| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://huggingface.co/OrionStarAI/Orion-14B-Base/blob/main/modeling_orion.py |
| 6 | # Copyright (c) OrionStar Inc. |
| 7 | # LICENSE: https://huggingface.co/OrionStarAI/Orion-14B-Base/blob/main/LICENSE |
| 8 | """Inference-only Orion-14B model compatible with HuggingFace weights.""" |
| 9 | |
| 10 | from collections.abc import Iterable |
| 11 | from itertools import islice |
| 12 | from typing import Any |
| 13 | |
| 14 | import torch |
| 15 | from torch import nn |
| 16 | from transformers import PreTrainedConfig |
| 17 | |
| 18 | from vllm.compilation.decorators import support_torch_compile |
| 19 | from vllm.config import CacheConfig, VllmConfig |
| 20 | from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size |
| 21 | from vllm.model_executor.layers.activation import SiluAndMul |
| 22 | from vllm.model_executor.layers.attention import Attention |
| 23 | from vllm.model_executor.layers.linear import ( |
| 24 | MergedColumnParallelLinear, |
| 25 | QKVParallelLinear, |
| 26 | RowParallelLinear, |
| 27 | ) |
| 28 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 29 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 30 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 31 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 32 | ParallelLMHead, |
| 33 | VocabParallelEmbedding, |
| 34 | ) |
| 35 | from vllm.sequence import IntermediateTensors |
| 36 | |
| 37 | from .interfaces import SupportsPP |
| 38 | from .utils import ( |
| 39 | AutoWeightsLoader, |
| 40 | WeightsMapper, |
| 41 | make_empty_intermediate_tensors_factory, |
| 42 | make_layers, |
| 43 | maybe_prefix, |
| 44 | ) |
| 45 | |
| 46 | |
| 47 | class OrionMLP(nn.Module): |
| 48 | def __init__( |
| 49 | self, |
| 50 | hidden_size: int, |
| 51 | intermediate_size: int, |
| 52 | hidden_act: str, |
| 53 | quant_config: QuantizationConfig | None = None, |
| 54 | prefix: str = "", |
| 55 | ) -> None: |
| 56 | super().__init__() |
| 57 | self.gate_up_proj = MergedColumnParallelLinear( |
| 58 | hidden_size, |
| 59 | [intermediate_size] * 2, |
| 60 | bias=False, |
| 61 | quant_config=quant_config, |
| 62 | prefix=f"{prefix}.gate_up_proj", |
| 63 | ) |
| 64 | self.down_proj = RowParallelLinear( |
| 65 | intermediate_size, |
| 66 | hidden_size, |
| 67 | bias=False, |
| 68 | quant_config=quant_config, |
| 69 | prefix=f"{prefix}.down_proj", |
| 70 | ) |
| 71 | if hidden_act != "silu": |
| 72 | raise ValueError( |
| 73 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 74 | ) |
| 75 | self.act_fn = SiluAndMul() |
| 76 | |
| 77 | def forward(self, x): |
| 78 | gate_up, _ = self.gate_up_proj(x) |
| 79 | x = self.act_fn(gate_up) |
| 80 | x, _ = self.down_proj(x) |
| 81 | return x |
| 82 | |
| 83 | |
| 84 | class OrionAttention(nn.Module): |
| 85 | def __init__( |
| 86 | self, |
| 87 | hidden_size: int, |
| 88 | num_heads: int, |
| 89 | num_kv_heads: int, |
| 90 | rope_parameters: dict[str, Any] | None = None, |
| 91 | max_position_embeddings: int = 8192, |
| 92 | cache_config: CacheConfig | None = None, |
| 93 | quant_config: QuantizationConfig | None = None, |
| 94 | prefix: str = "", |
| 95 | ) -> None: |
| 96 | super().__init__() |
| 97 | self.hidden_size = hidden_size |
| 98 | tp_size = get_tensor_model_parallel_world_size() |
| 99 | self.total_num_heads = num_heads |
| 100 | assert self.total_num_heads % tp_size == 0 |
| 101 | self.num_heads = self.total_num_heads // tp_size |
| 102 | self.total_num_kv_heads = num_kv_heads |
| 103 | if self.total_num_kv_heads >= tp_size: |
| 104 | # Number of KV heads is greater than TP size, so we partition |
| 105 | # the KV heads across multiple tensor parallel GPUs. |
| 106 | assert self.total_num_kv_heads % tp_size == 0 |
| 107 | else: |
| 108 | # Number of KV heads is less than TP size, so we replicate |
| 109 | # the KV heads across multiple tensor parallel GPUs. |
| 110 | assert tp_size % self.total_num_kv_heads == 0 |
| 111 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 112 | self.head_dim = hidden_size // self.total_num_heads |
| 113 | self.q_size = self.num_heads * self.head_dim |
| 114 | self.kv_size = self.num_kv_heads * self.head_dim |
| 115 | self.scaling = self.head_dim**-0.5 |
| 116 | self.max_position_embeddings = max_position_embeddings |
| 117 | |
| 118 | self.qkv_proj = QKVParallelLinear( |
| 119 | hidden_size, |
| 120 | self.head_dim, |
| 121 | self.total_num_heads, |
| 122 | self.total_num_kv_heads, |
| 123 | bias=False, |
| 124 | quant_config=quant_config, |
| 125 | prefix=f"{prefix}.qkv_proj", |
| 126 | ) |
| 127 | self.o_proj = RowParallelLinear( |
| 128 | self.total_num_heads * self.head_dim, |
| 129 | hidden_size, |
| 130 | bias=False, |
| 131 | quant_config=quant_config, |
| 132 | prefix=f"{prefix}.o_proj", |
| 133 | ) |
| 134 | |
| 135 | self.rotary_emb = get_rope( |
| 136 | self.head_dim, |
| 137 | max_position=max_position_embeddings, |
| 138 | rope_parameters=rope_parameters, |
| 139 | ) |
| 140 | self.attn = Attention( |
| 141 | self.num_heads, |
| 142 | self.head_dim, |
| 143 | self.scaling, |
| 144 | num_kv_heads=self.num_kv_heads, |
| 145 | cache_config=cache_config, |
| 146 | quant_config=quant_config, |
| 147 | prefix=f"{prefix}.attn", |
| 148 | ) |
| 149 | |
| 150 | def forward( |
| 151 | self, |
| 152 | positions: torch.Tensor, |
| 153 | hidden_states: torch.Tensor, |
| 154 | ) -> torch.Tensor: |
| 155 | qkv, _ = self.qkv_proj(hidden_states) |
| 156 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 157 | q, k = self.rotary_emb(positions, q, k) |
| 158 | attn_output = self.attn(q, k, v) |
| 159 | output, _ = self.o_proj(attn_output) |
| 160 | return output |
| 161 | |
| 162 | |
| 163 | class OrionDecoderLayer(nn.Module): |
| 164 | def __init__( |
| 165 | self, |
| 166 | config: PreTrainedConfig, |
| 167 | cache_config: CacheConfig | None = None, |
| 168 | quant_config: QuantizationConfig | None = None, |
| 169 | prefix: str = "", |
| 170 | ) -> None: |
| 171 | super().__init__() |
| 172 | self.hidden_size = config.hidden_size |
| 173 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 174 | self.self_attn = OrionAttention( |
| 175 | hidden_size=self.hidden_size, |
| 176 | num_heads=config.num_attention_heads, |
| 177 | num_kv_heads=config.num_key_value_heads, |
| 178 | rope_parameters=config.rope_parameters, |
| 179 | max_position_embeddings=max_position_embeddings, |
| 180 | cache_config=cache_config, |
| 181 | quant_config=quant_config, |
| 182 | prefix=f"{prefix}.self_attn", |
| 183 | ) |
| 184 | self.mlp = OrionMLP( |
| 185 | hidden_size=self.hidden_size, |
| 186 | intermediate_size=config.intermediate_size, |
| 187 | hidden_act=config.hidden_act, |
| 188 | quant_config=quant_config, |
| 189 | prefix=f"{prefix}.mlp", |
| 190 | ) |
| 191 | |
| 192 | self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 193 | self.post_attention_layernorm = nn.LayerNorm( |
| 194 | config.hidden_size, eps=config.rms_norm_eps |
| 195 | ) |
| 196 | |
| 197 | def forward( |
| 198 | self, |
| 199 | positions: torch.Tensor, |
| 200 | hidden_states: torch.Tensor, |
| 201 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 202 | # Self Attention |
| 203 | residual = hidden_states |
| 204 | hidden_states = self.input_layernorm(hidden_states) |
| 205 | hidden_states = self.self_attn( |
| 206 | positions=positions, |
| 207 | hidden_states=hidden_states, |
| 208 | ) |
| 209 | |
| 210 | hidden_states = residual + hidden_states |
| 211 | |
| 212 | # Fully Connected |
| 213 | residual = hidden_states |
| 214 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 215 | hidden_states = self.mlp(hidden_states) |
| 216 | hidden_states = residual + hidden_states |
| 217 | return hidden_states |
| 218 | |
| 219 | |
| 220 | @support_torch_compile |
| 221 | class OrionModel(nn.Module): |
| 222 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 223 | super().__init__() |
| 224 | |
| 225 | config = vllm_config.model_config.hf_config |
| 226 | cache_config = vllm_config.cache_config |
| 227 | quant_config = vllm_config.quant_config |
| 228 | |
| 229 | self.config = config |
| 230 | self.vocab_size = config.vocab_size |
| 231 | self.embed_tokens = VocabParallelEmbedding( |
| 232 | config.vocab_size, |
| 233 | config.hidden_size, |
| 234 | ) |
| 235 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 236 | config.num_hidden_layers, |
| 237 | lambda prefix: OrionDecoderLayer( |
| 238 | config, cache_config, quant_config, prefix=prefix |
| 239 | ), |
| 240 | prefix=f"{prefix}.layers", |
| 241 | ) |
| 242 | self.norm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 243 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 244 | [ |
| 245 | "hidden_states", |
| 246 | ], |
| 247 | config.hidden_size, |
| 248 | ) |
| 249 | |
| 250 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 251 | return self.embed_tokens(input_ids) |
| 252 | |
| 253 | def forward( |
| 254 | self, |
| 255 | input_ids: torch.Tensor | None, |
| 256 | positions: torch.Tensor, |
| 257 | intermediate_tensors: IntermediateTensors | None, |
| 258 | inputs_embeds: torch.Tensor | None = None, |
| 259 | ) -> torch.Tensor | IntermediateTensors: |
| 260 | if get_pp_group().is_first_rank: |
| 261 | if inputs_embeds is not None: |
| 262 | hidden_states = inputs_embeds |
| 263 | else: |
| 264 | hidden_states = self.embed_input_ids(input_ids) |
| 265 | else: |
| 266 | assert intermediate_tensors is not None |
| 267 | hidden_states = intermediate_tensors["hidden_states"] |
| 268 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 269 | hidden_states = layer(positions, hidden_states) |
| 270 | if not get_pp_group().is_last_rank: |
| 271 | return IntermediateTensors( |
| 272 | { |
| 273 | "hidden_states": hidden_states, |
| 274 | } |
| 275 | ) |
| 276 | hidden_states = self.norm(hidden_states) |
| 277 | return hidden_states |
| 278 | |
| 279 | |
| 280 | class OrionForCausalLM(nn.Module, SupportsPP): |
| 281 | hf_to_vllm_mapper = WeightsMapper( |
| 282 | orig_to_new_stacked={ |
| 283 | # weight_name: (param_name, shard_id) |
| 284 | ".q_proj": (".qkv_proj", "q"), |
| 285 | ".k_proj": (".qkv_proj", "k"), |
| 286 | ".v_proj": (".qkv_proj", "v"), |
| 287 | ".gate_proj": (".gate_up_proj", 0), |
| 288 | ".up_proj": (".gate_up_proj", 1), |
| 289 | } |
| 290 | ) |
| 291 | |
| 292 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 293 | super().__init__() |
| 294 | config = vllm_config.model_config.hf_config |
| 295 | quant_config = vllm_config.quant_config |
| 296 | self.config = config |
| 297 | self.quant_config = quant_config |
| 298 | self.model = OrionModel( |
| 299 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 300 | ) |
| 301 | self.lm_head = ParallelLMHead( |
| 302 | config.vocab_size, |
| 303 | config.hidden_size, |
| 304 | quant_config=quant_config, |
| 305 | prefix=maybe_prefix(prefix, "lm_head"), |
| 306 | ) |
| 307 | if self.config.tie_word_embeddings: |
| 308 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 309 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 310 | self.make_empty_intermediate_tensors = ( |
| 311 | self.model.make_empty_intermediate_tensors |
| 312 | ) |
| 313 | |
| 314 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 315 | return self.model.embed_input_ids(input_ids) |
| 316 | |
| 317 | def forward( |
| 318 | self, |
| 319 | input_ids: torch.Tensor | None, |
| 320 | positions: torch.Tensor, |
| 321 | intermediate_tensors: IntermediateTensors | None = None, |
| 322 | inputs_embeds: torch.Tensor | None = None, |
| 323 | ) -> torch.Tensor | IntermediateTensors: |
| 324 | hidden_states = self.model( |
| 325 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 326 | ) |
| 327 | return hidden_states |
| 328 | |
| 329 | def compute_logits( |
| 330 | self, |
| 331 | hidden_states: torch.Tensor, |
| 332 | ) -> torch.Tensor | None: |
| 333 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 334 | return logits |
| 335 | |
| 336 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 337 | loader = AutoWeightsLoader(self) |
| 338 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |