| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only Jamba model.""" |
| 4 | |
| 5 | from collections.abc import Iterable |
| 6 | from itertools import islice |
| 7 | |
| 8 | import torch |
| 9 | from torch import nn |
| 10 | from transformers import JambaConfig |
| 11 | |
| 12 | from vllm.compilation.decorators import support_torch_compile |
| 13 | from vllm.config import CacheConfig, ModelConfig, VllmConfig |
| 14 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 15 | from vllm.distributed.parallel_state import get_pp_group |
| 16 | from vllm.model_executor.layers.attention import Attention |
| 17 | from vllm.model_executor.layers.fused_moe import ( |
| 18 | FusedMoEFactory, |
| 19 | GateLinear, |
| 20 | ) |
| 21 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 22 | from vllm.model_executor.layers.linear import ( |
| 23 | QKVParallelLinear, |
| 24 | RowParallelLinear, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 27 | from vllm.model_executor.layers.mamba.mamba_mixer import MambaMixer |
| 28 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 29 | MambaStateCopyFunc, |
| 30 | MambaStateCopyFuncCalculator, |
| 31 | MambaStateDtypeCalculator, |
| 32 | MambaStateShapeCalculator, |
| 33 | ) |
| 34 | from vllm.model_executor.layers.pooler import DispatchPooler |
| 35 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 36 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 37 | ParallelLMHead, |
| 38 | VocabParallelEmbedding, |
| 39 | ) |
| 40 | from vllm.model_executor.models.llama import LlamaMLP as JambaMLP |
| 41 | from vllm.sequence import IntermediateTensors |
| 42 | |
| 43 | from .interfaces import ( |
| 44 | HasInnerState, |
| 45 | IsHybrid, |
| 46 | SupportsLoRA, |
| 47 | SupportsMambaPrefixCaching, |
| 48 | SupportsPP, |
| 49 | ) |
| 50 | from .utils import ( |
| 51 | AutoWeightsLoader, |
| 52 | WeightsMapper, |
| 53 | make_empty_intermediate_tensors_factory, |
| 54 | make_layers, |
| 55 | maybe_prefix, |
| 56 | ) |
| 57 | |
| 58 | |
| 59 | class JambaMoE(nn.Module): |
| 60 | def __init__( |
| 61 | self, |
| 62 | config: JambaConfig, |
| 63 | num_experts: int | None = None, |
| 64 | top_k: int | None = None, |
| 65 | params_dtype: torch.dtype | None = None, |
| 66 | tp_size: int | None = None, |
| 67 | quant_config: QuantizationConfig | None = None, |
| 68 | prefix: str = "", |
| 69 | ): |
| 70 | super().__init__() |
| 71 | self.num_total_experts = num_experts or config.num_experts |
| 72 | self.top_k = top_k or config.num_experts_per_tok |
| 73 | self.hidden_size = config.hidden_size |
| 74 | self.intermediate_size = config.intermediate_size |
| 75 | |
| 76 | if self.num_total_experts > 1: |
| 77 | self.router = GateLinear( |
| 78 | self.hidden_size, |
| 79 | self.num_total_experts, |
| 80 | params_dtype=params_dtype, |
| 81 | prefix=f"{prefix}.router", |
| 82 | ) |
| 83 | |
| 84 | self.experts = FusedMoEFactory( |
| 85 | self.num_total_experts, |
| 86 | self.top_k, |
| 87 | self.hidden_size, |
| 88 | self.intermediate_size, |
| 89 | tp_size=tp_size, |
| 90 | params_dtype=params_dtype, |
| 91 | renormalize=False, |
| 92 | use_grouped_topk=False, |
| 93 | quant_config=quant_config, |
| 94 | prefix=f"{prefix}.experts", |
| 95 | ) |
| 96 | |
| 97 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 98 | orig_shape = hidden_states.shape |
| 99 | hidden_states = hidden_states.view(-1, self.hidden_size) |
| 100 | # router_logits: (batch * sequence_length, n_experts) |
| 101 | if self.num_total_experts > 1: |
| 102 | router_logits, _ = self.router(hidden_states) |
| 103 | else: |
| 104 | router_logits = torch.ones( |
| 105 | (hidden_states.shape[0], 1), |
| 106 | device=hidden_states.device, |
| 107 | dtype=hidden_states.dtype, |
| 108 | ) |
| 109 | hidden_states = self.experts(hidden_states, router_logits) |
| 110 | return hidden_states.view(orig_shape) |
| 111 | |
| 112 | |
| 113 | class JambaMambaDecoderLayer(nn.Module): |
| 114 | def __init__( |
| 115 | self, |
| 116 | config: JambaConfig, |
| 117 | layer_idx: int, |
| 118 | model_config: ModelConfig | None = None, |
| 119 | cache_config: CacheConfig | None = None, |
| 120 | quant_config: QuantizationConfig | None = None, |
| 121 | is_lora_enabled: bool = False, |
| 122 | prefix: str = "", |
| 123 | **kwargs, |
| 124 | ) -> None: |
| 125 | super().__init__() |
| 126 | self.config = config |
| 127 | self.is_lora_enabled = is_lora_enabled |
| 128 | self.mamba = MambaMixer( |
| 129 | hidden_size=config.hidden_size, |
| 130 | ssm_state_size=config.mamba_d_state, |
| 131 | conv_kernel_size=config.mamba_d_conv, |
| 132 | intermediate_size=config.mamba_expand * config.hidden_size, |
| 133 | time_step_rank=config.mamba_dt_rank, |
| 134 | use_conv_bias=config.mamba_conv_bias, |
| 135 | use_bias=config.mamba_proj_bias, |
| 136 | use_rms_norm=True, |
| 137 | rms_norm_eps=config.rms_norm_eps, |
| 138 | activation=config.hidden_act, |
| 139 | is_lora_enabled=self.is_lora_enabled, |
| 140 | model_config=model_config, |
| 141 | cache_config=cache_config, |
| 142 | prefix=f"{prefix}.mixer", |
| 143 | ) |
| 144 | |
| 145 | num_experts = config.layers_num_experts[layer_idx] |
| 146 | if num_experts > 1: |
| 147 | self.feed_forward = JambaMoE( |
| 148 | config, |
| 149 | quant_config=quant_config, |
| 150 | prefix=f"{prefix}.feed_forward", |
| 151 | ) |
| 152 | else: |
| 153 | self.feed_forward = JambaMLP( |
| 154 | config.hidden_size, |
| 155 | config.intermediate_size, |
| 156 | config.hidden_act, |
| 157 | quant_config=quant_config, |
| 158 | prefix=f"{prefix}.feed_forward", |
| 159 | ) |
| 160 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 161 | self.pre_ff_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 162 | |
| 163 | def forward( |
| 164 | self, |
| 165 | hidden_states: torch.Tensor, |
| 166 | residual: torch.Tensor | None, |
| 167 | **kwargs, |
| 168 | ): |
| 169 | if residual is None: |
| 170 | residual = hidden_states |
| 171 | hidden_states = self.input_layernorm(hidden_states) |
| 172 | else: |
| 173 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 174 | |
| 175 | output = torch.empty_like(hidden_states) |
| 176 | self.mamba(hidden_states, output) |
| 177 | # Fully Connected |
| 178 | hidden_states, residual = self.pre_ff_layernorm(output, residual) |
| 179 | hidden_states = self.feed_forward(hidden_states) |
| 180 | return hidden_states, residual |
| 181 | |
| 182 | |
| 183 | class JambaAttentionDecoderLayer(nn.Module): |
| 184 | def __init__( |
| 185 | self, |
| 186 | config: JambaConfig, |
| 187 | layer_idx: int, |
| 188 | model_config: ModelConfig | None = None, |
| 189 | cache_config: CacheConfig | None = None, |
| 190 | quant_config: QuantizationConfig | None = None, |
| 191 | prefix: str = "", |
| 192 | **kwargs, |
| 193 | ) -> None: |
| 194 | super().__init__() |
| 195 | self.hidden_size = config.hidden_size |
| 196 | tp_size = get_tensor_model_parallel_world_size() |
| 197 | self.total_num_heads = config.num_attention_heads |
| 198 | assert self.total_num_heads % tp_size == 0 |
| 199 | self.num_heads = self.total_num_heads // tp_size |
| 200 | self.total_num_kv_heads = config.num_key_value_heads |
| 201 | if self.total_num_kv_heads >= tp_size: |
| 202 | # Number of KV heads is greater than TP size, so we partition |
| 203 | # the KV heads across multiple tensor parallel GPUs. |
| 204 | assert self.total_num_kv_heads % tp_size == 0 |
| 205 | else: |
| 206 | # Number of KV heads is less than TP size, so we replicate |
| 207 | # the KV heads across multiple tensor parallel GPUs. |
| 208 | assert tp_size % self.total_num_kv_heads == 0 |
| 209 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 210 | self.head_dim = config.hidden_size // self.total_num_heads |
| 211 | self.q_size = self.num_heads * self.head_dim |
| 212 | self.kv_size = self.num_kv_heads * self.head_dim |
| 213 | self.scaling = self.head_dim**-0.5 |
| 214 | |
| 215 | self.qkv_proj = QKVParallelLinear( |
| 216 | config.hidden_size, |
| 217 | self.head_dim, |
| 218 | self.total_num_heads, |
| 219 | self.total_num_kv_heads, |
| 220 | bias=False, |
| 221 | quant_config=quant_config, |
| 222 | prefix=f"{prefix}.qkv_proj", |
| 223 | ) |
| 224 | self.o_proj = RowParallelLinear( |
| 225 | self.total_num_heads * self.head_dim, |
| 226 | config.hidden_size, |
| 227 | bias=False, |
| 228 | quant_config=quant_config, |
| 229 | prefix=f"{prefix}.o_proj", |
| 230 | ) |
| 231 | |
| 232 | self.attn = Attention( |
| 233 | self.num_heads, |
| 234 | self.head_dim, |
| 235 | self.scaling, |
| 236 | num_kv_heads=self.num_kv_heads, |
| 237 | cache_config=cache_config, |
| 238 | prefix=f"{prefix}.attn", |
| 239 | ) |
| 240 | |
| 241 | num_experts = config.layers_num_experts[layer_idx] |
| 242 | if num_experts > 1: |
| 243 | self.feed_forward = JambaMoE( |
| 244 | config, |
| 245 | quant_config=quant_config, |
| 246 | prefix=f"{prefix}.feed_forward", |
| 247 | ) |
| 248 | else: |
| 249 | self.feed_forward = JambaMLP( |
| 250 | config.hidden_size, |
| 251 | config.intermediate_size, |
| 252 | config.hidden_act, |
| 253 | quant_config=quant_config, |
| 254 | prefix=f"{prefix}.feed_forward", |
| 255 | ) |
| 256 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 257 | self.pre_ff_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 258 | |
| 259 | def self_attention( |
| 260 | self, |
| 261 | positions: torch.Tensor, |
| 262 | hidden_states: torch.Tensor, |
| 263 | **kwargs, |
| 264 | ) -> torch.Tensor: |
| 265 | qkv, _ = self.qkv_proj(hidden_states) |
| 266 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 267 | attn_output = self.attn(q, k, v) |
| 268 | output, _ = self.o_proj(attn_output) |
| 269 | return output |
| 270 | |
| 271 | def forward( |
| 272 | self, |
| 273 | positions: torch.Tensor, |
| 274 | hidden_states: torch.Tensor, |
| 275 | residual: torch.Tensor | None, |
| 276 | **kwargs, |
| 277 | ): |
| 278 | if residual is None: |
| 279 | residual = hidden_states |
| 280 | hidden_states = self.input_layernorm(hidden_states) |
| 281 | else: |
| 282 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 283 | |
| 284 | hidden_states = self.self_attention( |
| 285 | positions=positions, |
| 286 | hidden_states=hidden_states, |
| 287 | ) |
| 288 | # Fully Connected |
| 289 | hidden_states, residual = self.pre_ff_layernorm(hidden_states, residual) |
| 290 | hidden_states = self.feed_forward(hidden_states) |
| 291 | return hidden_states, residual |
| 292 | |
| 293 | |
| 294 | ALL_DECODER_LAYER_TYPES = { |
| 295 | "attention": JambaAttentionDecoderLayer, |
| 296 | "mamba": JambaMambaDecoderLayer, |
| 297 | } |
| 298 | |
| 299 | |
| 300 | @support_torch_compile |
| 301 | class JambaModel(nn.Module): |
| 302 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 303 | super().__init__() |
| 304 | |
| 305 | config = vllm_config.model_config.hf_config |
| 306 | model_config = vllm_config.model_config |
| 307 | cache_config = vllm_config.cache_config |
| 308 | quant_config = vllm_config.quant_config |
| 309 | |
| 310 | self.config = config |
| 311 | |
| 312 | self.vocab_size = config.vocab_size |
| 313 | |
| 314 | self.embed_tokens = VocabParallelEmbedding( |
| 315 | self.vocab_size, |
| 316 | config.hidden_size, |
| 317 | ) |
| 318 | |
| 319 | extra_kwargs = {"is_lora_enabled": bool(vllm_config.lora_config)} |
| 320 | |
| 321 | def get_layer(prefix: str): |
| 322 | layer_idx = int(prefix.rsplit(".", 1)[1]) |
| 323 | layer_class = ALL_DECODER_LAYER_TYPES[config.layers_block_type[layer_idx]] |
| 324 | return layer_class( |
| 325 | config, |
| 326 | layer_idx, |
| 327 | model_config, |
| 328 | cache_config, |
| 329 | quant_config=quant_config, |
| 330 | prefix=prefix, |
| 331 | **extra_kwargs, |
| 332 | ) |
| 333 | |
| 334 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 335 | config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers" |
| 336 | ) |
| 337 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 338 | ["hidden_states", "residual"], config.hidden_size |
| 339 | ) |
| 340 | |
| 341 | self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 342 | |
| 343 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 344 | return self.embed_tokens(input_ids) |
| 345 | |
| 346 | def forward( |
| 347 | self, |
| 348 | input_ids: torch.Tensor | None, |
| 349 | positions: torch.Tensor, |
| 350 | intermediate_tensors: IntermediateTensors | None = None, |
| 351 | inputs_embeds: torch.Tensor | None = None, |
| 352 | ) -> torch.Tensor: |
| 353 | if get_pp_group().is_first_rank: |
| 354 | if inputs_embeds is not None: |
| 355 | hidden_states = inputs_embeds |
| 356 | else: |
| 357 | hidden_states = self.embed_input_ids(input_ids) |
| 358 | residual = None |
| 359 | else: |
| 360 | assert intermediate_tensors is not None |
| 361 | hidden_states = intermediate_tensors["hidden_states"] |
| 362 | residual = intermediate_tensors["residual"] |
| 363 | |
| 364 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 365 | hidden_states, residual = layer( |
| 366 | positions=positions, hidden_states=hidden_states, residual=residual |
| 367 | ) |
| 368 | |
| 369 | if not get_pp_group().is_last_rank: |
| 370 | return IntermediateTensors( |
| 371 | {"hidden_states": hidden_states, "residual": residual} |
| 372 | ) |
| 373 | hidden_states, _ = self.final_layernorm(hidden_states, residual) |
| 374 | return hidden_states |
| 375 | |
| 376 | |
| 377 | class JambaForCausalLM( |
| 378 | nn.Module, |
| 379 | HasInnerState, |
| 380 | SupportsLoRA, |
| 381 | SupportsPP, |
| 382 | IsHybrid, |
| 383 | SupportsMambaPrefixCaching, |
| 384 | ): |
| 385 | hf_to_vllm_mapper = WeightsMapper( |
| 386 | orig_to_new_substr={".self_attn.": ".", ".A_log": ".A"}, |
| 387 | orig_to_new_stacked={ |
| 388 | ".q_proj": (".qkv_proj", "q"), |
| 389 | ".k_proj": (".qkv_proj", "k"), |
| 390 | ".v_proj": (".qkv_proj", "v"), |
| 391 | # .experts.gate_up_proj must be handled by MoERunner.load_weights for EP |
| 392 | ".feed_forward.gate_proj": (".feed_forward.gate_up_proj", 0), |
| 393 | ".feed_forward.up_proj": (".feed_forward.gate_up_proj", 1), |
| 394 | }, |
| 395 | ) |
| 396 | packed_modules_mapping = { |
| 397 | "qkv_proj": [ |
| 398 | "q_proj", |
| 399 | "k_proj", |
| 400 | "v_proj", |
| 401 | ], |
| 402 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 403 | "in_proj": ["in_proj"], |
| 404 | } |
| 405 | |
| 406 | # LoRA specific attributes |
| 407 | embedding_modules = { |
| 408 | "embed_tokens": "input_embeddings", |
| 409 | "lm_head": "output_embeddings", |
| 410 | } |
| 411 | |
| 412 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 413 | config = vllm_config.model_config.hf_config |
| 414 | |
| 415 | scheduler_config = vllm_config.scheduler_config |
| 416 | |
| 417 | super().__init__() |
| 418 | self.config = config |
| 419 | self.vllm_config = vllm_config |
| 420 | self.model_config = vllm_config.model_config |
| 421 | self.scheduler_config = scheduler_config |
| 422 | self.model = JambaModel( |
| 423 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 424 | ) |
| 425 | |
| 426 | self.lm_head = ParallelLMHead( |
| 427 | config.vocab_size, |
| 428 | config.hidden_size, |
| 429 | prefix=maybe_prefix(prefix, "lm_head"), |
| 430 | ) |
| 431 | |
| 432 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 433 | |
| 434 | self.make_empty_intermediate_tensors = ( |
| 435 | self.model.make_empty_intermediate_tensors |
| 436 | ) |
| 437 | |
| 438 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 439 | return self.model.embed_input_ids(input_ids) |
| 440 | |
| 441 | def forward( |
| 442 | self, |
| 443 | input_ids: torch.Tensor | None, |
| 444 | positions: torch.Tensor, |
| 445 | intermediate_tensors: IntermediateTensors | None = None, |
| 446 | inputs_embeds: torch.Tensor | None = None, |
| 447 | **kwargs, |
| 448 | ): |
| 449 | hidden_states = self.model( |
| 450 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 451 | ) |
| 452 | return hidden_states |
| 453 | |
| 454 | def copy_inputs_before_cuda_graphs(self, input_buffers, **kwargs): |
| 455 | return self.mamba_cache.copy_inputs_before_cuda_graphs(input_buffers, **kwargs) |
| 456 | |
| 457 | def get_seqlen_agnostic_capture_inputs(self, batch_size: int): |
| 458 | return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size) |
| 459 | |
| 460 | @classmethod |
| 461 | def get_mamba_state_dtype_from_config( |
| 462 | cls, |
| 463 | vllm_config: "VllmConfig", |
| 464 | ) -> tuple[torch.dtype, torch.dtype]: |
| 465 | return MambaStateDtypeCalculator.mamba1_state_dtype( |
| 466 | vllm_config.model_config.dtype, |
| 467 | vllm_config.cache_config.mamba_cache_dtype, |
| 468 | vllm_config.cache_config.mamba_ssm_cache_dtype, |
| 469 | ) |
| 470 | |
| 471 | @classmethod |
| 472 | def get_mamba_state_shape_from_config( |
| 473 | cls, |
| 474 | vllm_config: "VllmConfig", |
| 475 | ) -> tuple[tuple[int, int], tuple[int, int]]: |
| 476 | parallel_config = vllm_config.parallel_config |
| 477 | hf_config = vllm_config.model_config.hf_config |
| 478 | hidden_size = hf_config.hidden_size |
| 479 | |
| 480 | return MambaStateShapeCalculator.mamba1_state_shape( |
| 481 | tp_world_size=parallel_config.tensor_parallel_size, |
| 482 | intermediate_size=hf_config.mamba_expand * hidden_size, |
| 483 | state_size=hf_config.mamba_d_state, |
| 484 | conv_kernel=hf_config.mamba_d_conv, |
| 485 | ) |
| 486 | |
| 487 | @classmethod |
| 488 | def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]: |
| 489 | return MambaStateCopyFuncCalculator.mamba1_state_copy_func() |
| 490 | |
| 491 | def compute_logits( |
| 492 | self, |
| 493 | hidden_states: torch.Tensor, |
| 494 | ) -> torch.Tensor | None: |
| 495 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 496 | return logits |
| 497 | |
| 498 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 499 | loader = AutoWeightsLoader(self) |
| 500 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 501 | |
| 502 | |
| 503 | class JambaForSequenceClassification(JambaForCausalLM): |
| 504 | is_pooling_model = True |
| 505 | |
| 506 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 507 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 508 | |
| 509 | config = vllm_config.model_config.hf_config |
| 510 | num_labels: int = config.num_labels |
| 511 | score_bias: bool = getattr(config, "score_bias", False) |
| 512 | |
| 513 | # TODO: The original reward weights have float32 accuracy data, we |
| 514 | # would like to load them in fp32 to get that extra precision. |
| 515 | # Currently weight_loader passes the weight which is already in bf16 |
| 516 | self.score = nn.Linear( |
| 517 | config.hidden_size, |
| 518 | num_labels, |
| 519 | bias=score_bias, |
| 520 | dtype=vllm_config.model_config.head_dtype, |
| 521 | ) |
| 522 | |
| 523 | pooler_config = vllm_config.model_config.pooler_config |
| 524 | assert pooler_config is not None |
| 525 | |
| 526 | self.pooler = DispatchPooler.for_seq_cls(pooler_config, classifier=self.score) |