| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Iterable |
| 4 | |
| 5 | import numpy as np |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | from einops import rearrange |
| 9 | from transformers import PreTrainedConfig |
| 10 | |
| 11 | from vllm.distributed import parallel_state |
| 12 | from vllm.distributed import utils as dist_utils |
| 13 | from vllm.model_executor.layers.activation import get_act_fn |
| 14 | from vllm.model_executor.layers.attention.mm_encoder_attention import ( |
| 15 | MMEncoderAttention, |
| 16 | ) |
| 17 | from vllm.model_executor.layers.conv import Conv3dLayer |
| 18 | from vllm.model_executor.layers.linear import ( |
| 19 | ColumnParallelLinear, |
| 20 | QKVParallelLinear, |
| 21 | RowParallelLinear, |
| 22 | ) |
| 23 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 24 | from vllm.model_executor.layers.rotary_embedding.mrope import triton_mrope |
| 25 | from vllm.model_executor.layers.rotary_embedding.mrope_vit_setup import ( |
| 26 | vit_mrope_setup, |
| 27 | ) |
| 28 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 29 | from vllm.model_executor.models.utils import maybe_prefix |
| 30 | from vllm.model_executor.models.vision import ( |
| 31 | get_vit_attn_backend, |
| 32 | is_vit_use_data_parallel, |
| 33 | ) |
| 34 | |
| 35 | |
| 36 | class MiniMaxVLPatchEmbed(nn.Module): |
| 37 | """Conv3d-based patch embedding. |
| 38 | |
| 39 | Takes flat tokens of shape (N, C * temporal_patch_size * patch_size²) |
| 40 | and projects each to a hidden-size embedding. |
| 41 | """ |
| 42 | |
| 43 | def __init__(self, config: PreTrainedConfig) -> None: |
| 44 | super().__init__() |
| 45 | compression = config.img_token_compression_config |
| 46 | temporal_patch_size = compression.get("temporal_patch_size", 2) |
| 47 | patch_size = config.patch_size |
| 48 | num_channels = config.num_channels |
| 49 | |
| 50 | self.patch_size = patch_size |
| 51 | self.temporal_patch_size = temporal_patch_size |
| 52 | self.num_channels = num_channels |
| 53 | self.hidden_size = config.hidden_size |
| 54 | |
| 55 | self.patch_embedding = Conv3dLayer( |
| 56 | in_channels=num_channels, |
| 57 | out_channels=config.hidden_size, |
| 58 | kernel_size=(temporal_patch_size, patch_size, patch_size), |
| 59 | stride=(temporal_patch_size, patch_size, patch_size), |
| 60 | bias=False, |
| 61 | ) |
| 62 | |
| 63 | def forward(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 64 | # pixel_values: (N, C * temporal_patch_size * patch_size²) |
| 65 | if self.patch_embedding.weight.dtype != pixel_values.dtype: |
| 66 | self.patch_embedding = self.patch_embedding.to(pixel_values.dtype) |
| 67 | x = pixel_values.reshape( |
| 68 | pixel_values.shape[0], |
| 69 | self.num_channels, |
| 70 | self.temporal_patch_size, |
| 71 | self.patch_size, |
| 72 | self.patch_size, |
| 73 | ) |
| 74 | return self.patch_embedding(x).reshape(x.shape[0], -1) |
| 75 | |
| 76 | |
| 77 | class MiniMaxVLAttention(nn.Module): |
| 78 | """Multi-head attention with MiniMax's partial 3D RoPE. |
| 79 | |
| 80 | Partial means only the first ``rot_dim`` (< head_dim) dimensions of |
| 81 | Q and K are rotated; the remaining dims are passed through unchanged. |
| 82 | """ |
| 83 | |
| 84 | def __init__( |
| 85 | self, |
| 86 | embed_dim: int, |
| 87 | num_heads: int, |
| 88 | quant_config: QuantizationConfig | None = None, |
| 89 | prefix: str = "", |
| 90 | ) -> None: |
| 91 | super().__init__() |
| 92 | use_data_parallel = is_vit_use_data_parallel() |
| 93 | self.tp_size = ( |
| 94 | 1 |
| 95 | if use_data_parallel |
| 96 | else parallel_state.get_tensor_model_parallel_world_size() |
| 97 | ) |
| 98 | self.head_dim = embed_dim // num_heads |
| 99 | self.num_heads_per_partition = dist_utils.divide(num_heads, self.tp_size) |
| 100 | |
| 101 | self.qkv_proj = QKVParallelLinear( |
| 102 | hidden_size=embed_dim, |
| 103 | head_size=self.head_dim, |
| 104 | total_num_heads=num_heads, |
| 105 | total_num_kv_heads=num_heads, |
| 106 | bias=True, |
| 107 | quant_config=quant_config, |
| 108 | prefix=f"{prefix}.qkv_proj", |
| 109 | disable_tp=use_data_parallel, |
| 110 | ) |
| 111 | self.out_proj = RowParallelLinear( |
| 112 | input_size=embed_dim, |
| 113 | output_size=embed_dim, |
| 114 | bias=True, |
| 115 | quant_config=quant_config, |
| 116 | prefix=f"{prefix}.out_proj", |
| 117 | disable_tp=use_data_parallel, |
| 118 | ) |
| 119 | self.attn = MMEncoderAttention( |
| 120 | num_heads=self.num_heads_per_partition, |
| 121 | head_size=self.head_dim, |
| 122 | prefix=f"{prefix}.attn", |
| 123 | ) |
| 124 | # Partial 3D RoPE geometry: rot_dim is split evenly across t/h/w |
| 125 | # (same formula as MiniMaxVLVisionTransformer; rot_dim may be < |
| 126 | # head_dim). mrope_section is in half-dim units for triton_mrope. |
| 127 | rope_dims = 2 * (self.head_dim // 2) |
| 128 | axis_dim = 2 * ((rope_dims // 3) // 2) |
| 129 | self.rotary_dim = 3 * axis_dim |
| 130 | self.mrope_section = [axis_dim // 2] * 3 |
| 131 | |
| 132 | def forward( |
| 133 | self, |
| 134 | x: torch.Tensor, |
| 135 | cu_seqlens: torch.Tensor, |
| 136 | rotary_cos: torch.Tensor, |
| 137 | rotary_sin: torch.Tensor, |
| 138 | max_seqlen: torch.Tensor, |
| 139 | sequence_lengths: torch.Tensor | None = None, |
| 140 | ) -> torch.Tensor: |
| 141 | # x: (N, 1, embed_dim) [seq=N, batch=1, chan=embed_dim] |
| 142 | x_qkv, _ = self.qkv_proj(x) # (N, 1, 3 * heads_per_part * head_dim) |
| 143 | seq_len, batch_size, _ = x_qkv.shape |
| 144 | |
| 145 | qkv = x_qkv.view(seq_len, 3, self.num_heads_per_partition, self.head_dim) |
| 146 | v = qkv[:, 2].unsqueeze(0) # (b=1, N, heads, head_dim) strided view |
| 147 | |
| 148 | # In-place partial 3D RoPE on strided q/k views of the packed qkv |
| 149 | # tensor: one kernel launch, no repack/upcast copies. fp32 cos/sin |
| 150 | # make the kernel compute in fp32, matching the reference |
| 151 | # ``_minimax_rope_applier`` precision. |
| 152 | q = qkv[:, 0].reshape(seq_len, -1) |
| 153 | k = qkv[:, 1].reshape(seq_len, -1) |
| 154 | triton_mrope( |
| 155 | q, |
| 156 | k, |
| 157 | rotary_cos, |
| 158 | rotary_sin, |
| 159 | self.mrope_section, |
| 160 | self.head_dim, |
| 161 | self.rotary_dim, |
| 162 | mrope_interleaved=False, |
| 163 | is_neox_style=True, |
| 164 | ) |
| 165 | q = q.view(batch_size, seq_len, self.num_heads_per_partition, self.head_dim) |
| 166 | k = k.view(batch_size, seq_len, self.num_heads_per_partition, self.head_dim) |
| 167 | |
| 168 | # Flash attention → (b, N, heads, head_dim) |
| 169 | context = self.attn( |
| 170 | query=q, |
| 171 | key=k, |
| 172 | value=v, |
| 173 | cu_seqlens=cu_seqlens, |
| 174 | max_seqlen=max_seqlen, |
| 175 | sequence_lengths=sequence_lengths, |
| 176 | ) |
| 177 | |
| 178 | # Back to (N, 1, embed_dim) |
| 179 | context = rearrange(context, "b s h d -> s b (h d)", b=batch_size) |
| 180 | output, _ = self.out_proj(context) |
| 181 | return output |
| 182 | |
| 183 | |
| 184 | class MiniMaxVLEncoderLayer(nn.Module): |
| 185 | """Single CLIP-style transformer block.""" |
| 186 | |
| 187 | def __init__( |
| 188 | self, |
| 189 | config: PreTrainedConfig, |
| 190 | quant_config: QuantizationConfig | None = None, |
| 191 | prefix: str = "", |
| 192 | ) -> None: |
| 193 | super().__init__() |
| 194 | embed_dim = config.hidden_size |
| 195 | self.layer_norm1 = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) |
| 196 | self.self_attn = MiniMaxVLAttention( |
| 197 | embed_dim=embed_dim, |
| 198 | num_heads=config.num_attention_heads, |
| 199 | quant_config=quant_config, |
| 200 | prefix=f"{prefix}.self_attn", |
| 201 | ) |
| 202 | self.layer_norm2 = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) |
| 203 | use_data_parallel = is_vit_use_data_parallel() |
| 204 | self.fc1 = ColumnParallelLinear( |
| 205 | config.hidden_size, |
| 206 | config.intermediate_size, |
| 207 | bias=True, |
| 208 | quant_config=quant_config, |
| 209 | prefix=f"{prefix}.fc1", |
| 210 | disable_tp=use_data_parallel, |
| 211 | ) |
| 212 | self.act = get_act_fn(getattr(config, "hidden_act", "gelu")) |
| 213 | self.fc2 = RowParallelLinear( |
| 214 | config.intermediate_size, |
| 215 | config.hidden_size, |
| 216 | bias=True, |
| 217 | quant_config=quant_config, |
| 218 | prefix=f"{prefix}.fc2", |
| 219 | disable_tp=use_data_parallel, |
| 220 | ) |
| 221 | |
| 222 | def forward( |
| 223 | self, |
| 224 | x: torch.Tensor, |
| 225 | cu_seqlens: torch.Tensor, |
| 226 | rotary_cos: torch.Tensor, |
| 227 | rotary_sin: torch.Tensor, |
| 228 | max_seqlen: torch.Tensor, |
| 229 | sequence_lengths: torch.Tensor | None = None, |
| 230 | ) -> torch.Tensor: |
| 231 | # x: (N, 1, hidden_size) |
| 232 | x = x + self.self_attn( |
| 233 | self.layer_norm1(x), |
| 234 | cu_seqlens, |
| 235 | rotary_cos, |
| 236 | rotary_sin, |
| 237 | max_seqlen, |
| 238 | sequence_lengths, |
| 239 | ) |
| 240 | residual = x |
| 241 | x, _ = self.fc1(self.layer_norm2(x)) |
| 242 | x = self.act(x) |
| 243 | x, _ = self.fc2(x) |
| 244 | return residual + x |
| 245 | |
| 246 | |
| 247 | class MiniMaxVLEncoder(nn.Module): |
| 248 | def __init__( |
| 249 | self, |
| 250 | config: PreTrainedConfig, |
| 251 | num_hidden_layers_override: int | None = None, |
| 252 | quant_config: QuantizationConfig | None = None, |
| 253 | prefix: str = "", |
| 254 | ) -> None: |
| 255 | super().__init__() |
| 256 | n = ( |
| 257 | config.num_hidden_layers |
| 258 | if num_hidden_layers_override is None |
| 259 | else num_hidden_layers_override |
| 260 | ) |
| 261 | self.layers = nn.ModuleList( |
| 262 | [ |
| 263 | MiniMaxVLEncoderLayer( |
| 264 | config=config, |
| 265 | quant_config=quant_config, |
| 266 | prefix=f"{prefix}.layers.{i}", |
| 267 | ) |
| 268 | for i in range(n) |
| 269 | ] |
| 270 | ) |
| 271 | |
| 272 | def forward( |
| 273 | self, |
| 274 | x: torch.Tensor, |
| 275 | cu_seqlens: torch.Tensor, |
| 276 | rotary_cos: torch.Tensor, |
| 277 | rotary_sin: torch.Tensor, |
| 278 | max_seqlen: torch.Tensor, |
| 279 | sequence_lengths: torch.Tensor | None = None, |
| 280 | ) -> torch.Tensor: |
| 281 | for layer in self.layers: |
| 282 | x = layer( |
| 283 | x, |
| 284 | cu_seqlens, |
| 285 | rotary_cos, |
| 286 | rotary_sin, |
| 287 | max_seqlen, |
| 288 | sequence_lengths, |
| 289 | ) |
| 290 | return x |
| 291 | |
| 292 | |
| 293 | class MiniMaxVLVisionTransformer(nn.Module): |
| 294 | """CLIP-based ViT with 3D RoPE (t/h/w decomposed). |
| 295 | |
| 296 | Faithfully mirrors the reference ``MiniMaxVLVisionTransformer``. |
| 297 | FLASHINFER backend is not supported; standard flash-attn is used. |
| 298 | """ |
| 299 | |
| 300 | def __init__( |
| 301 | self, |
| 302 | config: PreTrainedConfig, |
| 303 | num_hidden_layers_override: int | None = None, |
| 304 | require_post_norm: bool | None = None, |
| 305 | quant_config: QuantizationConfig | None = None, |
| 306 | prefix: str = "", |
| 307 | ) -> None: |
| 308 | super().__init__() |
| 309 | compression = config.img_token_compression_config |
| 310 | self.spatial_merge_size: int = compression.get("spatial_merge_size", 2) |
| 311 | self.temporal_patch_size: int = compression.get("temporal_patch_size", 2) |
| 312 | self.vision_segment_max_frames: int | None = getattr( |
| 313 | config, "vision_segment_max_frames", None |
| 314 | ) |
| 315 | self.use_data_parallel = is_vit_use_data_parallel() |
| 316 | |
| 317 | embed_dim = config.hidden_size |
| 318 | head_dim = embed_dim // config.num_attention_heads |
| 319 | # Backend selection + sharding info for building encoder metadata. |
| 320 | # Defaults to FLASH_ATTN on SM80+; --mm-encoder-attn-backend FLASHINFER |
| 321 | # selects the cuDNN ViT prefill path. |
| 322 | self.hidden_size = embed_dim |
| 323 | self.tp_size = ( |
| 324 | 1 |
| 325 | if self.use_data_parallel |
| 326 | else parallel_state.get_tensor_model_parallel_world_size() |
| 327 | ) |
| 328 | self.attn_backend = get_vit_attn_backend( |
| 329 | head_size=head_dim, dtype=torch.get_default_dtype() |
| 330 | ) |
| 331 | rope_dims = 2 * (head_dim // 2) |
| 332 | |
| 333 | # Split rope dims evenly across t/h/w (same formula as the reference) |
| 334 | self.t_dim = int(2 * ((rope_dims // 3) // 2)) |
| 335 | self.h_dim = int(2 * ((rope_dims // 3) // 2)) |
| 336 | self.w_dim = int(2 * ((rope_dims // 3) // 2)) |
| 337 | # rot_dim = t_dim + h_dim + w_dim (may be < head_dim) |
| 338 | |
| 339 | rope_theta: float = getattr(config, "rope_theta", 10000.0) |
| 340 | inv_freq_t = 1.0 / ( |
| 341 | rope_theta |
| 342 | ** (torch.arange(0, self.t_dim, 2, dtype=torch.float32) / self.t_dim) |
| 343 | ) |
| 344 | inv_freq_h = 1.0 / ( |
| 345 | rope_theta |
| 346 | ** (torch.arange(0, self.h_dim, 2, dtype=torch.float32) / self.h_dim) |
| 347 | ) |
| 348 | inv_freq_w = 1.0 / ( |
| 349 | rope_theta |
| 350 | ** (torch.arange(0, self.w_dim, 2, dtype=torch.float32) / self.w_dim) |
| 351 | ) |
| 352 | self.register_buffer("inv_freq_t", inv_freq_t, persistent=False) |
| 353 | self.register_buffer("inv_freq_h", inv_freq_h, persistent=False) |
| 354 | self.register_buffer("inv_freq_w", inv_freq_w, persistent=False) |
| 355 | |
| 356 | self.embeddings = MiniMaxVLPatchEmbed(config) |
| 357 | self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) |
| 358 | |
| 359 | n_layers = config.num_hidden_layers |
| 360 | if num_hidden_layers_override is None: |
| 361 | num_hidden_layers_override = n_layers |
| 362 | self.encoder = MiniMaxVLEncoder( |
| 363 | config=config, |
| 364 | num_hidden_layers_override=num_hidden_layers_override, |
| 365 | quant_config=quant_config, |
| 366 | prefix=f"{prefix}.encoder", |
| 367 | ) |
| 368 | |
| 369 | if require_post_norm is None: |
| 370 | require_post_norm = num_hidden_layers_override == n_layers |
| 371 | self.post_layernorm = ( |
| 372 | nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) |
| 373 | if require_post_norm |
| 374 | else None |
| 375 | ) |
| 376 | |
| 377 | # out_hidden_size needed by run_dp_sharded_mrope_vision_model |
| 378 | self.out_hidden_size = embed_dim |
| 379 | |
| 380 | # ── Frame-limit helper (mirrors the reference) ─────────────────────── |
| 381 | |
| 382 | def _apply_max_frames_limit(self, grid_thw: list[list[int]]) -> list[list[int]]: |
| 383 | if self.vision_segment_max_frames is None: |
| 384 | return grid_thw |
| 385 | max_f = self.vision_segment_max_frames |
| 386 | out: list[list[int]] = [] |
| 387 | for t, h, w in grid_thw: |
| 388 | if t <= max_f: |
| 389 | out.append([t, h, w]) |
| 390 | else: |
| 391 | for i in range(0, t, max_f): |
| 392 | out.append([min(max_f, t - i), h, w]) |
| 393 | return out |
| 394 | |
| 395 | # ── Encoder metadata (shared by eager forward and CUDA graph replay) ── |
| 396 | |
| 397 | def prepare_encoder_metadata( |
| 398 | self, |
| 399 | grid_thw: list[list[int]], |
| 400 | *, |
| 401 | device: torch.device, |
| 402 | max_batch_size: int | None = None, |
| 403 | max_seqlen_override: int | None = None, |
| 404 | ) -> dict[str, torch.Tensor | None]: |
| 405 | """Build all grid-dependent forward inputs as standalone tensors. |
| 406 | |
| 407 | The eager path calls this per forward; encoder CUDA graphs call it at |
| 408 | capture (dummy grids, padded to fixed buffer sizes) and before each |
| 409 | replay (real grids, unpadded — the manager's padding logic completes |
| 410 | the captured buffers). |
| 411 | |
| 412 | Args: |
| 413 | grid_thw: per-item [t, h, w] patch grids (host metadata). |
| 414 | device: device for the attention metadata tensors. |
| 415 | max_batch_size: pad cu_seqlens to this many segments with empty |
| 416 | (zero-length) trailing segments; None leaves it unpadded. |
| 417 | max_seqlen_override: bake this max_seqlen instead of deriving it |
| 418 | from the grids; CUDA graph capture passes a worst-case value |
| 419 | because max_seqlen is a CPU scalar baked into the graph. |
| 420 | |
| 421 | """ |
| 422 | limited = self._apply_max_frames_limit(grid_thw) |
| 423 | |
| 424 | # Token-level cumulative sequence lengths (one segment per limited grid). |
| 425 | lens = [t * h * w for t, h, w in limited] |
| 426 | cu_seqlens_np = np.zeros(len(lens) + 1, dtype=np.int32) |
| 427 | np.cumsum(np.array(lens, dtype=np.int32), out=cu_seqlens_np[1:]) |
| 428 | |
| 429 | if max_batch_size is not None and len(lens) < max_batch_size: |
| 430 | pad = np.full(max_batch_size - len(lens), cu_seqlens_np[-1], dtype=np.int32) |
| 431 | cu_seqlens_np = np.concatenate([cu_seqlens_np, pad]) |
| 432 | |
| 433 | # Backend-specific encoder metadata. For FLASH_ATTN this returns the raw |
| 434 | # token cu_seqlens, the max segment length, and sequence_lengths=None; |
| 435 | # for FLASHINFER (cuDNN) it repacks cu_seqlens into element-offset |
| 436 | # indptrs, buckets max_seqlen, and builds padded per-sequence lengths. |
| 437 | sequence_lengths = MMEncoderAttention.maybe_compute_seq_lens( |
| 438 | self.attn_backend, cu_seqlens_np, device |
| 439 | ) |
| 440 | max_seqlen = torch.tensor( |
| 441 | max_seqlen_override |
| 442 | if max_seqlen_override is not None |
| 443 | else MMEncoderAttention.compute_max_seqlen( |
| 444 | self.attn_backend, cu_seqlens_np |
| 445 | ), |
| 446 | dtype=torch.int32, |
| 447 | ) |
| 448 | cu_seqlens = MMEncoderAttention.maybe_recompute_cu_seqlens( |
| 449 | self.attn_backend, |
| 450 | cu_seqlens_np, |
| 451 | self.hidden_size, |
| 452 | self.tp_size, |
| 453 | device, |
| 454 | ) |
| 455 | |
| 456 | # 3D RoPE cos/sin: (3, total_N, half_rot_dim) fp32 t/h/w planes for |
| 457 | # triton_mrope from one fused setup kernel. Kept fp32 so the kernel |
| 458 | # computes the rotation in fp32, matching the reference. |
| 459 | rotary_cos, rotary_sin = vit_mrope_setup( |
| 460 | self.inv_freq_t, |
| 461 | self.inv_freq_h, |
| 462 | self.inv_freq_w, |
| 463 | limited, |
| 464 | self.spatial_merge_size, |
| 465 | ) |
| 466 | |
| 467 | return { |
| 468 | "rotary_cos": rotary_cos, |
| 469 | "rotary_sin": rotary_sin, |
| 470 | "cu_seqlens": cu_seqlens, |
| 471 | "max_seqlen": max_seqlen, |
| 472 | "sequence_lengths": sequence_lengths, |
| 473 | } |
| 474 | |
| 475 | # ── Forward ────────────────────────────────────────────────────────── |
| 476 | |
| 477 | def forward( |
| 478 | self, |
| 479 | pixel_values: torch.Tensor, |
| 480 | grid_thw: list[list[int]] | None = None, |
| 481 | *, |
| 482 | encoder_metadata: dict[str, torch.Tensor | None] | None = None, |
| 483 | ) -> torch.Tensor: |
| 484 | # pixel_values: (total_N, C * temporal_patch_size * patch_size²) |
| 485 | # Output: (total_N, hidden_size) |
| 486 | |
| 487 | hidden = self.embeddings(pixel_values) # (total_N, hidden_size) |
| 488 | hidden = self.pre_layrnorm(hidden) |
| 489 | |
| 490 | if encoder_metadata is None: |
| 491 | assert grid_thw is not None |
| 492 | encoder_metadata = self.prepare_encoder_metadata( |
| 493 | grid_thw, device=hidden.device |
| 494 | ) |
| 495 | |
| 496 | cu_seqlens = encoder_metadata["cu_seqlens"] |
| 497 | rotary_cos = encoder_metadata["rotary_cos"] |
| 498 | rotary_sin = encoder_metadata["rotary_sin"] |
| 499 | max_seqlen = encoder_metadata["max_seqlen"] |
| 500 | sequence_lengths = encoder_metadata.get("sequence_lengths") |
| 501 | assert cu_seqlens is not None |
| 502 | assert rotary_cos is not None |
| 503 | assert rotary_sin is not None |
| 504 | assert max_seqlen is not None |
| 505 | |
| 506 | # Encoder expects (N, 1, hidden_size) — add batch dim |
| 507 | hidden = hidden.unsqueeze(1) |
| 508 | |
| 509 | hidden = self.encoder( |
| 510 | hidden, |
| 511 | cu_seqlens, |
| 512 | rotary_cos, |
| 513 | rotary_sin, |
| 514 | max_seqlen, |
| 515 | sequence_lengths=sequence_lengths, |
| 516 | ) |
| 517 | hidden = hidden.squeeze(1) # back to (total_N, hidden_size) |
| 518 | |
| 519 | if self.post_layernorm is not None: |
| 520 | hidden = self.post_layernorm(hidden) |
| 521 | |
| 522 | return hidden |
| 523 | |
| 524 | |
| 525 | class MiniMaxVLMultiModalProjector(nn.Module): |
| 526 | """Two-layer MLP projector: vision_hidden → text_hidden.""" |
| 527 | |
| 528 | def __init__( |
| 529 | self, |
| 530 | vision_hidden_size: int, |
| 531 | text_hidden_size: int, |
| 532 | projector_hidden_size: int | None, |
| 533 | multimodal_projector_bias: bool, |
| 534 | projector_hidden_act: str = "gelu", |
| 535 | quant_config: QuantizationConfig | None = None, |
| 536 | prefix: str = "", |
| 537 | ) -> None: |
| 538 | super().__init__() |
| 539 | mid = projector_hidden_size if projector_hidden_size else text_hidden_size |
| 540 | use_dp = is_vit_use_data_parallel() |
| 541 | self.linear_1 = ColumnParallelLinear( |
| 542 | vision_hidden_size, |
| 543 | mid, |
| 544 | bias=multimodal_projector_bias, |
| 545 | quant_config=quant_config, |
| 546 | prefix=f"{prefix}.linear_1", |
| 547 | disable_tp=use_dp, |
| 548 | ) |
| 549 | self.act = get_act_fn(projector_hidden_act) |
| 550 | self.linear_2 = RowParallelLinear( |
| 551 | mid, |
| 552 | text_hidden_size, |
| 553 | bias=multimodal_projector_bias, |
| 554 | quant_config=quant_config, |
| 555 | prefix=f"{prefix}.linear_2", |
| 556 | disable_tp=use_dp, |
| 557 | ) |
| 558 | |
| 559 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 560 | x, _ = self.linear_1(x) |
| 561 | x = self.act(x) |
| 562 | x, _ = self.linear_2(x) |
| 563 | return x |
| 564 | |
| 565 | |
| 566 | class MiniMaxVLPatchMerger(nn.Module): |
| 567 | def __init__( |
| 568 | self, |
| 569 | spatial_merge_size: int, |
| 570 | text_hidden_size: int, |
| 571 | projector_hidden_size: int | None, |
| 572 | patch_merge_bias: bool, |
| 573 | projector_hidden_act: str = "gelu", |
| 574 | quant_config: QuantizationConfig | None = None, |
| 575 | prefix: str = "", |
| 576 | ) -> None: |
| 577 | super().__init__() |
| 578 | self.spatial_merge_size = spatial_merge_size |
| 579 | mid = projector_hidden_size if projector_hidden_size else text_hidden_size |
| 580 | merge_in = text_hidden_size * spatial_merge_size**2 |
| 581 | use_dp = is_vit_use_data_parallel() |
| 582 | self.linear_1 = ColumnParallelLinear( |
| 583 | merge_in, |
| 584 | mid, |
| 585 | bias=patch_merge_bias, |
| 586 | quant_config=quant_config, |
| 587 | prefix=f"{prefix}.linear_1", |
| 588 | disable_tp=use_dp, |
| 589 | ) |
| 590 | self.act = get_act_fn(projector_hidden_act) |
| 591 | self.linear_2 = RowParallelLinear( |
| 592 | mid, |
| 593 | text_hidden_size, |
| 594 | bias=patch_merge_bias, |
| 595 | quant_config=quant_config, |
| 596 | prefix=f"{prefix}.linear_2", |
| 597 | disable_tp=use_dp, |
| 598 | ) |
| 599 | |
| 600 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 601 | # x: (N, text_hidden_size) → (N // merge_size², text_hidden_size) |
| 602 | x = x.reshape(x.shape[0] // (self.spatial_merge_size**2), -1) |
| 603 | x, _ = self.linear_1(x) |
| 604 | x = self.act(x) |
| 605 | x, _ = self.linear_2(x) |
| 606 | return x |
| 607 | |
| 608 | |
| 609 | class MiniMaxVLVisionModel(nn.Module): |
| 610 | """Full vision model: ViT → projector → patch merger.""" |
| 611 | |
| 612 | def __init__( |
| 613 | self, |
| 614 | config: PreTrainedConfig, |
| 615 | text_hidden_size: int, |
| 616 | projector_hidden_size: int | None = None, |
| 617 | quant_config: QuantizationConfig | None = None, |
| 618 | prefix: str = "", |
| 619 | ) -> None: |
| 620 | super().__init__() |
| 621 | compression = config.img_token_compression_config |
| 622 | spatial_merge_size: int = compression.get("spatial_merge_size", 2) |
| 623 | self.spatial_merge_size = spatial_merge_size |
| 624 | self.use_data_parallel = is_vit_use_data_parallel() |
| 625 | |
| 626 | # The released checkpoint ships no ``post_layernorm`` weights and |
| 627 | # uses ``vision_feature_layer=-1`` with ``vision_feature_select_strategy |
| 628 | # ="full"``, i.e. the raw last encoder hidden state (CLIP's |
| 629 | # ``last_hidden_state`` is taken before the post layernorm). Applying an |
| 630 | # untrained post layernorm here would corrupt the visual features. |
| 631 | self.vision_model = MiniMaxVLVisionTransformer( |
| 632 | config=config, |
| 633 | require_post_norm=False, |
| 634 | quant_config=quant_config, |
| 635 | prefix=maybe_prefix(prefix, "vision_model"), |
| 636 | ) |
| 637 | self.multi_modal_projector = MiniMaxVLMultiModalProjector( |
| 638 | vision_hidden_size=config.hidden_size, |
| 639 | text_hidden_size=text_hidden_size, |
| 640 | projector_hidden_size=projector_hidden_size, |
| 641 | multimodal_projector_bias=getattr( |
| 642 | config, "multimodal_projector_bias", True |
| 643 | ), |
| 644 | projector_hidden_act=getattr(config, "projector_hidden_act", "gelu"), |
| 645 | quant_config=quant_config, |
| 646 | prefix=maybe_prefix(prefix, "multi_modal_projector"), |
| 647 | ) |
| 648 | self.patch_merge_mlp = MiniMaxVLPatchMerger( |
| 649 | spatial_merge_size=spatial_merge_size, |
| 650 | text_hidden_size=text_hidden_size, |
| 651 | projector_hidden_size=projector_hidden_size, |
| 652 | patch_merge_bias=getattr(config, "patch_merge_bias", True), |
| 653 | projector_hidden_act=getattr(config, "projector_hidden_act", "gelu"), |
| 654 | quant_config=quant_config, |
| 655 | prefix=maybe_prefix(prefix, "patch_merge_mlp"), |
| 656 | ) |
| 657 | |
| 658 | self.dtype = self.vision_model.embeddings.patch_embedding.weight.dtype |
| 659 | self.out_hidden_size = text_hidden_size |
| 660 | |
| 661 | def forward( |
| 662 | self, |
| 663 | pixel_values: torch.Tensor, |
| 664 | grid_thw: list[list[int]] | None = None, |
| 665 | *, |
| 666 | encoder_metadata: dict[str, torch.Tensor | None] | None = None, |
| 667 | ) -> torch.Tensor: |
| 668 | hidden = self.vision_model( |
| 669 | pixel_values=pixel_values, |
| 670 | grid_thw=grid_thw, |
| 671 | encoder_metadata=encoder_metadata, |
| 672 | ) |
| 673 | if hidden.dim() == 3: |
| 674 | hidden = hidden.squeeze(0) |
| 675 | hidden = self.multi_modal_projector(hidden) |
| 676 | hidden = self.patch_merge_mlp(hidden) |
| 677 | return hidden |
| 678 | |
| 679 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 680 | stacked_params_mapping = [ |
| 681 | # (param_name, shard_name, shard_id) |
| 682 | ("qkv_proj.", "q_proj.", "q"), |
| 683 | ("qkv_proj.", "k_proj.", "k"), |
| 684 | ("qkv_proj.", "v_proj.", "v"), |
| 685 | ] |
| 686 | params_dict = dict(self.named_parameters(remove_duplicate=False)) |
| 687 | loaded_params: set[str] = set() |
| 688 | |
| 689 | for name, loaded_weight in weights: |
| 690 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 691 | if weight_name not in name: |
| 692 | continue |
| 693 | name = name.replace(weight_name, param_name) |
| 694 | |
| 695 | param = params_dict[name] |
| 696 | weight_loader = param.weight_loader |
| 697 | weight_loader(param, loaded_weight, shard_id) |
| 698 | break |
| 699 | else: |
| 700 | param = params_dict[name] |
| 701 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 702 | weight_loader(param, loaded_weight) |
| 703 | loaded_params.add(name) |
| 704 | return loaded_params |