4 |
],
|
| 5 |
-
"
|
| 6 |
-
"AutoModel": "modeling_glm.GlmModel",
|
| 7 |
-
"AutoModelForCausalLM": "modeling_glm.GlmForCausalLM",
|
| 8 |
-
"AutoModelForSeq2SeqLM": "modeling_glm.GlmForTokenClassification",
|
| 9 |
-
"AutoModelForSequenceClassification": "modeling_glm.GlmForSequenceClassification"
|
| 10 |
-
},
|
| 11 |
-
"rotary_percent": 1.0,
|
| 12 |
"attention_bias": false,
|
| 13 |
"attention_dropout": 0.0,
|
| 14 |
"eos_token_id": [
|
|
|
|
| 2 |
"architectures": [
|
| 3 |
"GlmForCausalLM"
|
| 4 |
],
|
| 5 |
+
"partial_rotary_factor": 1.0,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
"attention_bias": false,
|
| 7 |
"attention_dropout": 0.0,
|
| 8 |
"eos_token_id": [
|
modeling_glm.py
DELETED
|
@@ -1,1304 +0,0 @@
|
|
| 1 |
-
import math
|
| 2 |
-
from typing import List, Optional, Tuple, Union
|
| 3 |
-
|
| 4 |
-
import torch
|
| 5 |
-
import torch.nn as nn
|
| 6 |
-
|
| 7 |
-
from transformers.activations import ACT2FN
|
| 8 |
-
from transformers.cache_utils import Cache, DynamicCache, StaticCache
|
| 9 |
-
from transformers.generation import GenerationMixin
|
| 10 |
-
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
|
| 11 |
-
from transformers.modeling_flash_attention_utils import (
|
| 12 |
-
FlashAttentionKwargs,
|
| 13 |
-
_flash_attention_forward,
|
| 14 |
-
)
|
| 15 |
-
from transformers.modeling_outputs import (
|
| 16 |
-
BaseModelOutputWithPast,
|
| 17 |
-
CausalLMOutputWithPast,
|
| 18 |
-
SequenceClassifierOutputWithPast,
|
| 19 |
-
TokenClassifierOutput,
|
| 20 |
-
)
|
| 21 |
-
from transformers.modeling_utils import PreTrainedModel
|
| 22 |
-
from transformers.processing_utils import Unpack
|
| 23 |
-
from transformers.utils import (
|
| 24 |
-
add_code_sample_docstrings,
|
| 25 |
-
add_start_docstrings,
|
| 26 |
-
add_start_docstrings_to_model_forward,
|
| 27 |
-
is_flash_attn_greater_or_equal_2_10,
|
| 28 |
-
logging,
|
| 29 |
-
replace_return_docstrings,
|
| 30 |
-
)
|
| 31 |
-
from transformers.models.glm.configuration_glm import GlmConfig
|
| 32 |
-
|
| 33 |
-
logger = logging.get_logger(__name__)
|
| 34 |
-
|
| 35 |
-
_CHECKPOINT_FOR_DOC = "THUDM/glm-edge-4b-chat"
|
| 36 |
-
_CONFIG_FOR_DOC = "GlmConfig"
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
class GlmRMSNorm(nn.Module):
|
| 40 |
-
def __init__(self, hidden_size, eps=1e-6):
|
| 41 |
-
"""
|
| 42 |
-
GlmRMSNorm is equivalent to T5LayerNorm
|
| 43 |
-
"""
|
| 44 |
-
super().__init__()
|
| 45 |
-
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 46 |
-
self.variance_epsilon = eps
|
| 47 |
-
|
| 48 |
-
def forward(self, hidden_states):
|
| 49 |
-
input_dtype = hidden_states.dtype
|
| 50 |
-
hidden_states = hidden_states.to(torch.float32)
|
| 51 |
-
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 52 |
-
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
| 53 |
-
return self.weight * hidden_states.to(input_dtype)
|
| 54 |
-
|
| 55 |
-
def extra_repr(self):
|
| 56 |
-
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
class GlmRotaryEmbedding(nn.Module):
|
| 60 |
-
def __init__(self, dim, max_position_embeddings=2048, base=10000, rotary_percent=0.5, device=None):
|
| 61 |
-
super().__init__()
|
| 62 |
-
self.rotary_percent = rotary_percent
|
| 63 |
-
self.dim = dim * rotary_percent
|
| 64 |
-
self.max_position_embeddings = max_position_embeddings
|
| 65 |
-
self.base = base
|
| 66 |
-
|
| 67 |
-
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
|
| 68 |
-
self.register_buffer("inv_freq", inv_freq)
|
| 69 |
-
|
| 70 |
-
def forward(self, x, position_ids=None):
|
| 71 |
-
batch_size, seq_len, head_dim = x.shape
|
| 72 |
-
device = x.device
|
| 73 |
-
dtype = x.dtype
|
| 74 |
-
|
| 75 |
-
seq_idx = torch.arange(0, self.max_position_embeddings, device=device).float()
|
| 76 |
-
idx_theta = torch.outer(seq_idx, self.inv_freq)
|
| 77 |
-
|
| 78 |
-
if position_ids is not None:
|
| 79 |
-
idx_theta = idx_theta[position_ids[0]]
|
| 80 |
-
else:
|
| 81 |
-
idx_theta = idx_theta[:seq_len]
|
| 82 |
-
if self.rotary_percent == 0.5:
|
| 83 |
-
idx_theta = torch.cat([idx_theta, idx_theta], dim=-1) # for glm-4-9b
|
| 84 |
-
|
| 85 |
-
device_type = device.type
|
| 86 |
-
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
| 87 |
-
with torch.autocast(device_type=device_type, enabled=False):
|
| 88 |
-
cos = torch.cos(idx_theta).to(dtype=dtype)
|
| 89 |
-
sin = torch.sin(idx_theta).to(dtype=dtype)
|
| 90 |
-
|
| 91 |
-
cos = cos[None, :, :].expand(batch_size, seq_len, -1)
|
| 92 |
-
sin = sin[None, :, :].expand(batch_size, seq_len, -1)
|
| 93 |
-
|
| 94 |
-
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
class GlmMLP(nn.Module):
|
| 98 |
-
def __init__(self, config):
|
| 99 |
-
super().__init__()
|
| 100 |
-
|
| 101 |
-
self.config = config
|
| 102 |
-
self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
|
| 103 |
-
self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
|
| 104 |
-
|
| 105 |
-
self.activation_fn = ACT2FN[config.hidden_act]
|
| 106 |
-
|
| 107 |
-
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
|
| 108 |
-
up_states = self.gate_up_proj(hidden_states)
|
| 109 |
-
|
| 110 |
-
gate, up_states = up_states.chunk(2, dim=-1)
|
| 111 |
-
up_states = up_states * self.activation_fn(gate)
|
| 112 |
-
|
| 113 |
-
return self.down_proj(up_states)
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 117 |
-
"""
|
| 118 |
-
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
| 119 |
-
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
| 120 |
-
"""
|
| 121 |
-
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
| 122 |
-
if n_rep == 1:
|
| 123 |
-
return hidden_states
|
| 124 |
-
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
| 125 |
-
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
def rotate_half(x):
|
| 129 |
-
"""Rotates half the hidden dims of the input."""
|
| 130 |
-
x1 = x[..., 0::2]
|
| 131 |
-
x2 = x[..., 1::2]
|
| 132 |
-
return torch.stack((-x2, x1), dim=-1).flatten(-2)
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1, rotary_percent=0.5):
|
| 136 |
-
"""
|
| 137 |
-
Applies Rotary Position Embedding to the query and key tensors.
|
| 138 |
-
rotary_percent is for glm-4-9b(0.5) or glm-edge(1.0)
|
| 139 |
-
"""
|
| 140 |
-
cos = cos.unsqueeze(unsqueeze_dim)
|
| 141 |
-
sin = sin.unsqueeze(unsqueeze_dim)
|
| 142 |
-
|
| 143 |
-
# Interleave them instead of usual shape
|
| 144 |
-
cos = cos[..., : int(cos.shape[-1] * rotary_percent)].repeat_interleave(2, dim=-1)
|
| 145 |
-
sin = sin[..., : int(sin.shape[-1] * rotary_percent)].repeat_interleave(2, dim=-1)
|
| 146 |
-
|
| 147 |
-
# Keep rotary_percent(half or not) for later concatenation
|
| 148 |
-
rotary_dim = int(q.shape[-1] * rotary_percent)
|
| 149 |
-
q, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
|
| 150 |
-
k, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
|
| 151 |
-
|
| 152 |
-
# Apply rotary embeddings on the first half or full tensor
|
| 153 |
-
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 154 |
-
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 155 |
-
|
| 156 |
-
# Concatenate back to full shape
|
| 157 |
-
q_embed = torch.cat([q_embed, q_pass], dim=-1)
|
| 158 |
-
k_embed = torch.cat([k_embed, k_pass], dim=-1)
|
| 159 |
-
return q_embed, k_embed
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
class GlmAttention(nn.Module):
|
| 163 |
-
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 164 |
-
|
| 165 |
-
def __init__(self, config: GlmConfig, layer_idx: Optional[int] = None):
|
| 166 |
-
super().__init__()
|
| 167 |
-
self.config = config
|
| 168 |
-
self.layer_idx = layer_idx
|
| 169 |
-
if layer_idx is None:
|
| 170 |
-
logger.warning_once(
|
| 171 |
-
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
| 172 |
-
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
| 173 |
-
"when creating this class."
|
| 174 |
-
)
|
| 175 |
-
|
| 176 |
-
self.attention_dropout = config.attention_dropout
|
| 177 |
-
self.hidden_size = config.hidden_size
|
| 178 |
-
self.num_heads = config.num_attention_heads
|
| 179 |
-
self.head_dim = self.hidden_size // self.num_heads
|
| 180 |
-
self.num_key_value_heads = config.num_key_value_heads
|
| 181 |
-
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
| 182 |
-
self.is_causal = True
|
| 183 |
-
self.scaling = 1 / math.sqrt(self.head_dim)
|
| 184 |
-
self.rotary_percent = config.rotary_percent if hasattr(config, "rotary_percent") else 0.5
|
| 185 |
-
|
| 186 |
-
if (self.head_dim * self.num_heads) != self.hidden_size:
|
| 187 |
-
raise ValueError(
|
| 188 |
-
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
| 189 |
-
f" and `num_heads`: {self.num_heads})."
|
| 190 |
-
)
|
| 191 |
-
|
| 192 |
-
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
|
| 193 |
-
self.k_proj = nn.Linear(
|
| 194 |
-
self.hidden_size,
|
| 195 |
-
self.num_key_value_heads * self.head_dim,
|
| 196 |
-
bias=config.attention_bias,
|
| 197 |
-
)
|
| 198 |
-
self.v_proj = nn.Linear(
|
| 199 |
-
self.hidden_size,
|
| 200 |
-
self.num_key_value_heads * self.head_dim,
|
| 201 |
-
bias=config.attention_bias,
|
| 202 |
-
)
|
| 203 |
-
self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
|
| 204 |
-
|
| 205 |
-
def forward(
|
| 206 |
-
self,
|
| 207 |
-
hidden_states: torch.Tensor,
|
| 208 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 209 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 210 |
-
past_key_value: Optional[Cache] = None,
|
| 211 |
-
output_attentions: bool = False,
|
| 212 |
-
use_cache: bool = False,
|
| 213 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 214 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
| 215 |
-
**kwargs,
|
| 216 |
-
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 217 |
-
bsz, q_len, _ = hidden_states.size()
|
| 218 |
-
|
| 219 |
-
query_states = self.q_proj(hidden_states)
|
| 220 |
-
key_states = self.k_proj(hidden_states)
|
| 221 |
-
value_states = self.v_proj(hidden_states)
|
| 222 |
-
|
| 223 |
-
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 224 |
-
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 225 |
-
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 226 |
-
|
| 227 |
-
cos, sin = position_embeddings
|
| 228 |
-
|
| 229 |
-
query_states, key_states = apply_rotary_pos_emb(
|
| 230 |
-
query_states,
|
| 231 |
-
key_states,
|
| 232 |
-
cos,
|
| 233 |
-
sin,
|
| 234 |
-
rotary_percent=self.rotary_percent,
|
| 235 |
-
)
|
| 236 |
-
if past_key_value is not None:
|
| 237 |
-
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 238 |
-
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
| 239 |
-
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 240 |
-
|
| 241 |
-
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 242 |
-
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 243 |
-
|
| 244 |
-
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scaling
|
| 245 |
-
|
| 246 |
-
if attention_mask is not None: # no matter the length, we just slice it
|
| 247 |
-
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
| 248 |
-
attn_weights = attn_weights + causal_mask
|
| 249 |
-
|
| 250 |
-
# upcast attention to fp32
|
| 251 |
-
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
| 252 |
-
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
| 253 |
-
attn_output = torch.matmul(attn_weights, value_states)
|
| 254 |
-
|
| 255 |
-
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
| 256 |
-
raise ValueError(
|
| 257 |
-
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
| 258 |
-
f" {attn_output.size()}"
|
| 259 |
-
)
|
| 260 |
-
|
| 261 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 262 |
-
|
| 263 |
-
attn_output = attn_output.view(bsz, q_len, -1)
|
| 264 |
-
attn_output = self.o_proj(attn_output)
|
| 265 |
-
|
| 266 |
-
if not output_attentions:
|
| 267 |
-
attn_weights = None
|
| 268 |
-
|
| 269 |
-
return attn_output, attn_weights, past_key_value
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
class GlmFlashAttention2(GlmAttention):
|
| 273 |
-
"""
|
| 274 |
-
Glm flash attention module. This module inherits from `GlmAttention` as the weights of the module stays
|
| 275 |
-
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
| 276 |
-
flash attention and deal with padding tokens in case the input contains any of them.
|
| 277 |
-
"""
|
| 278 |
-
|
| 279 |
-
def __init__(self, *args, **kwargs):
|
| 280 |
-
super().__init__(*args, **kwargs)
|
| 281 |
-
|
| 282 |
-
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
| 283 |
-
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
| 284 |
-
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
| 285 |
-
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
| 286 |
-
|
| 287 |
-
def forward(
|
| 288 |
-
self,
|
| 289 |
-
hidden_states: torch.Tensor,
|
| 290 |
-
attention_mask: Optional[torch.LongTensor] = None,
|
| 291 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 292 |
-
past_key_value: Optional[Cache] = None,
|
| 293 |
-
output_attentions: bool = False,
|
| 294 |
-
use_cache: bool = False,
|
| 295 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 296 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
| 297 |
-
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 298 |
-
output_attentions = False
|
| 299 |
-
|
| 300 |
-
bsz, q_len, _ = hidden_states.size()
|
| 301 |
-
|
| 302 |
-
query_states = self.q_proj(hidden_states)
|
| 303 |
-
key_states = self.k_proj(hidden_states)
|
| 304 |
-
value_states = self.v_proj(hidden_states)
|
| 305 |
-
|
| 306 |
-
# Flash attention requires the input to have the shape
|
| 307 |
-
# batch_size x seq_length x head_dim x hidden_dim
|
| 308 |
-
# therefore we just need to keep the original shape
|
| 309 |
-
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 310 |
-
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 311 |
-
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 312 |
-
|
| 313 |
-
cos, sin = position_embeddings
|
| 314 |
-
query_states, key_states = apply_rotary_pos_emb(
|
| 315 |
-
query_states,
|
| 316 |
-
key_states,
|
| 317 |
-
cos,
|
| 318 |
-
sin,
|
| 319 |
-
rotary_percent=self.rotary_percent,
|
| 320 |
-
)
|
| 321 |
-
|
| 322 |
-
if past_key_value is not None:
|
| 323 |
-
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 324 |
-
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
| 325 |
-
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 326 |
-
|
| 327 |
-
# TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
|
| 328 |
-
# to be able to avoid many of these transpose/reshape/view.
|
| 329 |
-
query_states = query_states.transpose(1, 2)
|
| 330 |
-
key_states = key_states.transpose(1, 2)
|
| 331 |
-
value_states = value_states.transpose(1, 2)
|
| 332 |
-
|
| 333 |
-
dropout_rate = self.attention_dropout if self.training else 0.0
|
| 334 |
-
|
| 335 |
-
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
| 336 |
-
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
| 337 |
-
# cast them back in the correct dtype just to be sure everything works as expected.
|
| 338 |
-
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
|
| 339 |
-
# in fp32. (GlmRMSNorm handles it correctly)
|
| 340 |
-
|
| 341 |
-
input_dtype = query_states.dtype
|
| 342 |
-
if input_dtype == torch.float32:
|
| 343 |
-
if torch.is_autocast_enabled():
|
| 344 |
-
target_dtype = torch.get_autocast_gpu_dtype()
|
| 345 |
-
# Handle the case where the model is quantized
|
| 346 |
-
elif hasattr(self.config, "_pre_quantization_dtype"):
|
| 347 |
-
target_dtype = self.config._pre_quantization_dtype
|
| 348 |
-
else:
|
| 349 |
-
target_dtype = self.q_proj.weight.dtype
|
| 350 |
-
|
| 351 |
-
logger.warning_once(
|
| 352 |
-
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
| 353 |
-
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
| 354 |
-
f" {target_dtype}."
|
| 355 |
-
)
|
| 356 |
-
|
| 357 |
-
query_states = query_states.to(target_dtype)
|
| 358 |
-
key_states = key_states.to(target_dtype)
|
| 359 |
-
value_states = value_states.to(target_dtype)
|
| 360 |
-
|
| 361 |
-
attn_output = _flash_attention_forward(
|
| 362 |
-
query_states,
|
| 363 |
-
key_states,
|
| 364 |
-
value_states,
|
| 365 |
-
attention_mask,
|
| 366 |
-
q_len,
|
| 367 |
-
position_ids=position_ids,
|
| 368 |
-
dropout=dropout_rate,
|
| 369 |
-
softmax_scale=self.scaling,
|
| 370 |
-
sliding_window=getattr(self, "sliding_window", None),
|
| 371 |
-
use_top_left_mask=self._flash_attn_uses_top_left_mask,
|
| 372 |
-
is_causal=self.is_causal,
|
| 373 |
-
)
|
| 374 |
-
|
| 375 |
-
attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
|
| 376 |
-
attn_output = self.o_proj(attn_output)
|
| 377 |
-
|
| 378 |
-
if not output_attentions:
|
| 379 |
-
attn_weights = None
|
| 380 |
-
|
| 381 |
-
return attn_output, attn_weights, past_key_value
|
| 382 |
-
|
| 383 |
-
|
| 384 |
-
class GlmSdpaAttention(GlmAttention):
|
| 385 |
-
"""
|
| 386 |
-
Glm attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
| 387 |
-
`GlmAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
| 388 |
-
SDPA API.
|
| 389 |
-
"""
|
| 390 |
-
|
| 391 |
-
# Adapted from GlmAttention.forward
|
| 392 |
-
def forward(
|
| 393 |
-
self,
|
| 394 |
-
hidden_states: torch.Tensor,
|
| 395 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 396 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 397 |
-
past_key_value: Optional[Cache] = None,
|
| 398 |
-
output_attentions: bool = False,
|
| 399 |
-
use_cache: bool = False,
|
| 400 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 401 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.45
|
| 402 |
-
**kwargs,
|
| 403 |
-
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 404 |
-
if output_attentions:
|
| 405 |
-
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
| 406 |
-
logger.warning_once(
|
| 407 |
-
"GlmModel is using GlmSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
| 408 |
-
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 409 |
-
)
|
| 410 |
-
return super().forward(
|
| 411 |
-
hidden_states=hidden_states,
|
| 412 |
-
attention_mask=attention_mask,
|
| 413 |
-
position_ids=position_ids,
|
| 414 |
-
past_key_value=past_key_value,
|
| 415 |
-
output_attentions=output_attentions,
|
| 416 |
-
use_cache=use_cache,
|
| 417 |
-
cache_position=cache_position,
|
| 418 |
-
position_embeddings=position_embeddings,
|
| 419 |
-
)
|
| 420 |
-
|
| 421 |
-
bsz, q_len, _ = hidden_states.size()
|
| 422 |
-
|
| 423 |
-
query_states = self.q_proj(hidden_states)
|
| 424 |
-
key_states = self.k_proj(hidden_states)
|
| 425 |
-
value_states = self.v_proj(hidden_states)
|
| 426 |
-
|
| 427 |
-
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 428 |
-
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 429 |
-
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 430 |
-
|
| 431 |
-
cos, sin = position_embeddings
|
| 432 |
-
query_states, key_states = apply_rotary_pos_emb(
|
| 433 |
-
query_states,
|
| 434 |
-
key_states,
|
| 435 |
-
cos,
|
| 436 |
-
sin,
|
| 437 |
-
rotary_percent=self.rotary_percent,
|
| 438 |
-
)
|
| 439 |
-
|
| 440 |
-
if past_key_value is not None:
|
| 441 |
-
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 442 |
-
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
| 443 |
-
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 444 |
-
|
| 445 |
-
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 446 |
-
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 447 |
-
|
| 448 |
-
causal_mask = attention_mask
|
| 449 |
-
if attention_mask is not None:
|
| 450 |
-
causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
|
| 451 |
-
|
| 452 |
-
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 453 |
-
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
| 454 |
-
if query_states.device.type == "cuda" and causal_mask is not None:
|
| 455 |
-
query_states = query_states.contiguous()
|
| 456 |
-
key_states = key_states.contiguous()
|
| 457 |
-
value_states = value_states.contiguous()
|
| 458 |
-
|
| 459 |
-
# We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
|
| 460 |
-
# in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
|
| 461 |
-
is_causal = True if causal_mask is None and q_len > 1 else False
|
| 462 |
-
|
| 463 |
-
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
| 464 |
-
query_states,
|
| 465 |
-
key_states,
|
| 466 |
-
value_states,
|
| 467 |
-
attn_mask=causal_mask,
|
| 468 |
-
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 469 |
-
is_causal=is_causal,
|
| 470 |
-
scale=self.scaling,
|
| 471 |
-
)
|
| 472 |
-
|
| 473 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 474 |
-
attn_output = attn_output.view(bsz, q_len, -1)
|
| 475 |
-
|
| 476 |
-
attn_output = self.o_proj(attn_output)
|
| 477 |
-
|
| 478 |
-
return attn_output, None, past_key_value
|
| 479 |
-
|
| 480 |
-
|
| 481 |
-
GLM_ATTENTION_CLASSES = {
|
| 482 |
-
"eager": GlmAttention,
|
| 483 |
-
"flash_attention_2": GlmFlashAttention2,
|
| 484 |
-
"sdpa": GlmSdpaAttention,
|
| 485 |
-
}
|
| 486 |
-
|
| 487 |
-
|
| 488 |
-
class GlmDecoderLayer(nn.Module):
|
| 489 |
-
def __init__(self, config: GlmConfig, layer_idx: Optional[int] = None):
|
| 490 |
-
super().__init__()
|
| 491 |
-
self.hidden_size = config.hidden_size
|
| 492 |
-
|
| 493 |
-
self.self_attn = GLM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
|
| 494 |
-
|
| 495 |
-
self.mlp = GlmMLP(config)
|
| 496 |
-
self.input_layernorm = GlmRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 497 |
-
self.post_attention_layernorm = GlmRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 498 |
-
|
| 499 |
-
def forward(
|
| 500 |
-
self,
|
| 501 |
-
hidden_states: torch.Tensor,
|
| 502 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 503 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 504 |
-
past_key_value: Optional[Cache] = None,
|
| 505 |
-
output_attentions: Optional[bool] = False,
|
| 506 |
-
use_cache: Optional[bool] = False,
|
| 507 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 508 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # will become mandatory in v4.46
|
| 509 |
-
**kwargs,
|
| 510 |
-
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
| 511 |
-
"""
|
| 512 |
-
Args:
|
| 513 |
-
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 514 |
-
attention_mask (`torch.FloatTensor`, *optional*):
|
| 515 |
-
attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
|
| 516 |
-
query_sequence_length, key_sequence_length)` if default attention is used.
|
| 517 |
-
output_attentions (`bool`, *optional*):
|
| 518 |
-
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 519 |
-
returned tensors for more detail.
|
| 520 |
-
use_cache (`bool`, *optional*):
|
| 521 |
-
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 522 |
-
(see `past_key_values`).
|
| 523 |
-
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
| 524 |
-
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 525 |
-
Indices depicting the position of the input sequence tokens in the sequence
|
| 526 |
-
position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
|
| 527 |
-
Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
|
| 528 |
-
with `head_dim` being the embedding dimension of each attention head.
|
| 529 |
-
kwargs (`dict`, *optional*):
|
| 530 |
-
Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
|
| 531 |
-
into the model
|
| 532 |
-
"""
|
| 533 |
-
residual = hidden_states
|
| 534 |
-
|
| 535 |
-
hidden_states = self.input_layernorm(hidden_states)
|
| 536 |
-
|
| 537 |
-
# Self Attention
|
| 538 |
-
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
| 539 |
-
hidden_states=hidden_states,
|
| 540 |
-
attention_mask=attention_mask,
|
| 541 |
-
position_ids=position_ids,
|
| 542 |
-
past_key_value=past_key_value,
|
| 543 |
-
output_attentions=output_attentions,
|
| 544 |
-
use_cache=use_cache,
|
| 545 |
-
cache_position=cache_position,
|
| 546 |
-
position_embeddings=position_embeddings,
|
| 547 |
-
**kwargs,
|
| 548 |
-
)
|
| 549 |
-
hidden_states = residual + hidden_states
|
| 550 |
-
|
| 551 |
-
# Fully Connected
|
| 552 |
-
residual = hidden_states
|
| 553 |
-
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 554 |
-
hidden_states = self.mlp(hidden_states)
|
| 555 |
-
hidden_states = residual + hidden_states
|
| 556 |
-
|
| 557 |
-
outputs = (hidden_states,)
|
| 558 |
-
|
| 559 |
-
if output_attentions:
|
| 560 |
-
outputs += (self_attn_weights,)
|
| 561 |
-
|
| 562 |
-
if use_cache:
|
| 563 |
-
outputs += (present_key_value,)
|
| 564 |
-
|
| 565 |
-
return outputs
|
| 566 |
-
|
| 567 |
-
|
| 568 |
-
GLM_START_DOCSTRING = r"""
|
| 569 |
-
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
| 570 |
-
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
| 571 |
-
etc.)
|
| 572 |
-
|
| 573 |
-
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
| 574 |
-
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
| 575 |
-
and behavior.
|
| 576 |
-
|
| 577 |
-
Parameters:
|
| 578 |
-
config ([`GlmConfig`]):
|
| 579 |
-
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
| 580 |
-
load the weights associated with the model, only the configuration. Check out the
|
| 581 |
-
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
| 582 |
-
"""
|
| 583 |
-
|
| 584 |
-
|
| 585 |
-
@add_start_docstrings(
|
| 586 |
-
"The bare Glm Model outputting raw hidden-states without any specific head on top.",
|
| 587 |
-
GLM_START_DOCSTRING,
|
| 588 |
-
)
|
| 589 |
-
class GlmPreTrainedModel(PreTrainedModel):
|
| 590 |
-
config_class = GlmConfig
|
| 591 |
-
base_model_prefix = "model"
|
| 592 |
-
supports_gradient_checkpointing = True
|
| 593 |
-
_no_split_modules = ["GlmDecoderLayer"]
|
| 594 |
-
_skip_keys_device_placement = ["past_key_values"]
|
| 595 |
-
_supports_flash_attn_2 = True
|
| 596 |
-
_supports_sdpa = True
|
| 597 |
-
_supports_cache_class = True
|
| 598 |
-
_supports_quantized_cache = True
|
| 599 |
-
_supports_static_cache = True
|
| 600 |
-
|
| 601 |
-
def _init_weights(self, module):
|
| 602 |
-
std = self.config.initializer_range
|
| 603 |
-
if isinstance(module, nn.Linear):
|
| 604 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
| 605 |
-
if module.bias is not None:
|
| 606 |
-
module.bias.data.zero_()
|
| 607 |
-
elif isinstance(module, nn.Embedding):
|
| 608 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
| 609 |
-
if module.padding_idx is not None:
|
| 610 |
-
module.weight.data[module.padding_idx].zero_()
|
| 611 |
-
|
| 612 |
-
|
| 613 |
-
GLM_INPUTS_DOCSTRING = r"""
|
| 614 |
-
Args:
|
| 615 |
-
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 616 |
-
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
| 617 |
-
it.
|
| 618 |
-
|
| 619 |
-
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 620 |
-
[`PreTrainedTokenizer.__call__`] for details.
|
| 621 |
-
|
| 622 |
-
[What are input IDs?](../glossary#input-ids)
|
| 623 |
-
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 624 |
-
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 625 |
-
|
| 626 |
-
- 1 for tokens that are **not masked**,
|
| 627 |
-
- 0 for tokens that are **masked**.
|
| 628 |
-
|
| 629 |
-
[What are attention masks?](../glossary#attention-mask)
|
| 630 |
-
|
| 631 |
-
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 632 |
-
[`PreTrainedTokenizer.__call__`] for details.
|
| 633 |
-
|
| 634 |
-
If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
|
| 635 |
-
`past_key_values`).
|
| 636 |
-
|
| 637 |
-
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
| 638 |
-
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
| 639 |
-
information on the default strategy.
|
| 640 |
-
|
| 641 |
-
- 1 indicates the head is **not masked**,
|
| 642 |
-
- 0 indicates the head is **masked**.
|
| 643 |
-
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 644 |
-
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
| 645 |
-
config.n_positions - 1]`.
|
| 646 |
-
|
| 647 |
-
[What are position IDs?](../glossary#position-ids)
|
| 648 |
-
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
| 649 |
-
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
| 650 |
-
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
| 651 |
-
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
| 652 |
-
|
| 653 |
-
Two formats are allowed:
|
| 654 |
-
- a [`~cache_utils.Cache`] instance, see our
|
| 655 |
-
[kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
|
| 656 |
-
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
| 657 |
-
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
| 658 |
-
cache format.
|
| 659 |
-
|
| 660 |
-
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
| 661 |
-
legacy cache format will be returned.
|
| 662 |
-
|
| 663 |
-
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
| 664 |
-
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
| 665 |
-
of shape `(batch_size, sequence_length)`.
|
| 666 |
-
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
| 667 |
-
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
| 668 |
-
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
| 669 |
-
model's internal embedding lookup matrix.
|
| 670 |
-
use_cache (`bool`, *optional*):
|
| 671 |
-
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
| 672 |
-
`past_key_values`).
|
| 673 |
-
output_attentions (`bool`, *optional*):
|
| 674 |
-
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 675 |
-
tensors for more detail.
|
| 676 |
-
output_hidden_states (`bool`, *optional*):
|
| 677 |
-
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 678 |
-
more detail.
|
| 679 |
-
return_dict (`bool`, *optional*):
|
| 680 |
-
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 681 |
-
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 682 |
-
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
|
| 683 |
-
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
|
| 684 |
-
the complete sequence length.
|
| 685 |
-
"""
|
| 686 |
-
|
| 687 |
-
|
| 688 |
-
@add_start_docstrings(
|
| 689 |
-
"The bare Glm Model outputting raw hidden-states without any specific head on top.",
|
| 690 |
-
GLM_START_DOCSTRING,
|
| 691 |
-
)
|
| 692 |
-
class GlmModel(GlmPreTrainedModel):
|
| 693 |
-
"""
|
| 694 |
-
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GlmDecoderLayer`]
|
| 695 |
-
|
| 696 |
-
Args:
|
| 697 |
-
config: GlmConfig
|
| 698 |
-
"""
|
| 699 |
-
|
| 700 |
-
def __init__(self, config: GlmConfig):
|
| 701 |
-
super().__init__(config)
|
| 702 |
-
self.padding_idx = config.pad_token_id
|
| 703 |
-
self.vocab_size = config.vocab_size
|
| 704 |
-
self.rotary_percent = config.rotary_percent if hasattr(config, "rotary_percent") else 0.5
|
| 705 |
-
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
| 706 |
-
self.layers = nn.ModuleList(
|
| 707 |
-
[GlmDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
| 708 |
-
)
|
| 709 |
-
self.norm = GlmRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 710 |
-
self.rotary_emb = GlmRotaryEmbedding(
|
| 711 |
-
dim=config.head_dim,
|
| 712 |
-
max_position_embeddings=config.max_position_embeddings,
|
| 713 |
-
base=config.rope_theta,
|
| 714 |
-
rotary_percent=self.rotary_percent,
|
| 715 |
-
)
|
| 716 |
-
self.gradient_checkpointing = False
|
| 717 |
-
|
| 718 |
-
# Initialize weights and apply final processing
|
| 719 |
-
self.post_init()
|
| 720 |
-
|
| 721 |
-
def get_input_embeddings(self):
|
| 722 |
-
return self.embed_tokens
|
| 723 |
-
|
| 724 |
-
def set_input_embeddings(self, value):
|
| 725 |
-
self.embed_tokens = value
|
| 726 |
-
|
| 727 |
-
@add_start_docstrings_to_model_forward(GLM_INPUTS_DOCSTRING)
|
| 728 |
-
def forward(
|
| 729 |
-
self,
|
| 730 |
-
input_ids: torch.LongTensor = None,
|
| 731 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 732 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 733 |
-
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
| 734 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 735 |
-
use_cache: Optional[bool] = None,
|
| 736 |
-
output_attentions: Optional[bool] = None,
|
| 737 |
-
output_hidden_states: Optional[bool] = None,
|
| 738 |
-
return_dict: Optional[bool] = None,
|
| 739 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 740 |
-
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 741 |
-
) -> Union[Tuple, BaseModelOutputWithPast]:
|
| 742 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 743 |
-
output_hidden_states = (
|
| 744 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 745 |
-
)
|
| 746 |
-
|
| 747 |
-
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 748 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 749 |
-
|
| 750 |
-
if (input_ids is None) ^ (inputs_embeds is not None):
|
| 751 |
-
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
| 752 |
-
|
| 753 |
-
if self.gradient_checkpointing and self.training and use_cache:
|
| 754 |
-
logger.warning_once(
|
| 755 |
-
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
| 756 |
-
)
|
| 757 |
-
use_cache = False
|
| 758 |
-
|
| 759 |
-
if inputs_embeds is None:
|
| 760 |
-
inputs_embeds = self.embed_tokens(input_ids)
|
| 761 |
-
|
| 762 |
-
# kept for BC (non `Cache` `past_key_values` inputs)
|
| 763 |
-
return_legacy_cache = False
|
| 764 |
-
if use_cache and not isinstance(past_key_values, Cache):
|
| 765 |
-
return_legacy_cache = True
|
| 766 |
-
if past_key_values is None:
|
| 767 |
-
past_key_values = DynamicCache()
|
| 768 |
-
else:
|
| 769 |
-
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
| 770 |
-
logger.warning_once(
|
| 771 |
-
"We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
|
| 772 |
-
"will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
|
| 773 |
-
"(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
|
| 774 |
-
)
|
| 775 |
-
|
| 776 |
-
if cache_position is None:
|
| 777 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 778 |
-
cache_position = torch.arange(
|
| 779 |
-
past_seen_tokens,
|
| 780 |
-
past_seen_tokens + inputs_embeds.shape[1],
|
| 781 |
-
device=inputs_embeds.device,
|
| 782 |
-
)
|
| 783 |
-
if position_ids is None:
|
| 784 |
-
position_ids = cache_position.unsqueeze(0)
|
| 785 |
-
|
| 786 |
-
causal_mask = self._update_causal_mask(
|
| 787 |
-
attention_mask,
|
| 788 |
-
inputs_embeds,
|
| 789 |
-
cache_position,
|
| 790 |
-
past_key_values,
|
| 791 |
-
output_attentions,
|
| 792 |
-
)
|
| 793 |
-
hidden_states = inputs_embeds
|
| 794 |
-
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 795 |
-
|
| 796 |
-
# decoder layers
|
| 797 |
-
all_hidden_states = () if output_hidden_states else None
|
| 798 |
-
all_self_attns = () if output_attentions else None
|
| 799 |
-
next_decoder_cache = None
|
| 800 |
-
|
| 801 |
-
for decoder_layer in self.layers:
|
| 802 |
-
if output_hidden_states:
|
| 803 |
-
all_hidden_states += (hidden_states,)
|
| 804 |
-
|
| 805 |
-
if self.gradient_checkpointing and self.training:
|
| 806 |
-
layer_outputs = self._gradient_checkpointing_func(
|
| 807 |
-
decoder_layer.__call__,
|
| 808 |
-
hidden_states,
|
| 809 |
-
causal_mask,
|
| 810 |
-
position_ids,
|
| 811 |
-
past_key_values,
|
| 812 |
-
output_attentions,
|
| 813 |
-
use_cache,
|
| 814 |
-
cache_position,
|
| 815 |
-
position_embeddings,
|
| 816 |
-
)
|
| 817 |
-
else:
|
| 818 |
-
layer_outputs = decoder_layer(
|
| 819 |
-
hidden_states,
|
| 820 |
-
attention_mask=causal_mask,
|
| 821 |
-
position_ids=position_ids,
|
| 822 |
-
past_key_value=past_key_values,
|
| 823 |
-
output_attentions=output_attentions,
|
| 824 |
-
use_cache=use_cache,
|
| 825 |
-
cache_position=cache_position,
|
| 826 |
-
position_embeddings=position_embeddings,
|
| 827 |
-
**flash_attn_kwargs,
|
| 828 |
-
)
|
| 829 |
-
|
| 830 |
-
hidden_states = layer_outputs[0]
|
| 831 |
-
|
| 832 |
-
if use_cache:
|
| 833 |
-
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
| 834 |
-
|
| 835 |
-
if output_attentions:
|
| 836 |
-
all_self_attns += (layer_outputs[1],)
|
| 837 |
-
|
| 838 |
-
hidden_states = self.norm(hidden_states)
|
| 839 |
-
|
| 840 |
-
# add hidden states from the last decoder layer
|
| 841 |
-
if output_hidden_states:
|
| 842 |
-
all_hidden_states += (hidden_states,)
|
| 843 |
-
|
| 844 |
-
next_cache = next_decoder_cache if use_cache else None
|
| 845 |
-
if return_legacy_cache:
|
| 846 |
-
next_cache = next_cache.to_legacy_cache()
|
| 847 |
-
|
| 848 |
-
if not return_dict:
|
| 849 |
-
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
| 850 |
-
return BaseModelOutputWithPast(
|
| 851 |
-
last_hidden_state=hidden_states,
|
| 852 |
-
past_key_values=next_cache,
|
| 853 |
-
hidden_states=all_hidden_states,
|
| 854 |
-
attentions=all_self_attns,
|
| 855 |
-
)
|
| 856 |
-
|
| 857 |
-
def _update_causal_mask(
|
| 858 |
-
self,
|
| 859 |
-
attention_mask: torch.Tensor,
|
| 860 |
-
input_tensor: torch.Tensor,
|
| 861 |
-
cache_position: torch.Tensor,
|
| 862 |
-
past_key_values: Cache,
|
| 863 |
-
output_attentions: bool,
|
| 864 |
-
):
|
| 865 |
-
if self.config._attn_implementation == "flash_attention_2":
|
| 866 |
-
if attention_mask is not None and 0.0 in attention_mask:
|
| 867 |
-
return attention_mask
|
| 868 |
-
return None
|
| 869 |
-
|
| 870 |
-
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
|
| 871 |
-
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
|
| 872 |
-
# to infer the attention mask.
|
| 873 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 874 |
-
using_static_cache = isinstance(past_key_values, StaticCache)
|
| 875 |
-
|
| 876 |
-
# When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
|
| 877 |
-
if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
|
| 878 |
-
if AttentionMaskConverter._ignore_causal_mask_sdpa(
|
| 879 |
-
attention_mask,
|
| 880 |
-
inputs_embeds=input_tensor,
|
| 881 |
-
past_key_values_length=past_seen_tokens,
|
| 882 |
-
is_training=self.training,
|
| 883 |
-
):
|
| 884 |
-
return None
|
| 885 |
-
|
| 886 |
-
dtype, device = input_tensor.dtype, input_tensor.device
|
| 887 |
-
sequence_length = input_tensor.shape[1]
|
| 888 |
-
if using_static_cache:
|
| 889 |
-
target_length = past_key_values.get_max_cache_shape()
|
| 890 |
-
else:
|
| 891 |
-
target_length = (
|
| 892 |
-
attention_mask.shape[-1]
|
| 893 |
-
if isinstance(attention_mask, torch.Tensor)
|
| 894 |
-
else past_seen_tokens + sequence_length + 1
|
| 895 |
-
)
|
| 896 |
-
|
| 897 |
-
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
| 898 |
-
causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
|
| 899 |
-
attention_mask,
|
| 900 |
-
sequence_length=sequence_length,
|
| 901 |
-
target_length=target_length,
|
| 902 |
-
dtype=dtype,
|
| 903 |
-
device=device,
|
| 904 |
-
cache_position=cache_position,
|
| 905 |
-
batch_size=input_tensor.shape[0],
|
| 906 |
-
)
|
| 907 |
-
|
| 908 |
-
if (
|
| 909 |
-
self.config._attn_implementation == "sdpa"
|
| 910 |
-
and attention_mask is not None
|
| 911 |
-
and attention_mask.device.type == "cuda"
|
| 912 |
-
and not output_attentions
|
| 913 |
-
):
|
| 914 |
-
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
| 915 |
-
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
| 916 |
-
# Details: https://github.com/pytorch/pytorch/issues/110213
|
| 917 |
-
min_dtype = torch.finfo(dtype).min
|
| 918 |
-
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
| 919 |
-
|
| 920 |
-
return causal_mask
|
| 921 |
-
|
| 922 |
-
@staticmethod
|
| 923 |
-
def _prepare_4d_causal_attention_mask_with_cache_position(
|
| 924 |
-
attention_mask: torch.Tensor,
|
| 925 |
-
sequence_length: int,
|
| 926 |
-
target_length: int,
|
| 927 |
-
dtype: torch.dtype,
|
| 928 |
-
device: torch.device,
|
| 929 |
-
cache_position: torch.Tensor,
|
| 930 |
-
batch_size: int,
|
| 931 |
-
**kwargs,
|
| 932 |
-
):
|
| 933 |
-
"""
|
| 934 |
-
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
| 935 |
-
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
| 936 |
-
|
| 937 |
-
Args:
|
| 938 |
-
attention_mask (`torch.Tensor`):
|
| 939 |
-
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
|
| 940 |
-
`(batch_size, 1, query_length, key_value_length)`.
|
| 941 |
-
sequence_length (`int`):
|
| 942 |
-
The sequence length being processed.
|
| 943 |
-
target_length (`int`):
|
| 944 |
-
The target length: when generating with static cache, the mask should be as long as the static cache,
|
| 945 |
-
to account for the 0 padding, the part of the cache that is not filled yet.
|
| 946 |
-
dtype (`torch.dtype`):
|
| 947 |
-
The dtype to use for the 4D attention mask.
|
| 948 |
-
device (`torch.device`):
|
| 949 |
-
The device to plcae the 4D attention mask on.
|
| 950 |
-
cache_position (`torch.Tensor`):
|
| 951 |
-
Indices depicting the position of the input sequence tokens in the sequence.
|
| 952 |
-
batch_size (`torch.Tensor`):
|
| 953 |
-
Batch size.
|
| 954 |
-
"""
|
| 955 |
-
if attention_mask is not None and attention_mask.dim() == 4:
|
| 956 |
-
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
| 957 |
-
causal_mask = attention_mask
|
| 958 |
-
else:
|
| 959 |
-
min_dtype = torch.finfo(dtype).min
|
| 960 |
-
causal_mask = torch.full(
|
| 961 |
-
(sequence_length, target_length),
|
| 962 |
-
fill_value=min_dtype,
|
| 963 |
-
dtype=dtype,
|
| 964 |
-
device=device,
|
| 965 |
-
)
|
| 966 |
-
if sequence_length != 1:
|
| 967 |
-
causal_mask = torch.triu(causal_mask, diagonal=1)
|
| 968 |
-
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
| 969 |
-
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
| 970 |
-
if attention_mask is not None:
|
| 971 |
-
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
| 972 |
-
mask_length = attention_mask.shape[-1]
|
| 973 |
-
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
|
| 974 |
-
padding_mask = padding_mask == 0
|
| 975 |
-
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
| 976 |
-
padding_mask, min_dtype
|
| 977 |
-
)
|
| 978 |
-
|
| 979 |
-
return causal_mask
|
| 980 |
-
|
| 981 |
-
|
| 982 |
-
class GlmForCausalLM(GlmPreTrainedModel, GenerationMixin):
|
| 983 |
-
_tied_weights_keys = ["lm_head.weight"]
|
| 984 |
-
|
| 985 |
-
def __init__(self, config: GlmConfig):
|
| 986 |
-
super().__init__(config)
|
| 987 |
-
self.model = GlmModel(config)
|
| 988 |
-
self.vocab_size = config.vocab_size
|
| 989 |
-
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 990 |
-
|
| 991 |
-
# Initialize weights and apply final processing
|
| 992 |
-
self.post_init()
|
| 993 |
-
|
| 994 |
-
def get_input_embeddings(self):
|
| 995 |
-
return self.model.embed_tokens
|
| 996 |
-
|
| 997 |
-
def set_input_embeddings(self, value):
|
| 998 |
-
self.model.embed_tokens = value
|
| 999 |
-
|
| 1000 |
-
def get_output_embeddings(self):
|
| 1001 |
-
return self.lm_head
|
| 1002 |
-
|
| 1003 |
-
def set_output_embeddings(self, new_embeddings):
|
| 1004 |
-
self.lm_head = new_embeddings
|
| 1005 |
-
|
| 1006 |
-
def set_decoder(self, decoder):
|
| 1007 |
-
self.model = decoder
|
| 1008 |
-
|
| 1009 |
-
def get_decoder(self):
|
| 1010 |
-
return self.model
|
| 1011 |
-
|
| 1012 |
-
@add_start_docstrings_to_model_forward(GLM_INPUTS_DOCSTRING)
|
| 1013 |
-
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
| 1014 |
-
def forward(
|
| 1015 |
-
self,
|
| 1016 |
-
input_ids: torch.LongTensor = None,
|
| 1017 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 1018 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 1019 |
-
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
| 1020 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1021 |
-
labels: Optional[torch.LongTensor] = None,
|
| 1022 |
-
use_cache: Optional[bool] = None,
|
| 1023 |
-
output_attentions: Optional[bool] = None,
|
| 1024 |
-
output_hidden_states: Optional[bool] = None,
|
| 1025 |
-
return_dict: Optional[bool] = None,
|
| 1026 |
-
cache_position: Optional[torch.LongTensor] = None,
|
| 1027 |
-
num_logits_to_keep: int = 0,
|
| 1028 |
-
**loss_kwargs,
|
| 1029 |
-
) -> Union[Tuple, CausalLMOutputWithPast]:
|
| 1030 |
-
r"""
|
| 1031 |
-
Args:
|
| 1032 |
-
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1033 |
-
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 1034 |
-
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 1035 |
-
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 1036 |
-
|
| 1037 |
-
num_logits_to_keep (`int`, *optional*):
|
| 1038 |
-
Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
|
| 1039 |
-
`input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
|
| 1040 |
-
token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
|
| 1041 |
-
|
| 1042 |
-
Returns:
|
| 1043 |
-
|
| 1044 |
-
Example:
|
| 1045 |
-
|
| 1046 |
-
```python
|
| 1047 |
-
>>> from transformers import AutoTokenizer, GlmForCausalLM
|
| 1048 |
-
|
| 1049 |
-
>>> model = GlmForCausalLM.from_pretrained("google/glm-7b")
|
| 1050 |
-
>>> tokenizer = AutoTokenizer.from_pretrained("google/glm-7b")
|
| 1051 |
-
|
| 1052 |
-
>>> prompt = "What is your favorite condiment?"
|
| 1053 |
-
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
| 1054 |
-
|
| 1055 |
-
>>> # Generate
|
| 1056 |
-
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
| 1057 |
-
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 1058 |
-
"What is your favorite condiment?"
|
| 1059 |
-
```"""
|
| 1060 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1061 |
-
output_hidden_states = (
|
| 1062 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1063 |
-
)
|
| 1064 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1065 |
-
|
| 1066 |
-
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
| 1067 |
-
outputs = self.model(
|
| 1068 |
-
input_ids=input_ids,
|
| 1069 |
-
attention_mask=attention_mask,
|
| 1070 |
-
position_ids=position_ids,
|
| 1071 |
-
past_key_values=past_key_values,
|
| 1072 |
-
inputs_embeds=inputs_embeds,
|
| 1073 |
-
use_cache=use_cache,
|
| 1074 |
-
output_attentions=output_attentions,
|
| 1075 |
-
output_hidden_states=output_hidden_states,
|
| 1076 |
-
return_dict=return_dict,
|
| 1077 |
-
cache_position=cache_position,
|
| 1078 |
-
)
|
| 1079 |
-
|
| 1080 |
-
hidden_states = outputs[0]
|
| 1081 |
-
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
| 1082 |
-
logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
|
| 1083 |
-
|
| 1084 |
-
loss = None
|
| 1085 |
-
if labels is not None:
|
| 1086 |
-
loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
|
| 1087 |
-
|
| 1088 |
-
if not return_dict:
|
| 1089 |
-
output = (logits,) + outputs[1:]
|
| 1090 |
-
return (loss,) + output if loss is not None else output
|
| 1091 |
-
|
| 1092 |
-
return CausalLMOutputWithPast(
|
| 1093 |
-
loss=loss,
|
| 1094 |
-
logits=logits,
|
| 1095 |
-
past_key_values=outputs.past_key_values,
|
| 1096 |
-
hidden_states=outputs.hidden_states,
|
| 1097 |
-
attentions=outputs.attentions,
|
| 1098 |
-
)
|
| 1099 |
-
|
| 1100 |
-
|
| 1101 |
-
@add_start_docstrings(
|
| 1102 |
-
"""
|
| 1103 |
-
The Glm Model transformer with a sequence classification head on top (linear layer).
|
| 1104 |
-
|
| 1105 |
-
[`GlmForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
| 1106 |
-
(e.g. GPT-2) do.
|
| 1107 |
-
|
| 1108 |
-
Since it does classification on the last token, it requires to know the position of the last token. If a
|
| 1109 |
-
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
| 1110 |
-
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
| 1111 |
-
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
| 1112 |
-
each row of the batch).
|
| 1113 |
-
""",
|
| 1114 |
-
GLM_START_DOCSTRING,
|
| 1115 |
-
)
|
| 1116 |
-
class GlmForSequenceClassification(GlmPreTrainedModel):
|
| 1117 |
-
def __init__(self, config: GlmConfig):
|
| 1118 |
-
super().__init__(config)
|
| 1119 |
-
self.num_labels = config.num_labels
|
| 1120 |
-
self.model = GlmModel(config)
|
| 1121 |
-
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
| 1122 |
-
|
| 1123 |
-
# Initialize weights and apply final processing
|
| 1124 |
-
self.post_init()
|
| 1125 |
-
|
| 1126 |
-
def get_input_embeddings(self):
|
| 1127 |
-
return self.model.embed_tokens
|
| 1128 |
-
|
| 1129 |
-
def set_input_embeddings(self, value):
|
| 1130 |
-
self.model.embed_tokens = value
|
| 1131 |
-
|
| 1132 |
-
@add_start_docstrings_to_model_forward(GLM_INPUTS_DOCSTRING)
|
| 1133 |
-
def forward(
|
| 1134 |
-
self,
|
| 1135 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 1136 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 1137 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 1138 |
-
past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
|
| 1139 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1140 |
-
labels: Optional[torch.LongTensor] = None,
|
| 1141 |
-
use_cache: Optional[bool] = None,
|
| 1142 |
-
output_attentions: Optional[bool] = None,
|
| 1143 |
-
output_hidden_states: Optional[bool] = None,
|
| 1144 |
-
return_dict: Optional[bool] = None,
|
| 1145 |
-
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
| 1146 |
-
r"""
|
| 1147 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1148 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 1149 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1150 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1151 |
-
"""
|
| 1152 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1153 |
-
|
| 1154 |
-
transformer_outputs = self.model(
|
| 1155 |
-
input_ids,
|
| 1156 |
-
attention_mask=attention_mask,
|
| 1157 |
-
position_ids=position_ids,
|
| 1158 |
-
past_key_values=past_key_values,
|
| 1159 |
-
inputs_embeds=inputs_embeds,
|
| 1160 |
-
use_cache=use_cache,
|
| 1161 |
-
output_attentions=output_attentions,
|
| 1162 |
-
output_hidden_states=output_hidden_states,
|
| 1163 |
-
return_dict=return_dict,
|
| 1164 |
-
)
|
| 1165 |
-
hidden_states = transformer_outputs[0]
|
| 1166 |
-
logits = self.score(hidden_states)
|
| 1167 |
-
|
| 1168 |
-
if input_ids is not None:
|
| 1169 |
-
batch_size = input_ids.shape[0]
|
| 1170 |
-
else:
|
| 1171 |
-
batch_size = inputs_embeds.shape[0]
|
| 1172 |
-
|
| 1173 |
-
if self.config.pad_token_id is None and batch_size != 1:
|
| 1174 |
-
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
|
| 1175 |
-
if self.config.pad_token_id is None:
|
| 1176 |
-
sequence_lengths = -1
|
| 1177 |
-
else:
|
| 1178 |
-
if input_ids is not None:
|
| 1179 |
-
# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
|
| 1180 |
-
sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
|
| 1181 |
-
sequence_lengths = sequence_lengths % input_ids.shape[-1]
|
| 1182 |
-
sequence_lengths = sequence_lengths.to(logits.device)
|
| 1183 |
-
else:
|
| 1184 |
-
sequence_lengths = -1
|
| 1185 |
-
|
| 1186 |
-
pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
|
| 1187 |
-
|
| 1188 |
-
loss = None
|
| 1189 |
-
if labels is not None:
|
| 1190 |
-
loss = self.loss_function(
|
| 1191 |
-
logits=logits,
|
| 1192 |
-
labels=labels,
|
| 1193 |
-
pooled_logits=pooled_logits,
|
| 1194 |
-
config=self.config,
|
| 1195 |
-
)
|
| 1196 |
-
|
| 1197 |
-
if not return_dict:
|
| 1198 |
-
output = (pooled_logits,) + transformer_outputs[1:]
|
| 1199 |
-
return ((loss,) + output) if loss is not None else output
|
| 1200 |
-
|
| 1201 |
-
return SequenceClassifierOutputWithPast(
|
| 1202 |
-
loss=loss,
|
| 1203 |
-
logits=pooled_logits,
|
| 1204 |
-
past_key_values=transformer_outputs.past_key_values,
|
| 1205 |
-
hidden_states=transformer_outputs.hidden_states,
|
| 1206 |
-
attentions=transformer_outputs.attentions,
|
| 1207 |
-
)
|
| 1208 |
-
|
| 1209 |
-
|
| 1210 |
-
@add_start_docstrings(
|
| 1211 |
-
"""
|
| 1212 |
-
The Glm Model transformer with a token classification head on top (a linear layer on top of the hidden-states
|
| 1213 |
-
output) e.g. for Named-Entity-Recognition (NER) tasks.
|
| 1214 |
-
""",
|
| 1215 |
-
GLM_START_DOCSTRING,
|
| 1216 |
-
)
|
| 1217 |
-
class GlmForTokenClassification(GlmPreTrainedModel):
|
| 1218 |
-
def __init__(self, config: GlmConfig):
|
| 1219 |
-
super().__init__(config)
|
| 1220 |
-
self.num_labels = config.num_labels
|
| 1221 |
-
self.model = GlmModel(config)
|
| 1222 |
-
if getattr(config, "classifier_dropout", None) is not None:
|
| 1223 |
-
classifier_dropout = config.classifier_dropout
|
| 1224 |
-
elif getattr(config, "hidden_dropout", None) is not None:
|
| 1225 |
-
classifier_dropout = config.hidden_dropout
|
| 1226 |
-
else:
|
| 1227 |
-
classifier_dropout = 0.1
|
| 1228 |
-
self.dropout = nn.Dropout(classifier_dropout)
|
| 1229 |
-
self.score = nn.Linear(config.hidden_size, config.num_labels)
|
| 1230 |
-
|
| 1231 |
-
# Initialize weights and apply final processing
|
| 1232 |
-
self.post_init()
|
| 1233 |
-
|
| 1234 |
-
def get_input_embeddings(self):
|
| 1235 |
-
return self.model.embed_tokens
|
| 1236 |
-
|
| 1237 |
-
def set_input_embeddings(self, value):
|
| 1238 |
-
self.model.embed_tokens = value
|
| 1239 |
-
|
| 1240 |
-
@add_start_docstrings_to_model_forward(GLM_INPUTS_DOCSTRING)
|
| 1241 |
-
@add_code_sample_docstrings(
|
| 1242 |
-
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 1243 |
-
output_type=TokenClassifierOutput,
|
| 1244 |
-
config_class=_CONFIG_FOR_DOC,
|
| 1245 |
-
)
|
| 1246 |
-
def forward(
|
| 1247 |
-
self,
|
| 1248 |
-
input_ids: Optional[torch.LongTensor] = None,
|
| 1249 |
-
attention_mask: Optional[torch.Tensor] = None,
|
| 1250 |
-
position_ids: Optional[torch.LongTensor] = None,
|
| 1251 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1252 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1253 |
-
labels: Optional[torch.LongTensor] = None,
|
| 1254 |
-
use_cache: Optional[bool] = None,
|
| 1255 |
-
output_attentions: Optional[bool] = None,
|
| 1256 |
-
output_hidden_states: Optional[bool] = None,
|
| 1257 |
-
return_dict: Optional[bool] = None,
|
| 1258 |
-
) -> Union[Tuple, TokenClassifierOutput]:
|
| 1259 |
-
r"""
|
| 1260 |
-
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1261 |
-
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
| 1262 |
-
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
| 1263 |
-
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
| 1264 |
-
"""
|
| 1265 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1266 |
-
|
| 1267 |
-
outputs = self.model(
|
| 1268 |
-
input_ids,
|
| 1269 |
-
attention_mask=attention_mask,
|
| 1270 |
-
position_ids=position_ids,
|
| 1271 |
-
past_key_values=past_key_values,
|
| 1272 |
-
inputs_embeds=inputs_embeds,
|
| 1273 |
-
use_cache=use_cache,
|
| 1274 |
-
output_attentions=output_attentions,
|
| 1275 |
-
output_hidden_states=output_hidden_states,
|
| 1276 |
-
return_dict=return_dict,
|
| 1277 |
-
)
|
| 1278 |
-
sequence_output = outputs[0]
|
| 1279 |
-
sequence_output = self.dropout(sequence_output)
|
| 1280 |
-
logits = self.score(sequence_output)
|
| 1281 |
-
|
| 1282 |
-
loss = None
|
| 1283 |
-
if labels is not None:
|
| 1284 |
-
loss = self.loss_function(logits, labels, self.config)
|
| 1285 |
-
|
| 1286 |
-
if not return_dict:
|
| 1287 |
-
output = (logits,) + outputs[2:]
|
| 1288 |
-
return ((loss,) + output) if loss is not None else output
|
| 1289 |
-
|
| 1290 |
-
return TokenClassifierOutput(
|
| 1291 |
-
loss=loss,
|
| 1292 |
-
logits=logits,
|
| 1293 |
-
hidden_states=outputs.hidden_states,
|
| 1294 |
-
attentions=outputs.attentions,
|
| 1295 |
-
)
|
| 1296 |
-
|
| 1297 |
-
|
| 1298 |
-
__all__ = [
|
| 1299 |
-
"GlmPreTrainedModel",
|
| 1300 |
-
"GlmModel",
|
| 1301 |
-
"GlmForCausalLM",
|
| 1302 |
-
"GlmForSequenceClassification",
|
| 1303 |
-
"GlmForTokenClassification",
|
| 1304 |
-
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|