| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inkling model configs for the text backbone and audio/vision towers.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | from dataclasses import dataclass, field |
| 8 | from typing import Any, ClassVar, Literal, cast |
| 9 | |
| 10 | import torch |
| 11 | from transformers.configuration_utils import PreTrainedConfig |
| 12 | |
| 13 | |
| 14 | class InklingModelConfig(PreTrainedConfig): |
| 15 | model_type = "inkling_model" |
| 16 | keys_to_ignore_at_inference = ["past_key_values"] |
| 17 | |
| 18 | def __init__( |
| 19 | self, |
| 20 | *, |
| 21 | vocab_size: int = 201024, |
| 22 | hidden_size: int = 1536, |
| 23 | intermediate_size: int = 768, |
| 24 | dense_intermediate_size: int | None = None, |
| 25 | moe_intermediate_size: int | None = None, |
| 26 | num_hidden_layers: int = 16, |
| 27 | num_attention_heads: int = 12, |
| 28 | num_key_value_heads: int = 4, |
| 29 | head_dim: int | None = None, |
| 30 | v_head_dim: int | None = None, |
| 31 | d_rel: int = 16, |
| 32 | rel_extent: int = 1024, |
| 33 | local_layer_ids: list[int] | None = None, |
| 34 | sliding_window_size: int = 512, |
| 35 | swa_num_attention_heads: int | None = None, |
| 36 | swa_num_key_value_heads: int | None = None, |
| 37 | swa_head_dim: int | None = None, |
| 38 | swa_v_head_dim: int | None = None, |
| 39 | rms_norm_eps: float = 1e-6, |
| 40 | hidden_act: str = "silu", |
| 41 | q_bias: bool = False, |
| 42 | o_bias: bool = False, |
| 43 | use_embed_norm: bool = False, |
| 44 | use_sconv: bool = False, |
| 45 | sconv_kernel_size: int | None = None, |
| 46 | conv_kernel_size: int | None = None, |
| 47 | dense_mlp_idx: int = 0, |
| 48 | n_routed_experts: int = 0, |
| 49 | n_shared_experts: int = 0, |
| 50 | num_experts_per_tok: int = 1, |
| 51 | route_scale: float = 1.0, |
| 52 | use_gate_bias: bool = False, |
| 53 | use_global_scale: bool = False, |
| 54 | norm_after_topk: bool = True, |
| 55 | gate_activation: Literal["sigmoid", "softmax"] = "sigmoid", |
| 56 | shared_expert_sink: bool = False, |
| 57 | shared_experts_size: int = 1, |
| 58 | inference_moe_w13_interleaved: bool = True, |
| 59 | log_scaling_n_floor: int | None = None, |
| 60 | log_scaling_alpha: float = 0.1, |
| 61 | unpadded_vocab_size: int | None = None, |
| 62 | padded_vocab_size: int | None = None, |
| 63 | logits_mup_width_multiplier: float | None = None, |
| 64 | final_logit_softcapping: float | None = None, |
| 65 | tie_word_embeddings: bool = False, |
| 66 | num_nextn_predict_layers: int = 0, |
| 67 | chain_hidden_post_norm: bool = False, |
| 68 | **kwargs: Any, |
| 69 | ) -> None: |
| 70 | if head_dim is None: |
| 71 | head_dim = hidden_size // num_attention_heads |
| 72 | if v_head_dim is None: |
| 73 | v_head_dim = head_dim |
| 74 | if swa_num_attention_heads is None: |
| 75 | swa_num_attention_heads = num_attention_heads |
| 76 | if swa_num_key_value_heads is None: |
| 77 | swa_num_key_value_heads = num_key_value_heads |
| 78 | if swa_head_dim is None: |
| 79 | swa_head_dim = head_dim |
| 80 | if swa_v_head_dim is None: |
| 81 | swa_v_head_dim = swa_head_dim |
| 82 | # OG format: https://huggingface.co/thinkingmachines/Inkling |
| 83 | # HF format: https://github.com/huggingface/transformers/blob/fe747d88a3296bd94d426db2717f232f9d4afdb7/src/transformers/models/inkling/configuration_inkling.py#L95 |
| 84 | # Huggingface Transformers compatibility: |
| 85 | # HF intermediate_size == OG dense_intermediate_size |
| 86 | # HF moe_intermediate_size == OG intermediate_size |
| 87 | # HF conv_kernel_size == OG sconv_kernel_size |
| 88 | if ( |
| 89 | moe_intermediate_size is not None |
| 90 | and intermediate_size is not None |
| 91 | and dense_intermediate_size is None |
| 92 | ): |
| 93 | dense_intermediate_size = intermediate_size |
| 94 | intermediate_size = moe_intermediate_size |
| 95 | if dense_intermediate_size is None: |
| 96 | dense_intermediate_size = intermediate_size |
| 97 | # HF conv_kernel_size == OG sconv_kernel_size |
| 98 | if sconv_kernel_size is None: |
| 99 | sconv_kernel_size = conv_kernel_size if conv_kernel_size is not None else 4 |
| 100 | if local_layer_ids is None: |
| 101 | local_layer_ids = [] |
| 102 | |
| 103 | if padded_vocab_size is None: |
| 104 | padded_vocab_size = vocab_size |
| 105 | vocab_size = ( |
| 106 | unpadded_vocab_size |
| 107 | if ( |
| 108 | unpadded_vocab_size is not None |
| 109 | and unpadded_vocab_size < padded_vocab_size |
| 110 | ) |
| 111 | else vocab_size |
| 112 | ) |
| 113 | |
| 114 | self.vocab_size = vocab_size |
| 115 | self.padded_vocab_size = padded_vocab_size |
| 116 | self.hidden_size = hidden_size |
| 117 | self.intermediate_size = intermediate_size |
| 118 | self.dense_intermediate_size = dense_intermediate_size |
| 119 | self.num_hidden_layers = num_hidden_layers |
| 120 | self.num_attention_heads = num_attention_heads |
| 121 | self.num_key_value_heads = num_key_value_heads |
| 122 | self.head_dim = head_dim |
| 123 | self.v_head_dim = v_head_dim |
| 124 | self.d_rel = d_rel |
| 125 | self.rel_extent = rel_extent |
| 126 | self.local_layer_ids = local_layer_ids |
| 127 | self.sliding_window_size = sliding_window_size |
| 128 | self.swa_num_attention_heads = swa_num_attention_heads |
| 129 | self.swa_num_key_value_heads = swa_num_key_value_heads |
| 130 | self.swa_head_dim = swa_head_dim |
| 131 | self.swa_v_head_dim = swa_v_head_dim |
| 132 | self.rms_norm_eps = rms_norm_eps |
| 133 | self.hidden_act = hidden_act |
| 134 | self.q_bias = q_bias |
| 135 | self.o_bias = o_bias |
| 136 | self.use_embed_norm = use_embed_norm |
| 137 | self.use_sconv = use_sconv |
| 138 | self.sconv_kernel_size = sconv_kernel_size |
| 139 | self.dense_mlp_idx = dense_mlp_idx |
| 140 | self.n_routed_experts = n_routed_experts |
| 141 | self.num_experts = n_routed_experts |
| 142 | self.n_shared_experts = n_shared_experts |
| 143 | self.num_shared_experts = n_shared_experts |
| 144 | self.num_experts_per_tok = num_experts_per_tok |
| 145 | self.route_scale = route_scale |
| 146 | self.use_gate_bias = use_gate_bias |
| 147 | self.use_global_scale = use_global_scale |
| 148 | self.norm_after_topk = norm_after_topk |
| 149 | self.gate_activation = gate_activation |
| 150 | self.shared_expert_sink = shared_expert_sink |
| 151 | self.shared_experts_size = shared_experts_size |
| 152 | self.inference_moe_w13_interleaved = inference_moe_w13_interleaved |
| 153 | self.log_scaling_n_floor = log_scaling_n_floor |
| 154 | self.log_scaling_alpha = log_scaling_alpha |
| 155 | self.unpadded_vocab_size = self.vocab_size |
| 156 | self.logits_mup_width_multiplier = logits_mup_width_multiplier |
| 157 | self.final_logit_softcapping = final_logit_softcapping |
| 158 | # MTP (multi-token prediction) draft head: number of depth layers in the |
| 159 | # checkpoint (0 if absent); chain_norm applied after each depth. |
| 160 | self.num_nextn_predict_layers = num_nextn_predict_layers |
| 161 | self.chain_hidden_post_norm = chain_hidden_post_norm |
| 162 | |
| 163 | super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs) |
| 164 | |
| 165 | @property |
| 166 | def conv_layer_ids(self) -> list[int]: |
| 167 | return list(range(self.num_hidden_layers)) |
| 168 | |
| 169 | @property |
| 170 | def linear_layer_ids(self) -> list[int]: |
| 171 | return self.conv_layer_ids |
| 172 | |
| 173 | @property |
| 174 | def full_attention_layer_ids(self) -> list[int]: |
| 175 | return list(range(self.num_hidden_layers)) |
| 176 | |
| 177 | @property |
| 178 | def mamba_chunk_size(self) -> int: |
| 179 | # Floor at 64: mamba_cache_chunk_size = max(mamba_chunk_size, page_size), |
| 180 | # and a floor of 1 lets the radix tree adopt another request's KV at |
| 181 | # tiny shared prefixes, whose different kernel-rounding perturbs decode |
| 182 | # logits. |
| 183 | return 64 |
| 184 | |
| 185 | @property |
| 186 | def mamba2_cache_params(self) -> TMLConvCacheParams | None: |
| 187 | try: |
| 188 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 189 | |
| 190 | tp_size = get_tensor_model_parallel_world_size() |
| 191 | except (AssertionError, RuntimeError): |
| 192 | tp_size = 1 |
| 193 | |
| 194 | def tp_local_kv_conv_dim(num_kv_heads: int, head_dim: int) -> int: |
| 195 | return max(1, num_kv_heads // tp_size) * head_dim |
| 196 | |
| 197 | full_kv_conv_dim = tp_local_kv_conv_dim(self.num_key_value_heads, self.head_dim) |
| 198 | local_kv_conv_dim = tp_local_kv_conv_dim( |
| 199 | self.swa_num_key_value_heads, self.swa_head_dim |
| 200 | ) |
| 201 | stream_dim = self.hidden_size |
| 202 | conv_len = self.sconv_kernel_size - 1 |
| 203 | shape = TMLConvStateShape( |
| 204 | conv=[ |
| 205 | (conv_len, full_kv_conv_dim), |
| 206 | (conv_len, full_kv_conv_dim), |
| 207 | (conv_len, local_kv_conv_dim), |
| 208 | (conv_len, local_kv_conv_dim), |
| 209 | (conv_len, stream_dim), |
| 210 | (conv_len, stream_dim), |
| 211 | ], |
| 212 | temporal=(0, 0, 0), |
| 213 | ) |
| 214 | dtype = TMLStateDType(conv=torch.bfloat16, temporal=torch.bfloat16) |
| 215 | return TMLConvCacheParams(shape=shape, layers=self.conv_layer_ids, dtype=dtype) |
| 216 | |
| 217 | |
| 218 | class InklingAudioConfig(PreTrainedConfig): |
| 219 | model_type = "inkling_audio_model" |
| 220 | |
| 221 | def __init__( |
| 222 | self, |
| 223 | *, |
| 224 | decoder_dmodel: int | None = None, |
| 225 | n_mel_bins: int | None = None, |
| 226 | mel_vocab_size: int | None = None, |
| 227 | dmel_min_value: float | None = None, |
| 228 | dmel_max_value: float | None = None, |
| 229 | use_audio_norm: bool | None = None, |
| 230 | audio_mode: Literal["dmel", "flow"] | None = None, |
| 231 | **kwargs: Any, |
| 232 | ) -> None: |
| 233 | values = { |
| 234 | "n_mel_bins": n_mel_bins, |
| 235 | "mel_vocab_size": mel_vocab_size, |
| 236 | "dmel_min_value": dmel_min_value, |
| 237 | "dmel_max_value": dmel_max_value, |
| 238 | "use_audio_norm": use_audio_norm, |
| 239 | "audio_mode": audio_mode, |
| 240 | } |
| 241 | if decoder_dmodel is not None and ( |
| 242 | missing := [name for name, value in values.items() if value is None] |
| 243 | ): |
| 244 | raise ValueError( |
| 245 | "Enabled Inkling audio tower is missing config fields: " |
| 246 | + ", ".join(missing) |
| 247 | ) |
| 248 | self.decoder_dmodel = decoder_dmodel |
| 249 | self.n_mel_bins = cast(int, n_mel_bins) |
| 250 | self.mel_vocab_size = cast(int, mel_vocab_size) |
| 251 | self.dmel_min_value = cast(float, dmel_min_value) |
| 252 | self.dmel_max_value = cast(float, dmel_max_value) |
| 253 | self.use_audio_norm = cast(bool, use_audio_norm) |
| 254 | self.audio_mode = cast(Literal["dmel", "flow"], audio_mode) |
| 255 | super().__init__(**kwargs) |
| 256 | |
| 257 | |
| 258 | class InklingVisionConfig(PreTrainedConfig): |
| 259 | model_type = "inkling_vision_model" |
| 260 | |
| 261 | def __init__( |
| 262 | self, |
| 263 | *, |
| 264 | vision_encoder_type: Literal["linear", "hmlp"] | None = None, |
| 265 | decoder_dmodel: int | None = None, |
| 266 | patch_size: int | None = None, |
| 267 | temporal_patch_size: int | None = None, |
| 268 | n_channels: int | None = None, |
| 269 | n_layers: int | None = None, |
| 270 | use_vision_norm: bool | None = None, |
| 271 | **kwargs: Any, |
| 272 | ) -> None: |
| 273 | values = { |
| 274 | "vision_encoder_type": vision_encoder_type, |
| 275 | "patch_size": patch_size, |
| 276 | "temporal_patch_size": temporal_patch_size, |
| 277 | "n_channels": n_channels, |
| 278 | "n_layers": n_layers, |
| 279 | "use_vision_norm": use_vision_norm, |
| 280 | } |
| 281 | if decoder_dmodel is not None and ( |
| 282 | missing := [name for name, value in values.items() if value is None] |
| 283 | ): |
| 284 | raise ValueError( |
| 285 | "Enabled Inkling vision tower is missing config fields: " |
| 286 | + ", ".join(missing) |
| 287 | ) |
| 288 | self.vision_encoder_type = cast(Literal["linear", "hmlp"], vision_encoder_type) |
| 289 | self.decoder_dmodel = decoder_dmodel |
| 290 | self.patch_size = cast(int, patch_size) |
| 291 | self.temporal_patch_size = cast(int, temporal_patch_size) |
| 292 | self.n_channels = cast(int, n_channels) |
| 293 | self.n_layers = cast(int, n_layers) |
| 294 | self.use_vision_norm = cast(bool, use_vision_norm) |
| 295 | super().__init__(**kwargs) |
| 296 | |
| 297 | |
| 298 | class InklingMMConfig(PreTrainedConfig): |
| 299 | model_type = "inkling_mm_model" |
| 300 | keys_to_ignore_at_inference = ["past_key_values"] |
| 301 | sub_configs: ClassVar[dict[str, type[PreTrainedConfig]]] = { |
| 302 | "text_config": InklingModelConfig, |
| 303 | "audio_config": InklingAudioConfig, |
| 304 | "vision_config": InklingVisionConfig, |
| 305 | } |
| 306 | |
| 307 | def __init__( |
| 308 | self, |
| 309 | *, |
| 310 | text_config: dict[str, Any] | InklingModelConfig | None = None, |
| 311 | audio_config: dict[str, Any] | InklingAudioConfig | None = None, |
| 312 | vision_config: dict[str, Any] | InklingVisionConfig | None = None, |
| 313 | tie_word_embeddings: bool = False, |
| 314 | **kwargs: Any, |
| 315 | ) -> None: |
| 316 | self.text_config = ( |
| 317 | text_config |
| 318 | if isinstance(text_config, InklingModelConfig) |
| 319 | else InklingModelConfig(**(text_config or {})) |
| 320 | ) |
| 321 | self.audio_config = ( |
| 322 | audio_config |
| 323 | if isinstance(audio_config, InklingAudioConfig) |
| 324 | else InklingAudioConfig(**(audio_config or {})) |
| 325 | ) |
| 326 | self.vision_config = ( |
| 327 | vision_config |
| 328 | if isinstance(vision_config, InklingVisionConfig) |
| 329 | else InklingVisionConfig(**(vision_config or {})) |
| 330 | ) |
| 331 | super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs) |
| 332 | |
| 333 | def get_text_config(self, *args: Any, **kwargs: Any) -> InklingModelConfig: |
| 334 | return self.text_config |
| 335 | |
| 336 | @property |
| 337 | def vocab_size(self) -> int: |
| 338 | return self.text_config.vocab_size |
| 339 | |
| 340 | @property |
| 341 | def hidden_size(self) -> int: |
| 342 | return self.text_config.hidden_size |
| 343 | |
| 344 | @property |
| 345 | def num_hidden_layers(self) -> int: |
| 346 | return self.text_config.num_hidden_layers |
| 347 | |
| 348 | @property |
| 349 | def num_attention_heads(self) -> int: |
| 350 | return self.text_config.num_attention_heads |
| 351 | |
| 352 | @property |
| 353 | def num_key_value_heads(self) -> int: |
| 354 | return self.text_config.num_key_value_heads |
| 355 | |
| 356 | @property |
| 357 | def head_dim(self) -> int: |
| 358 | return self.text_config.head_dim |
| 359 | |
| 360 | @property |
| 361 | def full_attention_layer_ids(self) -> list[int]: |
| 362 | return self.text_config.full_attention_layer_ids |
| 363 | |
| 364 | @property |
| 365 | def linear_layer_ids(self) -> list[int]: |
| 366 | return self.text_config.linear_layer_ids |
| 367 | |
| 368 | @property |
| 369 | def conv_layer_ids(self) -> list[int]: |
| 370 | return self.text_config.conv_layer_ids |
| 371 | |
| 372 | @property |
| 373 | def mamba_chunk_size(self) -> int: |
| 374 | return self.text_config.mamba_chunk_size |
| 375 | |
| 376 | @property |
| 377 | def mamba2_cache_params(self) -> TMLConvCacheParams | None: |
| 378 | return self.text_config.mamba2_cache_params |
| 379 | |
| 380 | |
| 381 | @dataclass(kw_only=True, frozen=True) |
| 382 | class TMLConvStateShape: |
| 383 | conv: list[tuple[int, int]] |
| 384 | temporal: tuple[int, int, int] |
| 385 | |
| 386 | |
| 387 | @dataclass(kw_only=True, frozen=True) |
| 388 | class TMLStateDType: |
| 389 | conv: torch.dtype = torch.bfloat16 |
| 390 | temporal: torch.dtype = torch.bfloat16 |
| 391 | |
| 392 | |
| 393 | @dataclass(kw_only=True, frozen=True) |
| 394 | class TMLConvCacheParams: |
| 395 | shape: TMLConvStateShape |
| 396 | layers: list[int] |
| 397 | dtype: TMLStateDType = field(default_factory=TMLStateDType) |