modeling_openelm.py
| 1 | # |
| 2 | # For licensing see accompanying LICENSE file. |
| 3 | # Copyright (C) 2024 Apple Inc. All Rights Reserved. |
| 4 | # |
| 5 | |
| 6 | from typing import List, Optional, Tuple, Union |
| 7 | |
| 8 | import torch |
| 9 | import torch.utils.checkpoint |
| 10 | from torch import Tensor, nn |
| 11 | from torch.nn import CrossEntropyLoss |
| 12 | from torch.nn import functional as F |
| 13 | from transformers import PreTrainedModel |
| 14 | from transformers.activations import ACT2FN |
| 15 | from transformers.cache_utils import Cache, DynamicCache, StaticCache |
| 16 | from transformers.modeling_outputs import ( |
| 17 | BaseModelOutputWithPast, |
| 18 | CausalLMOutputWithPast, |
| 19 | ) |
| 20 | from transformers.utils import logging |
| 21 | |
| 22 | logger = logging.get_logger(__name__) |
| 23 | |
| 24 | # this import has to be relative, otherwise, when setting trust_remote_code=True |
| 25 | # huggingface transformers won't be able to load the module correctly |
| 26 | from .configuration_openelm import OpenELMConfig, make_divisible |
| 27 | |
| 28 | |
| 29 | class OpenELMRMSNorm(nn.Module): |
| 30 | def __init__(self, num_features: int, eps: float = 1e-6): |
| 31 | """ |
| 32 | Initialize the OpenELMRMSNorm normalization layer. |
| 33 | |
| 34 | Args: |
| 35 | dim (int): The dimension of the input tensor. |
| 36 | eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6. |
| 37 | |
| 38 | Attributes: |
| 39 | eps (float): A small value added to the denominator for numerical stability. |
| 40 | weight (nn.Parameter): Learnable scaling parameter. |
| 41 | |
| 42 | """ |
| 43 | super().__init__() |
| 44 | self.eps = eps |
| 45 | self.weight = nn.Parameter(torch.ones(num_features)) |
| 46 | self.num_features = num_features |
| 47 | |
| 48 | def _norm(self, x: Tensor) -> Tensor: |
| 49 | """ |
| 50 | Apply the OpenELMRMSNorm normalization to the input tensor. |
| 51 | |
| 52 | Args: |
| 53 | x (torch.Tensor): The input tensor. |
| 54 | |
| 55 | Returns: |
| 56 | torch.Tensor: The normalized tensor. |
| 57 | |
| 58 | """ |
| 59 | return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) |
| 60 | |
| 61 | def forward(self, x: Tensor) -> Tensor: |
| 62 | """ |
| 63 | Forward pass through the OpenELMRMSNorm layer. |
| 64 | |
| 65 | Args: |
| 66 | x (torch.Tensor): The input tensor. |
| 67 | |
| 68 | Returns: |
| 69 | torch.Tensor: The output tensor after applying OpenELMRMSNorm. |
| 70 | |
| 71 | """ |
| 72 | output = self._norm(x.float()).type_as(x) |
| 73 | return output * self.weight |
| 74 | |
| 75 | def extra_repr(self) -> str: |
| 76 | return ( |
| 77 | super().extra_repr() + f"num_features={self.num_features}, eps={self.eps}" |
| 78 | ) |
| 79 | |
| 80 | |
| 81 | class OpenELMPreTrainedModel(PreTrainedModel): |
| 82 | config_class = OpenELMConfig |
| 83 | base_model_prefix = "transformer" |
| 84 | supports_gradient_checkpointing = True |
| 85 | _no_split_modules = ["OpenELMDecoderLayer"] |
| 86 | _skip_keys_device_placement = "past_key_values" |
| 87 | |
| 88 | def __init__(self, *inputs, **kwargs) -> None: |
| 89 | super().__init__(*inputs, **kwargs) |
| 90 | |
| 91 | def _init_weights(self, module: nn.Module) -> None: |
| 92 | """Initialize the weights.""" |
| 93 | if isinstance(module, nn.Linear): |
| 94 | # Slightly different from the TF version which uses truncated_normal for initialization |
| 95 | # cf https://github.com/pytorch/pytorch/pull/5617 |
| 96 | module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) |
| 97 | if module.bias is not None: |
| 98 | module.bias.data.zero_() |
| 99 | elif isinstance(module, nn.Embedding): |
| 100 | module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) |
| 101 | if module.padding_idx is not None: |
| 102 | module.weight.data[module.padding_idx].zero_() |
| 103 | elif isinstance(module, OpenELMRMSNorm): |
| 104 | module.weight.data.fill_(1.0) |
| 105 | |
| 106 | |
| 107 | def _rotate_half(x: Tensor) -> Tensor: |
| 108 | x1, x2 = x.chunk(2, dim=-1) |
| 109 | return torch.cat((-x2, x1), dim=-1) |
| 110 | |
| 111 | |
| 112 | def _apply_rotary_pos_emb(x: Tensor, pos_sin: Tensor, pos_cos: Tensor) -> Tensor: |
| 113 | return (x * pos_cos) + (_rotate_half(x) * pos_sin) |
| 114 | |
| 115 | |
| 116 | class OpenELMRotaryEmbedding(torch.nn.Module): |
| 117 | """ |
| 118 | The rotary position embeddings (aka RoPE) from `RoFormer <https://arxiv.org/abs/2104.09864>`_. |
| 119 | |
| 120 | RoPE encodes the position information of tokens using a rotation matrix, and is able to capture |
| 121 | explicit relative positional dependencies. |
| 122 | |
| 123 | Args: |
| 124 | model_dim: The dimensionality of the model's hidden state. |
| 125 | max_seq_length: Maximum sequence length. |
| 126 | freq_constant: A constant used for computing frequencies. |
| 127 | """ |
| 128 | |
| 129 | def __init__( |
| 130 | self, model_dim: int, max_seq_length: int, freq_constant: int = 10000 |
| 131 | ) -> None: |
| 132 | inv_freq = 1.0 / ( |
| 133 | freq_constant |
| 134 | ** (torch.arange(0, model_dim, 2, dtype=torch.float32) / model_dim) |
| 135 | ) |
| 136 | super().__init__() |
| 137 | |
| 138 | self.model_dim = model_dim |
| 139 | self.freq_constant = freq_constant |
| 140 | self.max_seq_length = max_seq_length |
| 141 | |
| 142 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 143 | self._cached_cos = None |
| 144 | self._cached_sin = None |
| 145 | self._cached_seq_length = max_seq_length |
| 146 | self._compute_sin_cos_embeddings(max_seq_length) |
| 147 | |
| 148 | def extra_repr(self) -> str: |
| 149 | return f"\tmodel_dim={self.model_dim}, max_seq_length={self.max_seq_length}, freq_constant={self.freq_constant}" |
| 150 | |
| 151 | def _compute_sin_cos_embeddings( |
| 152 | self, |
| 153 | key_len: int, |
| 154 | key_device: torch.device = torch.device("cpu"), |
| 155 | key_dtype: torch.dtype = torch.float32, |
| 156 | ) -> None: |
| 157 | """ |
| 158 | Compute sine and cos embeddings. |
| 159 | |
| 160 | Args: |
| 161 | key_len: Number of tokens in the key embeddings in the transformer model. |
| 162 | device: Device where the key embeddings are stored. |
| 163 | key_dtype: Data type of the key embeddings. |
| 164 | |
| 165 | Returns: |
| 166 | None |
| 167 | |
| 168 | ...note: |
| 169 | We recalculate the sine and cosine embeddings if any of the following conditions are met: |
| 170 | 1. The number of tokens in key embeddings are greater than the cached sequence length. |
| 171 | 2. Sine and cosine caches are empty. |
| 172 | 3. The device and data type of sine and cosine embeddings does not match with the key embeddings. |
| 173 | """ |
| 174 | if ( |
| 175 | key_len > self._cached_seq_length |
| 176 | or self._cached_cos is None |
| 177 | or (self._cached_cos is not None and self._cached_cos.device != key_device) |
| 178 | or (self._cached_cos is not None and self._cached_cos.dtype != key_dtype) |
| 179 | or self._cached_sin is None |
| 180 | or (self._cached_sin is not None and self._cached_sin.device != key_device) |
| 181 | or (self._cached_sin is not None and self._cached_sin.dtype != key_dtype) |
| 182 | ): |
| 183 | self._cached_seq_length = max(key_len, self._cached_seq_length) |
| 184 | |
| 185 | # The shape of 'pos_index' is [number of key tokens] |
| 186 | pos_index = torch.arange( |
| 187 | self._cached_seq_length, |
| 188 | dtype=torch.float32, |
| 189 | device=self.inv_freq.device, |
| 190 | ) |
| 191 | # The shape of 'pos_index_theta' is [number of key tokens, model dimension] |
| 192 | pos_index_theta = torch.einsum("i,j->ij", pos_index, self.inv_freq) |
| 193 | # The shape of 'emb' is [number of key tokens, model dimension] |
| 194 | emb = torch.cat((pos_index_theta, pos_index_theta), dim=-1) |
| 195 | |
| 196 | # the shape of cos and sin embeddings is [number of key tokens, model_dim] |
| 197 | cos_emb = emb.cos().to(dtype=key_dtype, device=key_device) |
| 198 | sin_emb = emb.sin().to(dtype=key_dtype, device=key_device) |
| 199 | |
| 200 | # the shape of cached cos and sin embeddings is [1, 1, number of key tokens, model_dim] |
| 201 | self._cached_cos = cos_emb[None, None, :, :] |
| 202 | self._cached_sin = sin_emb[None, None, :, :] |
| 203 | |
| 204 | def forward( |
| 205 | self, |
| 206 | query: torch.Tensor, |
| 207 | key: torch.Tensor, |
| 208 | ) -> Tuple[torch.Tensor, torch.Tensor]: |
| 209 | """ |
| 210 | The forward function of RoPE embeddings. |
| 211 | |
| 212 | Args: |
| 213 | query: Query embeddings in the transformer model. The shape of query embeddings is |
| 214 | [Batch, number of query heads, number of query tokens, model dimension]. |
| 215 | key: Key embeddings in the transformer model. The shape of key embeddings is |
| 216 | [Batch, number of key heads, number of key tokens, model dimension]. |
| 217 | |
| 218 | Returns: |
| 219 | A tuple containing the query and key embeddings with positional information. The shape of the returned query |
| 220 | and key embeddings is the same as the input query and key embeddings respectively. |
| 221 | |
| 222 | ...note: |
| 223 | The RoPE embedding computation is done in full-precision. After the computation, input query and key tensors |
| 224 | are casted to original input datatype. |
| 225 | """ |
| 226 | dim = key.shape[-1] |
| 227 | key_len = key.shape[2] |
| 228 | query_len = query.shape[2] |
| 229 | |
| 230 | assert dim == self.model_dim |
| 231 | assert key.device == query.device |
| 232 | assert key.dtype == query.dtype |
| 233 | |
| 234 | # In the context of self-attention, the lengths of keys and queries are equal. |
| 235 | # However, in generation tasks, such as predicting the next token in a sequence, the lengths of keys and queries |
| 236 | # can differ. For instance, when employing key-value (KV) caching for sequence prediction, the keys |
| 237 | # represent embeddings of previous tokens and the current token, while the query corresponds |
| 238 | # to the embedding of the current token only. |
| 239 | assert ( |
| 240 | key_len >= query_len |
| 241 | ), "Number of keys has to be greater than or equal to number of queries." |
| 242 | |
| 243 | query_float = query.float() |
| 244 | key_float = key.float() |
| 245 | |
| 246 | self._compute_sin_cos_embeddings( |
| 247 | key_len, key_device=key_float.device, key_dtype=key_float.dtype |
| 248 | ) |
| 249 | query_float = _apply_rotary_pos_emb( |
| 250 | x=query_float, |
| 251 | pos_sin=self._cached_sin[..., key_len - query_len : key_len, :], |
| 252 | pos_cos=self._cached_cos[..., key_len - query_len : key_len, :], |
| 253 | ) |
| 254 | key_float = _apply_rotary_pos_emb( |
| 255 | x=key_float, |
| 256 | pos_sin=self._cached_sin[..., :key_len, :], |
| 257 | pos_cos=self._cached_cos[..., :key_len, :], |
| 258 | ) |
| 259 | |
| 260 | return query_float.type_as(query), key_float.type_as(key) |
| 261 | |
| 262 | |
| 263 | class OpenELMMultiHeadCausalAttention(nn.Module): |
| 264 | def __init__(self, config: OpenELMConfig, layer_idx: int) -> None: |
| 265 | super().__init__() |
| 266 | self.layer_idx = layer_idx |
| 267 | head_dim = config.head_dim |
| 268 | q_heads = config.num_query_heads[layer_idx] |
| 269 | k_heads = config.num_kv_heads[layer_idx] |
| 270 | v_heads = config.num_kv_heads[layer_idx] |
| 271 | |
| 272 | self.qkv_proj = nn.Linear( |
| 273 | in_features=config.model_dim, |
| 274 | out_features=(q_heads + k_heads + v_heads) * head_dim, |
| 275 | bias=False, |
| 276 | ) |
| 277 | |
| 278 | self.pos_embedding = OpenELMRotaryEmbedding( |
| 279 | model_dim=config.head_dim, |
| 280 | max_seq_length=config.rope_max_length, |
| 281 | freq_constant=config.rope_freq_constant, |
| 282 | ) |
| 283 | |
| 284 | if config.normalize_qk_projections: |
| 285 | self.q_norm = OpenELMRMSNorm( |
| 286 | num_features=config.head_dim, |
| 287 | ) |
| 288 | self.k_norm = OpenELMRMSNorm( |
| 289 | num_features=config.head_dim, |
| 290 | ) |
| 291 | else: |
| 292 | self.q_norm = None |
| 293 | self.k_norm = None |
| 294 | |
| 295 | self.out_proj = nn.Linear( |
| 296 | in_features=q_heads * head_dim, |
| 297 | out_features=config.model_dim, |
| 298 | bias=False, |
| 299 | ) |
| 300 | |
| 301 | self.head_dim = config.head_dim |
| 302 | self.num_q_heads = q_heads |
| 303 | self.num_k_heads = k_heads |
| 304 | self.num_v_heads = v_heads |
| 305 | self.transformer_dim = config.model_dim |
| 306 | self.num_groups = self.num_q_heads // self.num_k_heads |
| 307 | |
| 308 | def extra_repr(self) -> str: |
| 309 | return ( |
| 310 | super().extra_repr() |
| 311 | + f"query_heads={self.num_q_heads}, key_heads={self.num_k_heads}, value_heads={self.num_v_heads}" |
| 312 | ) |
| 313 | |
| 314 | def forward( |
| 315 | self, |
| 316 | hidden_states: torch.Tensor, |
| 317 | attention_mask: Optional[torch.Tensor] = None, |
| 318 | past_key_value: Optional[Cache] = None, |
| 319 | output_attentions: bool = False, |
| 320 | use_cache: bool = False, |
| 321 | cache_position: Optional[torch.LongTensor] = None, |
| 322 | ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: |
| 323 | """ |
| 324 | Forward pass of multi-head self-attention. |
| 325 | |
| 326 | Args: |
| 327 | hidden_states: Input tensor of the shape [batch size, sequence length, model dimension]. |
| 328 | past_key_value: Tensor storing the cached keys and values. |
| 329 | output_attentions: output attention weights. |
| 330 | use_cache: Specifies whether to use kv-cache for generation. |
| 331 | cache_position: used for updating the kv-cache. |
| 332 | |
| 333 | Returns: |
| 334 | The output of the same shape as the input, optionally with a tensor containing cached keys and values. |
| 335 | """ |
| 336 | |
| 337 | # scaled_dot_product_attention does not return attention weights, set output_attentions to False |
| 338 | output_attentions = False |
| 339 | batch_size, seq_length, d_model = hidden_states.size() |
| 340 | |
| 341 | # [B, S, d] --> [B, S, (q_h + k_h + v_h) * h] |
| 342 | qkv = self.qkv_proj(hidden_states) |
| 343 | # [B, S, (q_h + k_h + v_h) * h] --> [B, S, (q_h + k_h + v_h), h] |
| 344 | qkv = qkv.reshape( |
| 345 | batch_size, |
| 346 | seq_length, |
| 347 | self.num_q_heads + self.num_k_heads + self.num_v_heads, |
| 348 | self.head_dim, |
| 349 | ) |
| 350 | # [B, S, (q_h + k_h + v_h), h] --> [B, (q_h + k_h + v_h), S, h] |
| 351 | qkv = qkv.transpose(1, 2) |
| 352 | # [B, (q_h + k_h + v_h), S, h] --> [B, q_h, S h], [B, k_h, S, h], [B, v_h, S, h] |
| 353 | queries, keys, values = qkv.split( |
| 354 | [self.num_q_heads, self.num_k_heads, self.num_v_heads], dim=1 |
| 355 | ) |
| 356 | |
| 357 | if self.q_norm is not None: |
| 358 | queries = self.q_norm(queries) |
| 359 | |
| 360 | if self.k_norm is not None: |
| 361 | keys = self.k_norm(keys) |
| 362 | |
| 363 | past_key_value = getattr(self, "past_key_value", past_key_value) |
| 364 | |
| 365 | if past_key_value is not None: |
| 366 | # sin and cos are specific to RoPE models; position_ids needed for the static cache |
| 367 | # cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} |
| 368 | cache_kwargs = {"cache_position": cache_position} |
| 369 | keys, values = past_key_value.update( |
| 370 | keys, values, self.layer_idx, cache_kwargs |
| 371 | ) |
| 372 | |
| 373 | # Add positional embedding |
| 374 | queries, keys = self.pos_embedding(queries, keys) |
| 375 | |
| 376 | if self.num_groups != 1: |
| 377 | # GQA |
| 378 | # [B, k_h, S, h] --> [B, q_h, S, h] |
| 379 | keys = keys.repeat_interleave(self.num_groups, dim=1) |
| 380 | # [B, v_h, S, h] --> [B, q_h, S, h] |
| 381 | values = values.repeat_interleave(self.num_groups, dim=1) |
| 382 | |
| 383 | causal_mask = attention_mask |
| 384 | if attention_mask is not None and cache_position is not None: |
| 385 | causal_mask = causal_mask[:, :, cache_position, : keys.shape[-2]] |
| 386 | |
| 387 | attn_output = F.scaled_dot_product_attention( |
| 388 | queries, |
| 389 | keys, |
| 390 | values, |
| 391 | attn_mask=causal_mask, |
| 392 | dropout_p=0, |
| 393 | ) |
| 394 | |
| 395 | attn_output = attn_output.transpose(1, 2).contiguous() |
| 396 | attn_output = attn_output.reshape( |
| 397 | batch_size, seq_length, self.num_q_heads * self.head_dim |
| 398 | ) |
| 399 | attn_output = self.out_proj(attn_output) |
| 400 | if not output_attentions: |
| 401 | attn_weights = None |
| 402 | return attn_output, attn_weights, past_key_value |
| 403 | |
| 404 | |
| 405 | class OpenELMFeedForwardNetwork(nn.Module): |
| 406 | def __init__(self, config: OpenELMConfig, layer_idx: int) -> None: |
| 407 | super().__init__() |
| 408 | ffn_multiplier = config.ffn_multipliers[layer_idx] |
| 409 | intermediate_dim = int( |
| 410 | make_divisible( |
| 411 | ffn_multiplier * config.model_dim, |
| 412 | divisor=config.ffn_dim_divisor, |
| 413 | ) |
| 414 | ) |
| 415 | if config.ffn_with_glu: |
| 416 | # FFN with Gated linear unit, as described in https://arxiv.org/abs/2002.05202v1. |
| 417 | self.proj_1 = nn.Linear( |
| 418 | in_features=config.model_dim, |
| 419 | out_features=2 * intermediate_dim, |
| 420 | bias=False, |
| 421 | ) |
| 422 | self.proj_2 = nn.Linear( |
| 423 | in_features=intermediate_dim, |
| 424 | out_features=config.model_dim, |
| 425 | bias=False, |
| 426 | ) |
| 427 | self.ffn_with_glu = True |
| 428 | else: |
| 429 | # Standard FFN, as described in https://arxiv.org/abs/1706.03762 |
| 430 | self.proj_1 = nn.Linear( |
| 431 | in_features=config.model_dim, |
| 432 | out_features=intermediate_dim, |
| 433 | bias=False, |
| 434 | ) |
| 435 | self.proj_2 = nn.Linear( |
| 436 | in_features=intermediate_dim, |
| 437 | out_features=config.model_dim, |
| 438 | bias=False, |
| 439 | ) |
| 440 | self.ffn_with_glu = False |
| 441 | |
| 442 | self.act = ACT2FN[config.activation_fn_name] |
| 443 | |
| 444 | def extra_repr(self) -> str: |
| 445 | return super().extra_repr() + f"(ffn_with_glu) : {self.ffn_with_glu}" |
| 446 | |
| 447 | def forward(self, x: Tensor) -> Tensor: |
| 448 | """Forward function of FFN layer. |
| 449 | |
| 450 | Args: |
| 451 | x: Input tensor of the shape [batch size, sequence length, model dimension]. |
| 452 | |
| 453 | Returns: |
| 454 | A tensor of the same shape as the input. |
| 455 | """ |
| 456 | if self.ffn_with_glu: |
| 457 | y_12 = self.proj_1(x) |
| 458 | y_1, y_2 = y_12.chunk(2, dim=-1) |
| 459 | y = self.act(y_1) * y_2 |
| 460 | return self.proj_2(y) |
| 461 | else: |
| 462 | return self.proj_2(self.act(self.proj_1(x))) |
| 463 | |
| 464 | |
| 465 | class OpenELMDecoderLayer(nn.Module): |
| 466 | def __init__(self, config: OpenELMConfig, layer_idx: int) -> None: |
| 467 | super().__init__() |
| 468 | self.attn = OpenELMMultiHeadCausalAttention(config=config, layer_idx=layer_idx) |
| 469 | self.ffn = OpenELMFeedForwardNetwork(config=config, layer_idx=layer_idx) |
| 470 | self.ffn_norm = OpenELMRMSNorm( |
| 471 | num_features=config.model_dim, |
| 472 | ) |
| 473 | self.attn_norm = OpenELMRMSNorm( |
| 474 | num_features=config.model_dim, |
| 475 | ) |
| 476 | |
| 477 | def forward( |
| 478 | self, |
| 479 | hidden_states: torch.Tensor, |
| 480 | attention_mask: Optional[torch.Tensor] = None, |
| 481 | position_ids: Optional[torch.LongTensor] = None, |
| 482 | past_key_value: Optional[Tuple[torch.Tensor]] = None, |
| 483 | output_attentions: Optional[bool] = False, |
| 484 | use_cache: Optional[bool] = False, |
| 485 | cache_position: Optional[torch.LongTensor] = None, |
| 486 | **kwargs, |
| 487 | ) -> Tuple[ |
| 488 | torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]] |
| 489 | ]: |
| 490 | """ |
| 491 | Args: |
| 492 | hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` |
| 493 | attention_mask (`torch.FloatTensor`, *optional*): |
| 494 | attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1, |
| 495 | query_sequence_length, key_sequence_length)` if default attention is used. |
| 496 | output_attentions (`bool`, *optional*): |
| 497 | Whether or not to return the attentions tensors of all attention layers. See `attentions` under |
| 498 | returned tensors for more detail. |
| 499 | use_cache (`bool`, *optional*): |
| 500 | If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding |
| 501 | (see `past_key_values`). |
| 502 | past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states |
| 503 | """ |
| 504 | residual = hidden_states |
| 505 | hidden_states = self.attn_norm(hidden_states) |
| 506 | |
| 507 | # Self Attention |
| 508 | hidden_states, self_attn_weights, present_key_value = self.attn( |
| 509 | hidden_states=hidden_states, |
| 510 | attention_mask=attention_mask, |
| 511 | past_key_value=past_key_value, |
| 512 | output_attentions=output_attentions, |
| 513 | use_cache=use_cache, |
| 514 | cache_position=cache_position, |
| 515 | **kwargs, |
| 516 | ) |
| 517 | hidden_states = residual + hidden_states |
| 518 | |
| 519 | # Fully Connected |
| 520 | residual = hidden_states |
| 521 | hidden_states = self.ffn_norm(hidden_states) |
| 522 | hidden_states = self.ffn(hidden_states) |
| 523 | hidden_states = residual + hidden_states |
| 524 | |
| 525 | outputs = (hidden_states,) |
| 526 | |
| 527 | if output_attentions: |
| 528 | outputs += (self_attn_weights,) |
| 529 | |
| 530 | if use_cache: |
| 531 | outputs += (present_key_value,) |
| 532 | |
| 533 | return outputs |
| 534 | |
| 535 | |
| 536 | class OpenELMModel(OpenELMPreTrainedModel): |
| 537 | config_class = OpenELMConfig |
| 538 | |
| 539 | def __init__(self, config: OpenELMConfig): |
| 540 | super().__init__(config) |
| 541 | self.config = config |
| 542 | |
| 543 | self.token_embeddings = nn.Embedding( |
| 544 | embedding_dim=config.model_dim, |
| 545 | num_embeddings=config.vocab_size, |
| 546 | ) |
| 547 | |
| 548 | self.layers = nn.ModuleList( |
| 549 | OpenELMDecoderLayer(config=config, layer_idx=layer_idx) |
| 550 | for layer_idx in range(config.num_transformer_layers) |
| 551 | ) |
| 552 | self.norm = OpenELMRMSNorm(num_features=config.model_dim) |
| 553 | if config.share_input_output_layers: |
| 554 | self.classifier = None |
| 555 | else: |
| 556 | self.classifier = nn.Linear( |
| 557 | in_features=config.model_dim, |
| 558 | out_features=config.vocab_size, |
| 559 | bias=False, |
| 560 | ) |
| 561 | self.num_transformer_layers = config.num_transformer_layers |
| 562 | self.gradient_checkpointing = False |
| 563 | |
| 564 | # Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class. |
| 565 | # NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_context_length`. |
| 566 | causal_mask = torch.full( |
| 567 | (config.max_context_length, config.max_context_length), |
| 568 | fill_value=True, |
| 569 | dtype=torch.bool, |
| 570 | ) |
| 571 | self.register_buffer( |
| 572 | "causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False |
| 573 | ) |
| 574 | |
| 575 | # Initialize weights and apply final processing |
| 576 | self.post_init() |
| 577 | self.reset_parameters(config=config) |
| 578 | |
| 579 | def get_input_embeddings(self): |
| 580 | return self.token_embeddings |
| 581 | |
| 582 | def set_input_embeddings(self, new_embeddings: torch.Tensor): |
| 583 | self.token_embeddings = new_embeddings |
| 584 | |
| 585 | def reset_parameters(self, config: OpenELMConfig) -> None: |
| 586 | """Initialize the layers in Language Model |
| 587 | |
| 588 | The initialization scheme is followed, following `OPT <https://arxiv.org/pdf/2205.01068.pdf>`_. |
| 589 | |
| 590 | Args: |
| 591 | use_megatron_std: Use standard deviation as described in Megatron-LM. |
| 592 | |
| 593 | Returns: |
| 594 | None |
| 595 | """ |
| 596 | for module in self.modules(): |
| 597 | if isinstance(module, nn.Linear): |
| 598 | std = module.in_features**-0.5 |
| 599 | torch.nn.init.normal_(module.weight, mean=0.0, std=std) |
| 600 | if module.bias is not None: |
| 601 | torch.nn.init.zeros_(module.bias) |
| 602 | elif isinstance(module, nn.Embedding): |
| 603 | std = module.embedding_dim**-0.5 |
| 604 | torch.nn.init.normal_(module.weight, mean=0.0, std=std) |
| 605 | elif isinstance(module, OpenELMRMSNorm): |
| 606 | if module.weight is not None: |
| 607 | torch.nn.init.ones_(module.weight) |
| 608 | if hasattr(module, "bias") and module.bias is not None: |
| 609 | torch.nn.init.zeros_(module.bias) |
| 610 | |
| 611 | model_dim = config.model_dim |
| 612 | n_layers = config.num_transformer_layers |
| 613 | std = (model_dim**-0.5) * ((2 * n_layers) ** -0.5) |
| 614 | for param_name, param in self.named_parameters(): |
| 615 | if param_name.endswith("out_proj.weight") or param_name.endswith( |
| 616 | "ffn.proj_2.weight" |
| 617 | ): |
| 618 | torch.nn.init.normal_(param, mean=0.0, std=std) |
| 619 | |
| 620 | def forward( |
| 621 | self, |
| 622 | input_ids: torch.LongTensor = None, |
| 623 | attention_mask: Optional[torch.Tensor] = None, |
| 624 | position_ids: Optional[torch.LongTensor] = None, |
| 625 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 626 | inputs_embeds: Optional[torch.FloatTensor] = None, |
| 627 | use_cache: Optional[bool] = None, |
| 628 | output_attentions: Optional[bool] = None, |
| 629 | output_hidden_states: Optional[bool] = None, |
| 630 | return_dict: Optional[bool] = None, |
| 631 | cache_position: Optional[torch.LongTensor] = None, |
| 632 | ) -> Union[Tuple, BaseModelOutputWithPast]: |
| 633 | output_attentions = ( |
| 634 | output_attentions |
| 635 | if output_attentions is not None |
| 636 | else self.config.output_attentions |
| 637 | ) |
| 638 | output_hidden_states = ( |
| 639 | output_hidden_states |
| 640 | if output_hidden_states is not None |
| 641 | else self.config.output_hidden_states |
| 642 | ) |
| 643 | use_cache = use_cache if use_cache is not None else self.config.use_cache |
| 644 | return_dict = ( |
| 645 | return_dict if return_dict is not None else self.config.use_return_dict |
| 646 | ) |
| 647 | |
| 648 | if (input_ids is None) ^ (inputs_embeds is not None): |
| 649 | raise ValueError( |
| 650 | "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" |
| 651 | ) |
| 652 | |
| 653 | if self.gradient_checkpointing and self.training and use_cache: |
| 654 | logger.warning_once( |
| 655 | "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." |
| 656 | ) |
| 657 | use_cache = False |
| 658 | |
| 659 | if inputs_embeds is None: |
| 660 | inputs_embeds = self.token_embeddings(input_ids) |
| 661 | |
| 662 | past_seen_tokens = 0 |
| 663 | if use_cache: # kept for BC (cache positions) |
| 664 | if not isinstance(past_key_values, StaticCache): |
| 665 | past_key_values = DynamicCache.from_legacy_cache(past_key_values) |
| 666 | past_seen_tokens = past_key_values.get_seq_length() |
| 667 | |
| 668 | if cache_position is None: |
| 669 | cache_position = torch.arange( |
| 670 | past_seen_tokens, |
| 671 | past_seen_tokens + inputs_embeds.shape[1], |
| 672 | device=inputs_embeds.device, |
| 673 | ) |
| 674 | |
| 675 | if position_ids is None: |
| 676 | position_ids = cache_position.unsqueeze(0) |
| 677 | |
| 678 | causal_mask = self._update_causal_mask(attention_mask, inputs_embeds) |
| 679 | |
| 680 | # embed positions |
| 681 | hidden_states = inputs_embeds |
| 682 | |
| 683 | # decoder layers |
| 684 | all_hidden_states = () if output_hidden_states else None |
| 685 | all_self_attns = () if output_attentions else None |
| 686 | next_decoder_cache = None |
| 687 | |
| 688 | for decoder_layer in self.layers: |
| 689 | if output_hidden_states: |
| 690 | all_hidden_states += (hidden_states,) |
| 691 | |
| 692 | if self.gradient_checkpointing and self.training: |
| 693 | layer_outputs = self._gradient_checkpointing_func( |
| 694 | decoder_layer.__call__, |
| 695 | hidden_states, |
| 696 | causal_mask, |
| 697 | position_ids, |
| 698 | past_key_values, |
| 699 | output_attentions, |
| 700 | use_cache, |
| 701 | cache_position, |
| 702 | ) |
| 703 | else: |
| 704 | layer_outputs = decoder_layer( |
| 705 | hidden_states, |
| 706 | attention_mask=causal_mask, |
| 707 | position_ids=position_ids, |
| 708 | past_key_value=past_key_values, |
| 709 | output_attentions=output_attentions, |
| 710 | use_cache=use_cache, |
| 711 | cache_position=cache_position, |
| 712 | ) |
| 713 | |
| 714 | hidden_states = layer_outputs[0] |
| 715 | |
| 716 | if use_cache: |
| 717 | next_decoder_cache = layer_outputs[2 if output_attentions else 1] |
| 718 | |
| 719 | if output_attentions: |
| 720 | all_self_attns += (layer_outputs[1],) |
| 721 | |
| 722 | hidden_states = self.norm(hidden_states) |
| 723 | |
| 724 | # add hidden states from the last decoder layer |
| 725 | if output_hidden_states: |
| 726 | all_hidden_states += (hidden_states,) |
| 727 | |
| 728 | next_cache = None |
| 729 | if use_cache: |
| 730 | next_cache = ( |
| 731 | next_decoder_cache.to_legacy_cache() |
| 732 | if isinstance(next_decoder_cache, Cache) |
| 733 | else next_decoder_cache |
| 734 | ) |
| 735 | if not return_dict: |
| 736 | return tuple( |
| 737 | v |
| 738 | for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] |
| 739 | if v is not None |
| 740 | ) |
| 741 | return BaseModelOutputWithPast( |
| 742 | last_hidden_state=hidden_states, |
| 743 | past_key_values=next_cache, |
| 744 | hidden_states=all_hidden_states, |
| 745 | attentions=all_self_attns, |
| 746 | ) |
| 747 | |
| 748 | def _update_causal_mask(self, attention_mask, input_tensor): |
| 749 | if self.config._attn_implementation == "flash_attention_2": |
| 750 | if attention_mask is not None and 0.0 in attention_mask: |
| 751 | return attention_mask |
| 752 | return None |
| 753 | |
| 754 | batch_size, seq_length = input_tensor.shape[:2] |
| 755 | dtype = input_tensor.dtype |
| 756 | device = input_tensor.device |
| 757 | |
| 758 | # support going beyond cached `max_position_embedding` |
| 759 | if seq_length > self.causal_mask.shape[-1]: |
| 760 | causal_mask = torch.full( |
| 761 | (2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]), |
| 762 | fill_value=1, |
| 763 | ) |
| 764 | self.register_buffer( |
| 765 | "causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False |
| 766 | ) |
| 767 | |
| 768 | # We use the current dtype to avoid any overflows |
| 769 | min_dtype = torch.finfo(dtype).min |
| 770 | causal_mask = ( |
| 771 | self.causal_mask[None, None, :, :].repeat(batch_size, 1, 1, 1).to(dtype) |
| 772 | * min_dtype |
| 773 | ) |
| 774 | |
| 775 | causal_mask = causal_mask.to(dtype=dtype, device=device) |
| 776 | if attention_mask is not None and attention_mask.dim() == 2: |
| 777 | mask_length = attention_mask.shape[-1] |
| 778 | padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[ |
| 779 | :, None, None, : |
| 780 | ].eq(0.0) |
| 781 | causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill( |
| 782 | padding_mask, min_dtype |
| 783 | ) |
| 784 | |
| 785 | if self.config._attn_implementation == "sdpa" and attention_mask is not None: |
| 786 | # For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400). |
| 787 | is_tracing = ( |
| 788 | torch.jit.is_tracing() |
| 789 | or isinstance(input_tensor, torch.fx.Proxy) |
| 790 | or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling()) |
| 791 | ) |
| 792 | if not is_tracing and torch.any(attention_mask != 1): |
| 793 | # Attend to all tokens in masked rows from the causal_mask, for example the relevant first rows when |
| 794 | # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. |
| 795 | # Details: https://github.com/pytorch/pytorch/issues/110213 |
| 796 | causal_mask = causal_mask.mul( |
| 797 | ~torch.all(causal_mask == min_dtype, dim=-1, keepdim=True) |
| 798 | ).to(dtype) |
| 799 | |
| 800 | return causal_mask |
| 801 | |
| 802 | |
| 803 | class OpenELMForCausalLM(OpenELMPreTrainedModel): |
| 804 | _tied_weights_keys = ["lm_head.weight"] |
| 805 | |
| 806 | def __init__(self, config: OpenELMConfig): |
| 807 | super().__init__(config) |
| 808 | self.transformer = OpenELMModel(config) |
| 809 | self.vocab_size = config.vocab_size |
| 810 | if config.share_input_output_layers: |
| 811 | self.lm_head = None |
| 812 | else: |
| 813 | self.lm_head = nn.Linear(config.model_dim, config.vocab_size, bias=False) |
| 814 | |
| 815 | # Initialize weights and apply final processing |
| 816 | self.post_init() |
| 817 | |
| 818 | def get_input_embeddings(self): |
| 819 | return self.transformer.token_embeddings |
| 820 | |
| 821 | def set_input_embeddings(self, value): |
| 822 | self.transformer.token_embeddings = value |
| 823 | |
| 824 | def get_output_embeddings(self): |
| 825 | return self.lm_head |
| 826 | |
| 827 | def set_output_embeddings(self, new_embeddings): |
| 828 | self.lm_head = new_embeddings |
| 829 | |
| 830 | def set_decoder(self, decoder): |
| 831 | self.transformer = decoder |
| 832 | |
| 833 | def get_decoder(self): |
| 834 | return self.transformer |
| 835 | |
| 836 | def forward( |
| 837 | self, |
| 838 | input_ids: torch.LongTensor = None, |
| 839 | attention_mask: Optional[torch.Tensor] = None, |
| 840 | position_ids: Optional[torch.LongTensor] = None, |
| 841 | past_key_values: Optional[List[torch.FloatTensor]] = None, |
| 842 | inputs_embeds: Optional[torch.FloatTensor] = None, |
| 843 | labels: Optional[torch.LongTensor] = None, |
| 844 | use_cache: Optional[bool] = None, |
| 845 | output_attentions: Optional[bool] = None, |
| 846 | output_hidden_states: Optional[bool] = None, |
| 847 | return_dict: Optional[bool] = None, |
| 848 | cache_position: Optional[torch.LongTensor] = None, |
| 849 | ) -> Union[Tuple, CausalLMOutputWithPast]: |
| 850 | output_attentions = ( |
| 851 | output_attentions |
| 852 | if output_attentions is not None |
| 853 | else self.config.output_attentions |
| 854 | ) |
| 855 | output_hidden_states = ( |
| 856 | output_hidden_states |
| 857 | if output_hidden_states is not None |
| 858 | else self.config.output_hidden_states |
| 859 | ) |
| 860 | return_dict = ( |
| 861 | return_dict if return_dict is not None else self.config.use_return_dict |
| 862 | ) |
| 863 | # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) |
| 864 | outputs = self.transformer( |
| 865 | input_ids=input_ids, |
| 866 | attention_mask=attention_mask, |
| 867 | position_ids=position_ids, |
| 868 | past_key_values=past_key_values, |
| 869 | inputs_embeds=inputs_embeds, |
| 870 | use_cache=use_cache, |
| 871 | output_attentions=output_attentions, |
| 872 | output_hidden_states=output_hidden_states, |
| 873 | return_dict=return_dict, |
| 874 | cache_position=cache_position, |
| 875 | ) |
| 876 | |
| 877 | hidden_states = outputs[0] |
| 878 | if self.lm_head is None: |
| 879 | # shared |
| 880 | logits = F.linear( |
| 881 | hidden_states, weight=self.transformer.token_embeddings.weight |
| 882 | ) |
| 883 | else: |
| 884 | logits = self.lm_head(hidden_states) |
| 885 | logits = logits[:, : self.config.vocab_size] |
| 886 | loss = None |
| 887 | if labels is not None: |
| 888 | # Shift so that tokens < n predict n |
| 889 | shift_logits = logits[..., :-1, :].contiguous() |
| 890 | shift_labels = labels[..., 1:].contiguous() |
| 891 | # Flatten the tokens |
| 892 | loss_fct = CrossEntropyLoss() |
| 893 | shift_logits = shift_logits.view(-1, self.config.vocab_size) |
| 894 | shift_labels = shift_labels.view(-1) |
| 895 | # Enable model parallelism |
| 896 | shift_labels = shift_labels.to(shift_logits.device) |
| 897 | loss = loss_fct(shift_logits, shift_labels) |
| 898 | |
| 899 | if not return_dict: |
| 900 | output = (logits,) + outputs[1:] |
| 901 | return (loss,) + output if loss is not None else output |
| 902 | |
| 903 | return CausalLMOutputWithPast( |
| 904 | loss=loss, |
| 905 | logits=logits, |
| 906 | past_key_values=outputs.past_key_values, |
| 907 | hidden_states=outputs.hidden_states, |
| 908 | attentions=outputs.attentions, |
| 909 | ) |
| 910 | |
| 911 | def prepare_inputs_for_generation( |
| 912 | self, |
| 913 | input_ids, |
| 914 | past_key_values=None, |
| 915 | attention_mask=None, |
| 916 | inputs_embeds=None, |
| 917 | **kwargs, |
| 918 | ): |
| 919 | past_length = 0 |
| 920 | if past_key_values is not None: |
| 921 | if isinstance(past_key_values, Cache): |
| 922 | cache_length = past_key_values.get_seq_length() |
| 923 | past_length = past_key_values.seen_tokens |
| 924 | max_cache_length = past_key_values.get_max_length() |
| 925 | else: |
| 926 | cache_length = past_length = past_key_values[0][0].shape[2] |
| 927 | max_cache_length = None |
| 928 | |
| 929 | # Keep only the unprocessed tokens: |
| 930 | # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where |
| 931 | # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as |
| 932 | # input) |
| 933 | if ( |
| 934 | attention_mask is not None |
| 935 | and attention_mask.shape[1] > input_ids.shape[1] |
| 936 | ): |
| 937 | input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] |
| 938 | # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard |
| 939 | # input_ids based on the past_length. |
| 940 | elif past_length < input_ids.shape[1]: |
| 941 | input_ids = input_ids[:, past_length:] |
| 942 | # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. |
| 943 | |
| 944 | # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. |
| 945 | if ( |
| 946 | max_cache_length is not None |
| 947 | and attention_mask is not None |
| 948 | and cache_length + input_ids.shape[1] > max_cache_length |
| 949 | ): |
| 950 | attention_mask = attention_mask[:, -max_cache_length:] |
| 951 | |
| 952 | position_ids = kwargs.get("position_ids", None) |
| 953 | if attention_mask is not None and position_ids is None: |
| 954 | # create position_ids on the fly for batch generation |
| 955 | position_ids = attention_mask.long().cumsum(-1) - 1 |
| 956 | position_ids.masked_fill_(attention_mask == 0, 1) |
| 957 | if past_key_values: |
| 958 | position_ids = position_ids[:, -input_ids.shape[1] :] |
| 959 | |
| 960 | if self.generation_config.cache_implementation == "static": |
| 961 | # generation with static cache |
| 962 | cache_position = kwargs.get("cache_position", None) |
| 963 | if cache_position is None: |
| 964 | past_length = 0 |
| 965 | else: |
| 966 | past_length = cache_position[-1] + 1 |
| 967 | input_ids = input_ids[:, past_length:] |
| 968 | position_ids = position_ids[:, past_length:] |
| 969 | |
| 970 | # we should only keep a `cache_position` in generate, and do +=1. |
| 971 | # same goes for position ids. Could also help with continued generation. |
| 972 | cache_position = torch.arange( |
| 973 | past_length, |
| 974 | past_length + position_ids.shape[-1], |
| 975 | device=position_ids.device, |
| 976 | ) |
| 977 | |
| 978 | # if `inputs_embeds` are passed, we only want to use them in the 1st generation step |
| 979 | if inputs_embeds is not None and past_key_values is None: |
| 980 | model_inputs = {"inputs_embeds": inputs_embeds} |
| 981 | else: |
| 982 | # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise |
| 983 | # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114 |
| 984 | # We could use `next_tokens` directly instead. |
| 985 | model_inputs = {"input_ids": input_ids.contiguous()} |
| 986 | |
| 987 | model_inputs.update( |
| 988 | { |
| 989 | "position_ids": position_ids.contiguous(), |
| 990 | "cache_position": cache_position, |
| 991 | "past_key_values": past_key_values, |
| 992 | "use_cache": kwargs.get("use_cache"), |
| 993 | "attention_mask": attention_mask, |
| 994 | } |
| 995 | ) |
| 996 | return model_inputs |
| 997 | |
| 998 | @staticmethod |
| 999 | def _reorder_cache(past_key_values, beam_idx): |
| 1000 | reordered_past = () |
| 1001 | for layer_past in past_key_values: |
| 1002 | reordered_past += ( |
| 1003 | tuple( |
| 1004 | past_state.index_select(0, beam_idx.to(past_state.device)) |
| 1005 | for past_state in layer_past |
| 1006 | ), |
| 1007 | ) |
| 1008 | return reordered_past |
| 1009 | |