| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Adapted from |
| 4 | # https://github.com/zai-org/ChatGLM2-6B |
| 5 | """Inference-only ChatGLM model compatible with THUDM weights.""" |
| 6 | |
| 7 | import json |
| 8 | from collections.abc import Iterable |
| 9 | from itertools import islice |
| 10 | |
| 11 | import torch |
| 12 | from torch import nn |
| 13 | from torch.nn import LayerNorm |
| 14 | |
| 15 | from vllm.compilation.decorators import support_torch_compile |
| 16 | from vllm.config import CacheConfig, VllmConfig |
| 17 | from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size |
| 18 | from vllm.model_executor.layers.activation import SiluAndMul |
| 19 | from vllm.model_executor.layers.attention import Attention |
| 20 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 21 | from vllm.model_executor.layers.linear import ( |
| 22 | MergedColumnParallelLinear, |
| 23 | QKVParallelLinear, |
| 24 | RowParallelLinear, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 27 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 28 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 29 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 30 | ParallelLMHead, |
| 31 | VocabParallelEmbedding, |
| 32 | ) |
| 33 | from vllm.sequence import IntermediateTensors |
| 34 | from vllm.transformers_utils.configs.chatglm import ChatGLMConfig |
| 35 | |
| 36 | from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant |
| 37 | from .utils import ( |
| 38 | AutoWeightsLoader, |
| 39 | WeightsMapper, |
| 40 | make_empty_intermediate_tensors_factory, |
| 41 | make_layers, |
| 42 | maybe_prefix, |
| 43 | ) |
| 44 | |
| 45 | |
| 46 | class GLMAttention(nn.Module): |
| 47 | def __init__( |
| 48 | self, |
| 49 | config: ChatGLMConfig, |
| 50 | cache_config: CacheConfig | None = None, |
| 51 | quant_config: QuantizationConfig | None = None, |
| 52 | prefix: str = "", |
| 53 | ): |
| 54 | super().__init__() |
| 55 | self.hidden_size = config.hidden_size |
| 56 | tp_size = get_tensor_model_parallel_world_size() |
| 57 | self.total_num_heads = config.num_attention_heads |
| 58 | assert self.total_num_heads % tp_size == 0 |
| 59 | self.num_heads = self.total_num_heads // tp_size |
| 60 | self.multi_query_attention = config.multi_query_attention |
| 61 | self.total_num_kv_heads = ( |
| 62 | config.multi_query_group_num |
| 63 | if config.multi_query_attention |
| 64 | else config.num_attention_heads |
| 65 | ) |
| 66 | if self.total_num_kv_heads >= tp_size: |
| 67 | # Number of KV heads is greater than TP size, so we partition |
| 68 | # the KV heads across multiple tensor parallel GPUs. |
| 69 | assert self.total_num_kv_heads % tp_size == 0 |
| 70 | else: |
| 71 | # Number of KV heads is less than TP size, so we replicate |
| 72 | # the KV heads across multiple tensor parallel GPUs. |
| 73 | assert tp_size % self.total_num_kv_heads == 0 |
| 74 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 75 | self.head_dim = config.hidden_size // self.total_num_heads |
| 76 | self.q_size = self.num_heads * self.head_dim |
| 77 | self.kv_size = self.num_kv_heads * self.head_dim |
| 78 | self.scaling = self.head_dim**-0.5 |
| 79 | |
| 80 | self.query_key_value = QKVParallelLinear( |
| 81 | self.hidden_size, |
| 82 | self.head_dim, |
| 83 | self.total_num_heads, |
| 84 | self.total_num_kv_heads, |
| 85 | bias=config.add_bias_linear or config.add_qkv_bias, |
| 86 | quant_config=quant_config, |
| 87 | prefix=f"{prefix}.query_key_value", |
| 88 | ) |
| 89 | self.dense = RowParallelLinear( |
| 90 | self.total_num_heads * self.head_dim, |
| 91 | config.hidden_size, |
| 92 | bias=config.add_bias_linear, |
| 93 | quant_config=quant_config, |
| 94 | prefix=f"{prefix}.dense", |
| 95 | ) |
| 96 | |
| 97 | # https://huggingface.co/zai-org/chatglm3-6b-32k/blob/e210410255278dd9d74463cf396ba559c0ef801c/modeling_chatglm.py#L141 |
| 98 | rope_ratio = getattr(config, "rope_ratio", 1.0) |
| 99 | max_positions = getattr(config, "seq_length", 8192) |
| 100 | rope_parameters = { |
| 101 | "rope_type": "default", |
| 102 | "rope_theta": 10000 * rope_ratio, |
| 103 | "partial_rotary_factor": 0.5, |
| 104 | } |
| 105 | # NOTE: zai-org/cogagent-9b-20241220 uses original_rope=False, |
| 106 | # which is equivalent to is_neox_style=True |
| 107 | is_neox_style = not config.original_rope |
| 108 | self.rotary_emb = get_rope( |
| 109 | self.head_dim, |
| 110 | max_position=max_positions, |
| 111 | rope_parameters=rope_parameters, |
| 112 | is_neox_style=is_neox_style, |
| 113 | ) |
| 114 | self.attn = Attention( |
| 115 | self.num_heads, |
| 116 | self.head_dim, |
| 117 | self.scaling, |
| 118 | num_kv_heads=self.num_kv_heads, |
| 119 | cache_config=cache_config, |
| 120 | quant_config=quant_config, |
| 121 | prefix=f"{prefix}.attn", |
| 122 | ) |
| 123 | |
| 124 | def forward( |
| 125 | self, |
| 126 | hidden_states: torch.Tensor, |
| 127 | position_ids: torch.Tensor, |
| 128 | ) -> torch.Tensor: |
| 129 | qkv, _ = self.query_key_value(hidden_states) |
| 130 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 131 | q, k = self.rotary_emb(position_ids, q, k) |
| 132 | context_layer = self.attn(q, k, v) |
| 133 | attn_output, _ = self.dense(context_layer) |
| 134 | return attn_output |
| 135 | |
| 136 | |
| 137 | class GLMMLP(nn.Module): |
| 138 | """MLP. |
| 139 | |
| 140 | MLP will take the input with h hidden state, project it to 4*h |
| 141 | hidden dimension, perform nonlinear transformation, and project the |
| 142 | state back into h hidden dimension. |
| 143 | """ |
| 144 | |
| 145 | def __init__( |
| 146 | self, |
| 147 | config: ChatGLMConfig, |
| 148 | quant_config: QuantizationConfig | None = None, |
| 149 | prefix: str = "", |
| 150 | ): |
| 151 | super().__init__() |
| 152 | |
| 153 | self.add_bias = config.add_bias_linear |
| 154 | |
| 155 | # Project to 4h. |
| 156 | self.dense_h_to_4h = MergedColumnParallelLinear( |
| 157 | config.hidden_size, |
| 158 | [config.ffn_hidden_size] * 2, |
| 159 | bias=config.add_bias_linear, |
| 160 | quant_config=quant_config, |
| 161 | prefix=f"{prefix}.dense_h_to_4h", |
| 162 | ) |
| 163 | |
| 164 | self.activation_func = SiluAndMul() |
| 165 | |
| 166 | # Project back to h. |
| 167 | self.dense_4h_to_h = RowParallelLinear( |
| 168 | config.ffn_hidden_size, |
| 169 | config.hidden_size, |
| 170 | bias=config.add_bias_linear, |
| 171 | quant_config=quant_config, |
| 172 | prefix=f"{prefix}.dense_4h_to_h", |
| 173 | ) |
| 174 | |
| 175 | def forward(self, hidden_states): |
| 176 | # [s, b, 4hp] |
| 177 | intermediate_parallel, _ = self.dense_h_to_4h(hidden_states) |
| 178 | intermediate_parallel = self.activation_func(intermediate_parallel) |
| 179 | # [s, b, h] |
| 180 | output, _ = self.dense_4h_to_h(intermediate_parallel) |
| 181 | return output |
| 182 | |
| 183 | |
| 184 | class GLMBlock(nn.Module): |
| 185 | """A single transformer layer. |
| 186 | |
| 187 | Transformer layer takes input with size [s, b, h] and returns an |
| 188 | output of the same size. |
| 189 | """ |
| 190 | |
| 191 | def __init__( |
| 192 | self, |
| 193 | config: ChatGLMConfig, |
| 194 | cache_config: CacheConfig | None = None, |
| 195 | quant_config: QuantizationConfig | None = None, |
| 196 | prefix: str = "", |
| 197 | ): |
| 198 | super().__init__() |
| 199 | self.apply_residual_connection_post_layernorm = ( |
| 200 | config.apply_residual_connection_post_layernorm |
| 201 | ) |
| 202 | |
| 203 | self.fp32_residual_connection = config.fp32_residual_connection |
| 204 | |
| 205 | layer_norm_func = RMSNorm if config.rmsnorm else LayerNorm |
| 206 | # Layernorm on the input data. |
| 207 | self.input_layernorm = layer_norm_func( |
| 208 | config.hidden_size, eps=config.layernorm_epsilon |
| 209 | ) |
| 210 | |
| 211 | # Self attention. |
| 212 | self.self_attention = GLMAttention( |
| 213 | config, cache_config, quant_config, prefix=f"{prefix}.self_attention" |
| 214 | ) |
| 215 | self.hidden_dropout = config.hidden_dropout |
| 216 | |
| 217 | # Layernorm on the attention output |
| 218 | self.post_attention_layernorm = layer_norm_func( |
| 219 | config.hidden_size, eps=config.layernorm_epsilon |
| 220 | ) |
| 221 | |
| 222 | # MLP |
| 223 | self.mlp = GLMMLP(config, quant_config, prefix=f"{prefix}.mlp") |
| 224 | |
| 225 | def forward( |
| 226 | self, |
| 227 | hidden_states: torch.Tensor, |
| 228 | position_ids: torch.Tensor, |
| 229 | ) -> torch.Tensor: |
| 230 | # hidden_states: [num_tokens, h] |
| 231 | # Layer norm at the beginning of the transformer layer. |
| 232 | layernorm_output = self.input_layernorm(hidden_states) |
| 233 | # Self attention. |
| 234 | attention_output = self.self_attention( |
| 235 | hidden_states=layernorm_output, |
| 236 | position_ids=position_ids, |
| 237 | ) |
| 238 | |
| 239 | # Residual connection. |
| 240 | if self.apply_residual_connection_post_layernorm: |
| 241 | residual = layernorm_output |
| 242 | else: |
| 243 | residual = hidden_states |
| 244 | |
| 245 | layernorm_input = residual + attention_output |
| 246 | |
| 247 | # Layer norm post the self attention. |
| 248 | layernorm_output = self.post_attention_layernorm(layernorm_input) |
| 249 | |
| 250 | # Second residual connection. |
| 251 | if self.apply_residual_connection_post_layernorm: |
| 252 | residual = layernorm_output |
| 253 | else: |
| 254 | residual = layernorm_input |
| 255 | |
| 256 | output = self.mlp(layernorm_output) + residual |
| 257 | |
| 258 | return output |
| 259 | |
| 260 | |
| 261 | class GLMTransformer(nn.Module): |
| 262 | """Transformer class.""" |
| 263 | |
| 264 | def __init__( |
| 265 | self, |
| 266 | config: ChatGLMConfig, |
| 267 | cache_config: CacheConfig | None = None, |
| 268 | quant_config: QuantizationConfig | None = None, |
| 269 | prefix: str = "", |
| 270 | ): |
| 271 | super().__init__() |
| 272 | self.post_layer_norm = config.post_layer_norm |
| 273 | |
| 274 | # Number of layers. |
| 275 | self.num_layers = config.num_layers |
| 276 | |
| 277 | # Transformer layers. |
| 278 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 279 | self.num_layers, |
| 280 | lambda prefix: GLMBlock(config, cache_config, quant_config, prefix=prefix), |
| 281 | prefix=f"{prefix}.layers", |
| 282 | ) |
| 283 | |
| 284 | if self.post_layer_norm: |
| 285 | layer_norm_func = RMSNorm if config.rmsnorm else LayerNorm |
| 286 | # Final layer norm before output. |
| 287 | self.final_layernorm = layer_norm_func( |
| 288 | config.hidden_size, eps=config.layernorm_epsilon |
| 289 | ) |
| 290 | |
| 291 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 292 | ["hidden_states"], config.hidden_size |
| 293 | ) |
| 294 | |
| 295 | def forward( |
| 296 | self, |
| 297 | hidden_states: torch.Tensor, |
| 298 | position_ids: torch.Tensor, |
| 299 | ) -> torch.Tensor | IntermediateTensors: |
| 300 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 301 | hidden_states = layer( |
| 302 | hidden_states=hidden_states, position_ids=position_ids |
| 303 | ) |
| 304 | |
| 305 | if not get_pp_group().is_last_rank: |
| 306 | return IntermediateTensors({"hidden_states": hidden_states}) |
| 307 | |
| 308 | # Final layer norm. |
| 309 | if self.post_layer_norm: |
| 310 | hidden_states = self.final_layernorm(hidden_states) |
| 311 | |
| 312 | return hidden_states |
| 313 | |
| 314 | |
| 315 | @support_torch_compile |
| 316 | class ChatGLMModel(nn.Module, SupportsQuant): |
| 317 | hf_to_vllm_mapper = WeightsMapper( |
| 318 | orig_to_new_substr={".word_embeddings": ""}, |
| 319 | ) |
| 320 | |
| 321 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 322 | super().__init__() |
| 323 | |
| 324 | config = vllm_config.model_config.hf_config |
| 325 | cache_config = vllm_config.cache_config |
| 326 | quant_config = vllm_config.quant_config |
| 327 | |
| 328 | self.config = config |
| 329 | |
| 330 | self.embedding = VocabParallelEmbedding( |
| 331 | config.padded_vocab_size, |
| 332 | config.hidden_size, |
| 333 | quant_config=quant_config, |
| 334 | prefix=f"{prefix}.embedding", |
| 335 | ) |
| 336 | |
| 337 | self.num_layers = config.num_layers |
| 338 | self.multi_query_group_num = config.multi_query_group_num |
| 339 | self.kv_channels = config.kv_channels |
| 340 | self.encoder = GLMTransformer( |
| 341 | config, cache_config, quant_config, prefix=f"{prefix}.encoder" |
| 342 | ) |
| 343 | |
| 344 | self.output_layer = ParallelLMHead( |
| 345 | config.padded_vocab_size, |
| 346 | config.hidden_size, |
| 347 | quant_config=quant_config, |
| 348 | prefix=f"{prefix}.output_layer", |
| 349 | ) |
| 350 | |
| 351 | self.make_empty_intermediate_tensors = ( |
| 352 | self.encoder.make_empty_intermediate_tensors |
| 353 | ) |
| 354 | |
| 355 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 356 | return self.embedding(input_ids) |
| 357 | |
| 358 | def forward( |
| 359 | self, |
| 360 | input_ids: torch.Tensor | None, |
| 361 | positions: torch.Tensor, |
| 362 | intermediate_tensors: IntermediateTensors | None = None, |
| 363 | inputs_embeds: torch.Tensor | None = None, |
| 364 | **kwargs: object, |
| 365 | ) -> torch.Tensor | IntermediateTensors: |
| 366 | if get_pp_group().is_first_rank: |
| 367 | if inputs_embeds is not None: |
| 368 | hidden_states = inputs_embeds |
| 369 | else: |
| 370 | hidden_states = self.embed_input_ids(input_ids) |
| 371 | else: |
| 372 | assert intermediate_tensors is not None |
| 373 | hidden_states = intermediate_tensors["hidden_states"] |
| 374 | |
| 375 | # Run encoder. |
| 376 | hidden_states = self.encoder( |
| 377 | hidden_states=hidden_states, |
| 378 | position_ids=positions, |
| 379 | ) |
| 380 | |
| 381 | return hidden_states |
| 382 | |
| 383 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 384 | loader = AutoWeightsLoader(self) |
| 385 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 386 | |
| 387 | |
| 388 | class ChatGLMBaseModel(nn.Module): |
| 389 | def __init__( |
| 390 | self, |
| 391 | *, |
| 392 | vllm_config: VllmConfig, |
| 393 | prefix: str = "", |
| 394 | transformer_type: type[ChatGLMModel] = ChatGLMModel, |
| 395 | ) -> None: |
| 396 | super().__init__() |
| 397 | config = vllm_config.model_config.hf_config |
| 398 | quant_config = vllm_config.quant_config |
| 399 | |
| 400 | multimodal_config = vllm_config.model_config.multimodal_config |
| 401 | self.config = config |
| 402 | self.multimodal_config = multimodal_config |
| 403 | |
| 404 | self.quant_config = quant_config |
| 405 | self.max_position_embeddings = getattr(config, "max_sequence_length", 8192) |
| 406 | self.transformer = transformer_type( |
| 407 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer") |
| 408 | ) |
| 409 | if self.config.tie_word_embeddings: |
| 410 | self.transformer.output_layer.weight = self.transformer.embedding.weight |
| 411 | self.lm_head = self.transformer.output_layer |
| 412 | self.logits_processor = LogitsProcessor(config.padded_vocab_size) |
| 413 | self.make_empty_intermediate_tensors = ( |
| 414 | self.transformer.make_empty_intermediate_tensors |
| 415 | ) |
| 416 | |
| 417 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 418 | return self.transformer.embed_input_ids(input_ids) |
| 419 | |
| 420 | def compute_logits( |
| 421 | self, |
| 422 | hidden_states: torch.Tensor, |
| 423 | ) -> torch.Tensor | None: |
| 424 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 425 | return logits |
| 426 | |
| 427 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 428 | loader = AutoWeightsLoader(self) |
| 429 | return loader.load_weights(weights) |
| 430 | |
| 431 | |
| 432 | class ChatGLMForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP, SupportsQuant): |
| 433 | packed_modules_mapping: dict[str, list[str]] = { |
| 434 | "query_key_value": ["query_key_value"], |
| 435 | "dense_h_to_4h": ["dense_h_to_4h"], |
| 436 | } |
| 437 | |
| 438 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 439 | config = vllm_config.model_config.hf_config |
| 440 | if hasattr(config, "vision_config"): |
| 441 | hf_overrides = {"architectures": ["GLM4VForCausalLM"]} |
| 442 | raise RuntimeError( |
| 443 | "The configuration of this model indicates that it supports " |
| 444 | "vision inputs, but you instantiated the text-only version " |
| 445 | "of this model. Please use the vision model by setting " |
| 446 | f"`--hf-overrides '{json.dumps(hf_overrides)}'`" |
| 447 | ) |
| 448 | |
| 449 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 450 | |
| 451 | def forward( |
| 452 | self, |
| 453 | input_ids: torch.Tensor | None, |
| 454 | positions: torch.Tensor, |
| 455 | intermediate_tensors: IntermediateTensors | None = None, |
| 456 | inputs_embeds: torch.Tensor | None = None, |
| 457 | ) -> torch.Tensor | IntermediateTensors: |
| 458 | hidden_states = self.transformer( |
| 459 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 460 | ) |
| 461 | return hidden_states |