| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Vision tower implementation for Kimi-K2.5 model. |
| 4 | |
| 5 | This module provides the vision encoder components for Kimi-K2.5, |
| 6 | including 3D patch embedding, RoPE position embedding, and |
| 7 | temporal pooling for video chunks. |
| 8 | """ |
| 9 | |
| 10 | from collections.abc import Sequence |
| 11 | from copy import deepcopy |
| 12 | from typing import Any |
| 13 | |
| 14 | import numpy as np |
| 15 | import torch |
| 16 | import torch.nn as nn |
| 17 | import torch.nn.functional as F |
| 18 | from transformers.activations import GELUActivation |
| 19 | |
| 20 | from vllm.distributed import divide, get_tensor_model_parallel_world_size |
| 21 | from vllm.logger import init_logger |
| 22 | from vllm.model_executor.layers.activation import get_act_fn |
| 23 | from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention |
| 24 | from vllm.model_executor.layers.conv import Conv2dLayer |
| 25 | from vllm.model_executor.layers.linear import ( |
| 26 | ColumnParallelLinear, |
| 27 | QKVParallelLinear, |
| 28 | ReplicatedLinear, |
| 29 | RowParallelLinear, |
| 30 | ) |
| 31 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 32 | from vllm.model_executor.layers.rotary_embedding.common import ApplyRotaryEmb |
| 33 | from vllm.model_executor.layers.rotary_embedding.packed_qk_rope import ( |
| 34 | packed_qk_rope_, |
| 35 | ) |
| 36 | from vllm.model_executor.models.utils import maybe_prefix |
| 37 | from vllm.model_executor.models.vision import ( |
| 38 | is_vit_use_data_parallel, |
| 39 | run_dp_sharded_mrope_vision_model, |
| 40 | ) |
| 41 | from vllm.platforms import current_platform |
| 42 | from vllm.transformers_utils.configs.kimi_k25 import KimiK25VisionConfig |
| 43 | from vllm.triton_utils import HAS_TRITON |
| 44 | from vllm.utils.torch_utils import async_tensor_h2d |
| 45 | |
| 46 | logger = init_logger(__name__) |
| 47 | |
| 48 | |
| 49 | def _apply_rope_input_validation(x, freqs_cis): |
| 50 | assert x.ndim == freqs_cis.ndim + 1, (x.shape, freqs_cis.shape) |
| 51 | assert x.shape[:-2] == freqs_cis.shape[:-1], (x.shape, freqs_cis.shape) |
| 52 | assert x.shape[-1] == 2 * freqs_cis.shape[-1], (x.shape, freqs_cis.shape) |
| 53 | assert freqs_cis.dtype == torch.complex64, freqs_cis.dtype |
| 54 | |
| 55 | |
| 56 | def get_rope_shape_decorate(func): |
| 57 | _get_rope_shape_first_call_flag = set() |
| 58 | |
| 59 | def wrapper(org, interpolation_mode, shape): |
| 60 | key = (org.requires_grad, torch.is_grad_enabled(), interpolation_mode) |
| 61 | if key not in _get_rope_shape_first_call_flag: |
| 62 | _get_rope_shape_first_call_flag.add(key) |
| 63 | _ = func(org, interpolation_mode, shape=(64, 64)) |
| 64 | return func(org, interpolation_mode, shape) |
| 65 | |
| 66 | return wrapper |
| 67 | |
| 68 | |
| 69 | @get_rope_shape_decorate |
| 70 | @torch.compile( |
| 71 | dynamic=True, |
| 72 | backend=current_platform.simple_compile_backend, |
| 73 | disable=current_platform.simple_compile_backend == "tpu", |
| 74 | ) |
| 75 | def get_rope_shape(org, interpolation_mode, shape): |
| 76 | return ( |
| 77 | F.interpolate( |
| 78 | org.permute((2, 0, 1)).unsqueeze(0), |
| 79 | size=shape, |
| 80 | mode=interpolation_mode, |
| 81 | ) |
| 82 | .squeeze(0) |
| 83 | .permute((1, 2, 0)) |
| 84 | .flatten(end_dim=1) |
| 85 | ) |
| 86 | |
| 87 | |
| 88 | def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): |
| 89 | """Generate 1D sincos positional embedding from grid positions.""" |
| 90 | assert embed_dim % 2 == 0 |
| 91 | omega = np.arange(embed_dim // 2, dtype=np.float32) |
| 92 | omega /= embed_dim / 2.0 |
| 93 | omega = 1.0 / 10000**omega # (D/2,) |
| 94 | |
| 95 | pos = pos.reshape(-1) # (M,) |
| 96 | out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product |
| 97 | |
| 98 | emb_sin = np.sin(out) # (M, D/2) |
| 99 | emb_cos = np.cos(out) # (M, D/2) |
| 100 | |
| 101 | emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) |
| 102 | return emb |
| 103 | |
| 104 | |
| 105 | def get_1d_sincos_pos_embed(embed_dim, t_size, cls_token=False): |
| 106 | """Generate 1D sincos positional embedding.""" |
| 107 | grid_t = np.arange(t_size, dtype=np.float32) |
| 108 | pos_embed = get_1d_sincos_pos_embed_from_grid(embed_dim, grid_t) |
| 109 | if cls_token: |
| 110 | pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0) |
| 111 | return pos_embed |
| 112 | |
| 113 | |
| 114 | class Learnable2DInterpPosEmbDivided_fixed(nn.Module): |
| 115 | """2D learnable position embedding with temporal extension.""" |
| 116 | |
| 117 | def __init__( |
| 118 | self, |
| 119 | height: int, |
| 120 | width: int, |
| 121 | num_frames: int, |
| 122 | dim: int, |
| 123 | interpolation_mode: str = "bicubic", |
| 124 | ) -> None: |
| 125 | super().__init__() |
| 126 | self.height = height |
| 127 | self.width = width |
| 128 | self.num_frames = num_frames |
| 129 | self.dim = dim |
| 130 | self.interpolation_mode = interpolation_mode |
| 131 | self.weight = nn.Parameter(torch.empty(height, width, dim)) |
| 132 | self.register_buffer( |
| 133 | "time_weight", |
| 134 | torch.from_numpy(get_1d_sincos_pos_embed(self.dim, self.num_frames)) |
| 135 | .float() |
| 136 | .unsqueeze(1), |
| 137 | persistent=False, |
| 138 | ) |
| 139 | |
| 140 | self.reset_parameters() |
| 141 | |
| 142 | def reset_parameters(self): |
| 143 | nn.init.normal_(self.weight) |
| 144 | |
| 145 | def get_pos_embeds(self, grid_thws: torch.Tensor | list[list[int]]) -> torch.Tensor: |
| 146 | pos_embs = [] |
| 147 | grid_thw_list = grid_thws if isinstance(grid_thws, list) else grid_thws.tolist() |
| 148 | for t, h, w in grid_thw_list: |
| 149 | assert t <= self.num_frames, f"t:{t} > self.num_frames:{self.num_frames}" |
| 150 | if (h, w) == self.weight.shape[:-1]: |
| 151 | pos_emb_2d = self.weight.flatten(end_dim=1) |
| 152 | else: |
| 153 | pos_emb_2d = get_rope_shape( |
| 154 | self.weight, |
| 155 | interpolation_mode=self.interpolation_mode, |
| 156 | shape=(h, w), |
| 157 | ) |
| 158 | |
| 159 | if t == 1: |
| 160 | pos_emb_3d = pos_emb_2d |
| 161 | else: |
| 162 | pos_emb_3d = ( |
| 163 | pos_emb_2d.unsqueeze(0).repeat(t, 1, 1) + self.time_weight[0:t] |
| 164 | ) |
| 165 | |
| 166 | pos_embs.append(pos_emb_3d.reshape(-1, pos_emb_3d.shape[-1])) |
| 167 | |
| 168 | return torch.cat(pos_embs) |
| 169 | |
| 170 | def forward( |
| 171 | self, x: torch.Tensor, grid_thws: torch.Tensor | list[list[int]] |
| 172 | ) -> torch.Tensor: |
| 173 | return x + self.get_pos_embeds(grid_thws) |
| 174 | |
| 175 | |
| 176 | class MoonVision3dPatchEmbed(nn.Module): |
| 177 | """3D patch embedding for vision tower.""" |
| 178 | |
| 179 | def __init__( |
| 180 | self, |
| 181 | out_dim: int, |
| 182 | in_dim: int = 3, |
| 183 | patch_size: int | tuple[int, int] = (14, 14), |
| 184 | pos_emb_height: int = 14, |
| 185 | pos_emb_width: int = 14, |
| 186 | pos_emb_time: int = 4, |
| 187 | pos_emb_type: str = "divided_fixed", |
| 188 | patch_embed_proj_bias: bool = True, |
| 189 | pos_emb_interpolation_mode: str = "bicubic", |
| 190 | ): |
| 191 | super().__init__() |
| 192 | assert isinstance(patch_size, int | Sequence), ( |
| 193 | f"Invalid patch_size type: {type(patch_size)}" |
| 194 | ) |
| 195 | if isinstance(patch_size, int): |
| 196 | patch_size = (patch_size, patch_size) |
| 197 | assert len(patch_size) == 2, ( |
| 198 | f"Expected patch_size to be a tuple of 2, got {patch_size}" |
| 199 | ) |
| 200 | self.patch_size = patch_size |
| 201 | |
| 202 | self.proj = Conv2dLayer( |
| 203 | in_dim, |
| 204 | out_dim, |
| 205 | kernel_size=patch_size, |
| 206 | stride=patch_size, |
| 207 | bias=patch_embed_proj_bias, |
| 208 | ) |
| 209 | |
| 210 | if pos_emb_type == "divided_fixed": |
| 211 | self.pos_emb = Learnable2DInterpPosEmbDivided_fixed( |
| 212 | height=pos_emb_height, |
| 213 | width=pos_emb_width, |
| 214 | num_frames=pos_emb_time, |
| 215 | dim=out_dim, |
| 216 | interpolation_mode=pos_emb_interpolation_mode, |
| 217 | ) |
| 218 | else: |
| 219 | raise NotImplementedError(f"Not support pos_emb_type: {pos_emb_type}") |
| 220 | |
| 221 | def forward( |
| 222 | self, |
| 223 | x: torch.Tensor, |
| 224 | grid_thws: torch.Tensor | list[list[int]] | None, |
| 225 | *, |
| 226 | pos_embeds: torch.Tensor | None = None, |
| 227 | ) -> torch.Tensor: |
| 228 | # forward_native dispatches this non-overlapping patch projection to GEMM. |
| 229 | x = self.proj.forward_native(x).view(x.size(0), self.proj.out_channels) |
| 230 | if pos_embeds is not None: |
| 231 | return x + pos_embeds |
| 232 | assert grid_thws is not None |
| 233 | return self.pos_emb(x, grid_thws) |
| 234 | |
| 235 | |
| 236 | class Rope2DPosEmbRepeated(nn.Module): |
| 237 | """2D rotary position embedding with multi-resolution support.""" |
| 238 | |
| 239 | def __init__(self, dim: int, max_height: int, max_width: int, theta_base=10000): |
| 240 | super().__init__() |
| 241 | self.dim = dim |
| 242 | assert self.dim % 4 == 0, "dim must be divisible by 4" |
| 243 | self.max_height = max_height |
| 244 | self.max_width = max_width |
| 245 | self.theta_base = theta_base |
| 246 | |
| 247 | def extra_repr(self): |
| 248 | return ( |
| 249 | f"dim={self.dim}, max_height={self.max_height}, " |
| 250 | f"max_width={self.max_width}, theta_base={self.theta_base}" |
| 251 | ) |
| 252 | |
| 253 | def _precompute_freqs_cis(self, device: torch.device) -> torch.Tensor: |
| 254 | """Calculate the cis(freqs) for each position in the 2D grid.""" |
| 255 | N = self.max_height * self.max_width |
| 256 | flat_pos = torch.arange(0, N).float().to(device) |
| 257 | x_pos = flat_pos % self.max_width |
| 258 | y_pos = flat_pos // self.max_width |
| 259 | dim_range = ( |
| 260 | torch.arange(0, self.dim, 4)[: (self.dim // 4)].float().to(device) |
| 261 | ) # C/4 |
| 262 | freqs = 1.0 / (self.theta_base ** (dim_range / self.dim)) |
| 263 | x_freqs = torch.outer(x_pos, freqs).float() # N, C/4 |
| 264 | y_freqs = torch.outer(y_pos, freqs).float() # N, C/4 |
| 265 | x_cis = torch.polar(torch.ones_like(x_freqs), x_freqs) # N, C/4 |
| 266 | y_cis = torch.polar(torch.ones_like(y_freqs), y_freqs) # N, C/4 |
| 267 | # N, C/4, 2 |
| 268 | freqs_cis = torch.cat( |
| 269 | [x_cis.unsqueeze(dim=-1), y_cis.unsqueeze(dim=-1)], dim=-1 |
| 270 | ) |
| 271 | # max_height, max_width, C/2 |
| 272 | freqs_cis = freqs_cis.reshape(self.max_height, self.max_width, -1) |
| 273 | return freqs_cis |
| 274 | |
| 275 | def get_freqs_cis( |
| 276 | self, grid_thws: torch.Tensor | list[list[int]], device: torch.device |
| 277 | ) -> torch.Tensor: |
| 278 | """Args: |
| 279 | grid_thws (torch.Tensor): grid time, height and width |
| 280 | |
| 281 | Returns: |
| 282 | freqs_cis: tensor of shape (sum(t * height * width), dim//2) |
| 283 | |
| 284 | """ |
| 285 | if not hasattr(self, "freqs_cis"): |
| 286 | self.register_buffer( |
| 287 | "freqs_cis", self._precompute_freqs_cis(device), persistent=False |
| 288 | ) |
| 289 | |
| 290 | shapes = grid_thws if isinstance(grid_thws, list) else grid_thws.tolist() |
| 291 | assert all( |
| 292 | 1 <= h <= self.max_height and 1 <= w <= self.max_width for t, h, w in shapes |
| 293 | ), ( |
| 294 | shapes, |
| 295 | self.max_height, |
| 296 | self.max_width, |
| 297 | ) |
| 298 | freqs_cis = torch.cat( |
| 299 | [ |
| 300 | self.freqs_cis[:h, :w].reshape(-1, self.dim // 2).repeat(t, 1) |
| 301 | for t, h, w in shapes |
| 302 | ], |
| 303 | dim=0, |
| 304 | ) |
| 305 | return freqs_cis |
| 306 | |
| 307 | |
| 308 | class MLP2(nn.Module): |
| 309 | """Two-layer MLP with tensor parallel support.""" |
| 310 | |
| 311 | def __init__( |
| 312 | self, |
| 313 | dims: list[int], |
| 314 | activation, |
| 315 | bias: bool = True, |
| 316 | quant_config: QuantizationConfig | None = None, |
| 317 | prefix: str = "", |
| 318 | use_data_parallel: bool = False, |
| 319 | ): |
| 320 | super().__init__() |
| 321 | assert len(dims) == 3 |
| 322 | self.use_data_parallel = use_data_parallel |
| 323 | self.fc0 = ColumnParallelLinear( |
| 324 | dims[0], |
| 325 | dims[1], |
| 326 | bias=bias, |
| 327 | quant_config=quant_config, |
| 328 | prefix=maybe_prefix(prefix, "fc0"), |
| 329 | disable_tp=self.use_data_parallel, |
| 330 | ) |
| 331 | self.fc1 = RowParallelLinear( |
| 332 | dims[1], |
| 333 | dims[2], |
| 334 | bias=bias, |
| 335 | quant_config=quant_config, |
| 336 | prefix=maybe_prefix(prefix, "fc1"), |
| 337 | disable_tp=self.use_data_parallel, |
| 338 | ) |
| 339 | self.activation = activation |
| 340 | |
| 341 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 342 | x, _ = self.fc0(x) |
| 343 | x = self.activation(x) |
| 344 | x, _ = self.fc1(x) |
| 345 | return x |
| 346 | |
| 347 | |
| 348 | def _make_vision_norm(norm_type: str, hidden_dim: int) -> nn.Module: |
| 349 | if norm_type == "layernorm": |
| 350 | return nn.LayerNorm(hidden_dim) |
| 351 | if norm_type == "rmsnorm": |
| 352 | return nn.RMSNorm(hidden_dim) |
| 353 | raise NotImplementedError(f"Not support norm_type: {norm_type}") |
| 354 | |
| 355 | |
| 356 | class MoonViTEncoderLayer(nn.Module): |
| 357 | """Single encoder layer for MoonViT with TP/DP support.""" |
| 358 | |
| 359 | def __init__( |
| 360 | self, |
| 361 | num_heads: int, |
| 362 | hidden_dim: int, |
| 363 | mlp_dim: int, |
| 364 | quant_config: QuantizationConfig | None = None, |
| 365 | prefix: str = "", |
| 366 | *, |
| 367 | activation=F.gelu, |
| 368 | attn_bias: bool = False, |
| 369 | qkv_hidden_size: int | None = None, |
| 370 | norm_type: str = "layernorm", |
| 371 | mlp_type: str = "mlp2", |
| 372 | linear_bias: bool = True, |
| 373 | ): |
| 374 | super().__init__() |
| 375 | self.use_data_parallel = is_vit_use_data_parallel(num_heads) |
| 376 | |
| 377 | self.num_heads = num_heads |
| 378 | self.hidden_dim = hidden_dim |
| 379 | self.qkv_hidden_size = ( |
| 380 | hidden_dim if qkv_hidden_size is None else qkv_hidden_size |
| 381 | ) |
| 382 | self.hidden_size_per_attention_head = self.qkv_hidden_size // self.num_heads |
| 383 | self.tp_size = ( |
| 384 | 1 if self.use_data_parallel else get_tensor_model_parallel_world_size() |
| 385 | ) |
| 386 | self.num_attention_heads_per_partition = divide(num_heads, self.tp_size) |
| 387 | |
| 388 | self.norm0 = _make_vision_norm(norm_type, hidden_dim) |
| 389 | self.norm1 = _make_vision_norm(norm_type, hidden_dim) |
| 390 | if mlp_type != "mlp2": |
| 391 | raise NotImplementedError(f"Not support mlp_type: {mlp_type}") |
| 392 | self.mlp = MLP2( |
| 393 | [hidden_dim, mlp_dim, hidden_dim], |
| 394 | activation, |
| 395 | bias=linear_bias, |
| 396 | quant_config=quant_config, |
| 397 | prefix=f"{prefix}.mlp", |
| 398 | use_data_parallel=self.use_data_parallel, |
| 399 | ) |
| 400 | self.wqkv = QKVParallelLinear( |
| 401 | hidden_size=hidden_dim, |
| 402 | head_size=self.hidden_size_per_attention_head, |
| 403 | total_num_heads=num_heads, |
| 404 | total_num_kv_heads=num_heads, |
| 405 | bias=attn_bias, |
| 406 | quant_config=quant_config, |
| 407 | prefix=f"{prefix}.wqkv", |
| 408 | disable_tp=self.use_data_parallel, |
| 409 | ) |
| 410 | self.wo = RowParallelLinear( |
| 411 | self.qkv_hidden_size, |
| 412 | hidden_dim, |
| 413 | bias=attn_bias, |
| 414 | quant_config=quant_config, |
| 415 | prefix=f"{prefix}.wo", |
| 416 | disable_tp=self.use_data_parallel, |
| 417 | ) |
| 418 | self.attn = MMEncoderAttention( |
| 419 | num_heads=self.num_attention_heads_per_partition, |
| 420 | head_size=self.hidden_size_per_attention_head, |
| 421 | scale=self.hidden_size_per_attention_head**-0.5, |
| 422 | prefix=f"{prefix}.attn", |
| 423 | ) |
| 424 | self.apply_rotary_emb = ApplyRotaryEmb( |
| 425 | enforce_enable=True, |
| 426 | is_neox_style=False, |
| 427 | enable_fp32_compute=True, |
| 428 | ) |
| 429 | |
| 430 | def attention_qkvpacked( |
| 431 | self, |
| 432 | x: torch.Tensor, |
| 433 | cu_seqlens: torch.Tensor, |
| 434 | rope_freqs_cis: torch.Tensor, |
| 435 | max_seqlen: torch.Tensor | None = None, |
| 436 | sequence_lengths: torch.Tensor | None = None, |
| 437 | ): |
| 438 | """Compute self-attention with packed QKV. |
| 439 | |
| 440 | Args: |
| 441 | x (torch.Tensor): (seqlen, hidden_dim) |
| 442 | cu_seqlens (torch.Tensor): cumulative sequence lengths |
| 443 | rope_freqs_cis (torch.Tensor): rotary embedding frequencies |
| 444 | max_seqlen (torch.Tensor | None): longest sequence in the batch |
| 445 | sequence_lengths (torch.Tensor | None): per-sequence lengths |
| 446 | |
| 447 | """ |
| 448 | seq_length = x.size(0) |
| 449 | xqkv, _ = self.wqkv(x) |
| 450 | |
| 451 | qkv_shape = xqkv.size()[:-1] + ( |
| 452 | 3, |
| 453 | self.num_attention_heads_per_partition, |
| 454 | self.hidden_size_per_attention_head, |
| 455 | ) |
| 456 | # xqkv: (seqlen, 3, nheads, headdim) |
| 457 | xqkv = xqkv.view(*qkv_shape) |
| 458 | |
| 459 | # xq/xk alias xqkv, so the in-place rotation below is visible through |
| 460 | # them; only the fallback needs to rebind. |
| 461 | xq, xk, xv = torch.unbind(xqkv, dim=-3) |
| 462 | if HAS_TRITON: |
| 463 | packed_qk_rope_(xqkv, rope_freqs_cis) |
| 464 | else: |
| 465 | _apply_rope_input_validation(xq, rope_freqs_cis) |
| 466 | _apply_rope_input_validation(xk, rope_freqs_cis) |
| 467 | rope_cos = rope_freqs_cis.real.contiguous() |
| 468 | rope_sin = rope_freqs_cis.imag.contiguous() |
| 469 | xq = self.apply_rotary_emb(xq, rope_cos, rope_sin) |
| 470 | xk = self.apply_rotary_emb(xk, rope_cos, rope_sin) |
| 471 | |
| 472 | if max_seqlen is None: |
| 473 | max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max() |
| 474 | attn_out = self.attn( |
| 475 | xq.unsqueeze(0), |
| 476 | xk.unsqueeze(0), |
| 477 | xv.unsqueeze(0), |
| 478 | cu_seqlens=cu_seqlens, |
| 479 | max_seqlen=max_seqlen, |
| 480 | sequence_lengths=sequence_lengths, |
| 481 | ) |
| 482 | attn_out = attn_out.reshape( |
| 483 | seq_length, |
| 484 | self.num_attention_heads_per_partition |
| 485 | * self.hidden_size_per_attention_head, |
| 486 | ) |
| 487 | attn_out, _ = self.wo(attn_out) |
| 488 | return attn_out |
| 489 | |
| 490 | def forward( |
| 491 | self, |
| 492 | hidden_states: torch.Tensor, |
| 493 | cu_seqlens: torch.Tensor, |
| 494 | rope_freqs_cis: torch.Tensor, |
| 495 | max_seqlen: torch.Tensor | None = None, |
| 496 | sequence_lengths: torch.Tensor | None = None, |
| 497 | ): |
| 498 | residual = hidden_states |
| 499 | hidden_states = self.norm0(hidden_states) |
| 500 | |
| 501 | hidden_states = self.attention_qkvpacked( |
| 502 | hidden_states, |
| 503 | cu_seqlens, |
| 504 | rope_freqs_cis, |
| 505 | max_seqlen=max_seqlen, |
| 506 | sequence_lengths=sequence_lengths, |
| 507 | ) |
| 508 | hidden_states = residual + hidden_states |
| 509 | |
| 510 | residual = hidden_states |
| 511 | hidden_states = self.norm1(hidden_states) |
| 512 | hidden_states = self.mlp(hidden_states) |
| 513 | hidden_states = residual + hidden_states |
| 514 | |
| 515 | return hidden_states |
| 516 | |
| 517 | |
| 518 | class MoonViT3dEncoder(nn.Module): |
| 519 | """Full encoder stack for MoonViT 3D.""" |
| 520 | |
| 521 | def __init__( |
| 522 | self, |
| 523 | hidden_dim: int, |
| 524 | num_layers: int, |
| 525 | block_cfg: dict, |
| 526 | video_attn_type: str = "spatial_temporal", |
| 527 | quant_config: QuantizationConfig | None = None, |
| 528 | prefix: str = "", |
| 529 | ) -> None: |
| 530 | super().__init__() |
| 531 | |
| 532 | assert video_attn_type == "spatial_temporal", ( |
| 533 | f'video_attn_type must be "spatial_temporal", got {video_attn_type}' |
| 534 | ) |
| 535 | self.video_attn_type = video_attn_type |
| 536 | qkv_hidden_size = block_cfg.get("qkv_hidden_size") or block_cfg["hidden_dim"] |
| 537 | self.rope_2d = Rope2DPosEmbRepeated( |
| 538 | qkv_hidden_size // block_cfg["num_heads"], 512, 512 |
| 539 | ) |
| 540 | self.blocks = nn.ModuleList( |
| 541 | [ |
| 542 | MoonViTEncoderLayer( |
| 543 | **block_cfg, |
| 544 | quant_config=quant_config, |
| 545 | prefix=f"{prefix}.blocks.{layer_idx}", |
| 546 | ) |
| 547 | for layer_idx in range(num_layers) |
| 548 | ] |
| 549 | ) |
| 550 | self.final_layernorm = _make_vision_norm( |
| 551 | block_cfg.get("norm_type", "layernorm"), hidden_dim |
| 552 | ) |
| 553 | |
| 554 | def prepare_encoder_metadata( |
| 555 | self, |
| 556 | grid_thw_list: list[list[int]], |
| 557 | *, |
| 558 | device: torch.device, |
| 559 | max_batch_size: int | None = None, |
| 560 | max_seqlen_override: int | None = None, |
| 561 | ) -> dict[str, torch.Tensor | None]: |
| 562 | metadata: dict[str, torch.Tensor | None] = {} |
| 563 | metadata["rope_freqs_cis"] = self.rope_2d.get_freqs_cis( |
| 564 | grid_thw_list, device=device |
| 565 | ) |
| 566 | |
| 567 | grid_thw_np = np.array(grid_thw_list, dtype=np.int32) |
| 568 | lengths = grid_thw_np[:, 0] * grid_thw_np[:, 1] * grid_thw_np[:, 2] |
| 569 | cu_seqlens = np.concatenate( |
| 570 | [np.zeros(1, dtype=np.int32), lengths.cumsum(dtype=np.int32)] |
| 571 | ) |
| 572 | if max_batch_size is not None: |
| 573 | num_seqs = len(cu_seqlens) - 1 |
| 574 | if num_seqs < max_batch_size: |
| 575 | cu_seqlens = np.concatenate( |
| 576 | [ |
| 577 | cu_seqlens, |
| 578 | np.full( |
| 579 | max_batch_size - num_seqs, |
| 580 | cu_seqlens[-1], |
| 581 | dtype=np.int32, |
| 582 | ), |
| 583 | ] |
| 584 | ) |
| 585 | |
| 586 | attn_backend = self.blocks[0].attn.attn_backend |
| 587 | metadata["sequence_lengths"] = MMEncoderAttention.maybe_compute_seq_lens( |
| 588 | attn_backend, cu_seqlens, device |
| 589 | ) |
| 590 | max_seqlen = ( |
| 591 | max_seqlen_override |
| 592 | if max_seqlen_override is not None |
| 593 | else MMEncoderAttention.compute_max_seqlen(attn_backend, cu_seqlens) |
| 594 | ) |
| 595 | metadata["max_seqlen"] = torch.tensor(max_seqlen, dtype=torch.int32) |
| 596 | metadata["cu_seqlens"] = MMEncoderAttention.maybe_recompute_cu_seqlens( |
| 597 | attn_backend, |
| 598 | cu_seqlens, |
| 599 | self.blocks[0].hidden_dim, |
| 600 | self.blocks[0].tp_size, |
| 601 | device, |
| 602 | ) |
| 603 | return metadata |
| 604 | |
| 605 | def forward( |
| 606 | self, |
| 607 | hidden_states: torch.Tensor, |
| 608 | grid_thws: torch.Tensor | list[list[int]] | None, |
| 609 | *, |
| 610 | encoder_metadata: dict[str, torch.Tensor | None] | None = None, |
| 611 | ) -> torch.Tensor: |
| 612 | if encoder_metadata is None: |
| 613 | assert grid_thws is not None |
| 614 | grid_thw_list = ( |
| 615 | grid_thws if isinstance(grid_thws, list) else grid_thws.tolist() |
| 616 | ) |
| 617 | encoder_metadata = self.prepare_encoder_metadata( |
| 618 | grid_thw_list, device=hidden_states.device |
| 619 | ) |
| 620 | |
| 621 | rope_freqs_cis = encoder_metadata["rope_freqs_cis"] |
| 622 | cu_seqlens = encoder_metadata["cu_seqlens"] |
| 623 | max_seqlen = encoder_metadata["max_seqlen"] |
| 624 | sequence_lengths = encoder_metadata.get("sequence_lengths") |
| 625 | assert rope_freqs_cis is not None |
| 626 | assert cu_seqlens is not None |
| 627 | assert max_seqlen is not None |
| 628 | |
| 629 | for block in self.blocks: |
| 630 | hidden_states = block( |
| 631 | hidden_states, |
| 632 | cu_seqlens, |
| 633 | rope_freqs_cis=rope_freqs_cis, |
| 634 | max_seqlen=max_seqlen, |
| 635 | sequence_lengths=sequence_lengths, |
| 636 | ) |
| 637 | |
| 638 | hidden_states = self.final_layernorm(hidden_states) |
| 639 | |
| 640 | return hidden_states |
| 641 | |
| 642 | |
| 643 | def tpool_patch_merger( |
| 644 | x: torch.Tensor, |
| 645 | grid_thws: torch.Tensor | list[list[int]], |
| 646 | merge_kernel_size: tuple[int, int] = (2, 2), |
| 647 | ) -> list[torch.Tensor]: |
| 648 | """Temporal pooling patch merger.""" |
| 649 | kh, kw = merge_kernel_size |
| 650 | grid_thw_list = grid_thws if isinstance(grid_thws, list) else grid_thws.tolist() |
| 651 | lengths = [t * h * w for t, h, w in grid_thw_list] |
| 652 | seqs = x.split(lengths, dim=0) |
| 653 | |
| 654 | outputs = [] |
| 655 | for seq, (t, h, w) in zip(seqs, grid_thw_list): |
| 656 | nh, nw = h // kh, w // kw |
| 657 | # Reshape: (t*h*w, d) -> (t, nh, kh, nw, kw, d) |
| 658 | v = seq.view(t, nh, kh, nw, kw, -1) |
| 659 | # Temporal pooling first (reduces tensor size before permute) |
| 660 | v = v.mean(dim=0) # (nh, kh, nw, kw, d) |
| 661 | # Spatial rearrangement: (nh, kh, nw, kw, d) -> (nh, nw, kh, kw, d) |
| 662 | out = v.permute(0, 2, 1, 3, 4).reshape(nh * nw, kh * kw, -1) |
| 663 | outputs.append(out) |
| 664 | |
| 665 | return outputs |
| 666 | |
| 667 | |
| 668 | def build_image_merge_gather_idx( |
| 669 | grid_thws: list[list[int]] | list[tuple[int, int, int]], |
| 670 | merge_kernel_size: tuple[int, int], |
| 671 | ) -> np.ndarray: |
| 672 | """Build packed spatial-merge indices for image-only CUDA graphs.""" |
| 673 | kh, kw = merge_kernel_size |
| 674 | parts: list[np.ndarray] = [] |
| 675 | offset = 0 |
| 676 | for t, h, w in grid_thws: |
| 677 | if t != 1: |
| 678 | raise ValueError("Image encoder CUDA graphs require grid T == 1") |
| 679 | idx = np.arange(h * w, dtype=np.int64).reshape(h, w) |
| 680 | idx = idx.reshape(h // kh, kh, w // kw, kw) |
| 681 | parts.append(idx.transpose(0, 2, 1, 3).reshape(-1, kh * kw) + offset) |
| 682 | offset += h * w |
| 683 | if not parts: |
| 684 | return np.empty((0, kh * kw), dtype=np.int64) |
| 685 | return np.concatenate(parts) |
| 686 | |
| 687 | |
| 688 | def tpool_patch_merger_packed( |
| 689 | x: torch.Tensor, |
| 690 | merge_gather_idx: torch.Tensor, |
| 691 | ) -> torch.Tensor: |
| 692 | """Apply the image-only spatial merge using precomputed tensor indices.""" |
| 693 | return x[merge_gather_idx] |
| 694 | |
| 695 | |
| 696 | class MoonViT3dPretrainedModel(nn.Module): |
| 697 | """Main vision tower model. |
| 698 | |
| 699 | Uses KimiK25VisionConfig directly from transformers_utils/configs/kimi_k25.py. |
| 700 | """ |
| 701 | |
| 702 | def __init__( |
| 703 | self, |
| 704 | config: KimiK25VisionConfig, |
| 705 | quant_config: QuantizationConfig | None = None, |
| 706 | prefix: str = "", |
| 707 | ): |
| 708 | super().__init__() |
| 709 | config = deepcopy(config) |
| 710 | self.config = config # Required for run_dp_sharded_mrope_vision_model |
| 711 | self.merge_kernel_size = config.merge_kernel_size |
| 712 | self.patch_size = config.patch_size |
| 713 | self.merge_type = config.merge_type |
| 714 | |
| 715 | self.patch_embed = MoonVision3dPatchEmbed( |
| 716 | out_dim=config.hidden_size, |
| 717 | patch_size=config.patch_size, |
| 718 | pos_emb_height=config.init_pos_emb_height, |
| 719 | pos_emb_width=config.init_pos_emb_width, |
| 720 | pos_emb_time=config.init_pos_emb_time, |
| 721 | pos_emb_type=config.pos_emb_type, |
| 722 | patch_embed_proj_bias=getattr(config, "patch_embed_proj_bias", True), |
| 723 | pos_emb_interpolation_mode=getattr( |
| 724 | config, "pos_emb_interpolation_mode", "bicubic" |
| 725 | ), |
| 726 | ) |
| 727 | |
| 728 | self.encoder = MoonViT3dEncoder( |
| 729 | hidden_dim=config.hidden_size, |
| 730 | num_layers=config.num_hidden_layers, |
| 731 | block_cfg={ |
| 732 | "num_heads": config.num_attention_heads, |
| 733 | "hidden_dim": config.hidden_size, |
| 734 | "qkv_hidden_size": getattr(config, "qkv_hidden_size", None), |
| 735 | "mlp_dim": config.intermediate_size, |
| 736 | "activation": get_act_fn( |
| 737 | getattr(config, "activation_func", "gelu_pytorch_tanh") |
| 738 | ), |
| 739 | "attn_bias": getattr(config, "attn_bias", True), |
| 740 | "norm_type": getattr(config, "norm_type", "layernorm"), |
| 741 | "mlp_type": getattr(config, "mlp_type", "mlp2"), |
| 742 | "linear_bias": getattr(config, "linear_bias", True), |
| 743 | }, |
| 744 | video_attn_type=config.video_attn_type, |
| 745 | quant_config=quant_config, |
| 746 | prefix=maybe_prefix(prefix, "encoder"), |
| 747 | ) |
| 748 | |
| 749 | def forward( |
| 750 | self, |
| 751 | pixel_values: torch.Tensor, |
| 752 | grid_thws: torch.Tensor | list[list[int]] | None, |
| 753 | *, |
| 754 | encoder_metadata: dict[str, torch.Tensor | None] | None = None, |
| 755 | ) -> torch.Tensor: |
| 756 | """Args: |
| 757 | pixel_values (torch.Tensor): The input pixel values. |
| 758 | grid_thws (torch.Tensor): Temporal, height and width. |
| 759 | |
| 760 | Returns: |
| 761 | torch.Tensor: The output tokens. |
| 762 | |
| 763 | """ |
| 764 | if encoder_metadata is not None and "pos_embeds" in encoder_metadata: |
| 765 | hidden_states = self.patch_embed( |
| 766 | pixel_values, |
| 767 | None, |
| 768 | pos_embeds=encoder_metadata["pos_embeds"], |
| 769 | ) |
| 770 | hidden_states = self.encoder( |
| 771 | hidden_states, |
| 772 | None, |
| 773 | encoder_metadata=encoder_metadata, |
| 774 | ) |
| 775 | merge_gather_idx = encoder_metadata["merge_gather_idx"] |
| 776 | assert merge_gather_idx is not None |
| 777 | return tpool_patch_merger_packed(hidden_states, merge_gather_idx) |
| 778 | |
| 779 | assert grid_thws is not None |
| 780 | grid_thw_list = grid_thws if isinstance(grid_thws, list) else grid_thws.tolist() |
| 781 | if encoder_metadata is None: |
| 782 | encoder_metadata = self.encoder.prepare_encoder_metadata( |
| 783 | grid_thw_list, device=pixel_values.device |
| 784 | ) |
| 785 | |
| 786 | hidden_states = self.patch_embed(pixel_values, grid_thw_list) |
| 787 | hidden_states = self.encoder( |
| 788 | hidden_states, |
| 789 | grid_thw_list, |
| 790 | encoder_metadata=encoder_metadata, |
| 791 | ) |
| 792 | if ( |
| 793 | self.merge_type == "sd2_tpool" |
| 794 | ): # spatial downsampling 2x with temporal pooling all |
| 795 | hidden_states = tpool_patch_merger( |
| 796 | hidden_states, grid_thw_list, merge_kernel_size=self.merge_kernel_size |
| 797 | ) |
| 798 | else: |
| 799 | raise NotImplementedError(f"Not support {self.merge_type}") |
| 800 | |
| 801 | return hidden_states |
| 802 | |
| 803 | def prepare_encoder_cudagraph_metadata( |
| 804 | self, |
| 805 | grid_thw_list: list[list[int]], |
| 806 | *, |
| 807 | max_batch_size: int | None, |
| 808 | max_seqlen_override: int | None = None, |
| 809 | device: torch.device, |
| 810 | ) -> dict[str, torch.Tensor | None]: |
| 811 | """Precompute fixed-buffer metadata for image encoder CUDA graphs.""" |
| 812 | grid_thw_list = [list(map(int, grid)) for grid in grid_thw_list] |
| 813 | metadata = self.encoder.prepare_encoder_metadata( |
| 814 | grid_thw_list, |
| 815 | device=device, |
| 816 | max_batch_size=max_batch_size, |
| 817 | max_seqlen_override=max_seqlen_override, |
| 818 | ) |
| 819 | metadata["pos_embeds"] = self.patch_embed.pos_emb.get_pos_embeds( |
| 820 | grid_thw_list |
| 821 | ).to(device=device) |
| 822 | merge_gather_idx = build_image_merge_gather_idx( |
| 823 | grid_thw_list, self.merge_kernel_size |
| 824 | ) |
| 825 | metadata["merge_gather_idx"] = async_tensor_h2d(merge_gather_idx, device=device) |
| 826 | return metadata |
| 827 | |
| 828 | |
| 829 | @torch.inference_mode() |
| 830 | def mm_projector_forward(mm_projector: torch.nn.Module, vt_output: list[torch.Tensor]): |
| 831 | """Apply MM projector to vision tower outputs.""" |
| 832 | num_embedding_list = [x.shape[0] for x in vt_output] |
| 833 | batched = torch.cat(vt_output, dim=0) |
| 834 | projector_dtype = next(mm_projector.parameters()).dtype |
| 835 | if batched.dtype != projector_dtype: |
| 836 | batched = batched.to(projector_dtype) |
| 837 | proj_out = mm_projector(batched) |
| 838 | proj_out = proj_out.reshape(-1, proj_out.shape[-1]) |
| 839 | proj_out = torch.split(proj_out, num_embedding_list) |
| 840 | return proj_out |
| 841 | |
| 842 | |
| 843 | @torch.inference_mode() |
| 844 | def vision_tower_forward( |
| 845 | vision_tower: Any, |
| 846 | pixel_values: torch.Tensor, |
| 847 | grid_thw: torch.Tensor, |
| 848 | mm_projector: Any, |
| 849 | use_data_parallel: bool, |
| 850 | ) -> list[torch.Tensor]: |
| 851 | """DP-sharded vision tower forward with mrope. |
| 852 | |
| 853 | Uses vLLM's standard data parallelism utility to shard the batch |
| 854 | across available GPUs, enabling parallel processing of vision features. |
| 855 | """ |
| 856 | if use_data_parallel: |
| 857 | grid_thw_list = grid_thw.tolist() |
| 858 | vt_outputs = run_dp_sharded_mrope_vision_model( |
| 859 | vision_model=vision_tower, |
| 860 | pixel_values=pixel_values, |
| 861 | grid_thw_list=grid_thw_list, |
| 862 | rope_type="rope_2d", |
| 863 | ) |
| 864 | else: |
| 865 | grid_thw_list = grid_thw.tolist() |
| 866 | encoder_metadata = vision_tower.encoder.prepare_encoder_metadata( |
| 867 | grid_thw_list, device=pixel_values.device |
| 868 | ) |
| 869 | vt_outputs = vision_tower( |
| 870 | pixel_values, |
| 871 | grid_thw_list, |
| 872 | encoder_metadata=encoder_metadata, |
| 873 | ) |
| 874 | tensors = mm_projector_forward(mm_projector, list(vt_outputs)) |
| 875 | return list(tensors) |
| 876 | |
| 877 | |
| 878 | class KimiK25MultiModalProjector(nn.Module): |
| 879 | """Multi-modal projector with patch merging for Kimi-K2.5.""" |
| 880 | |
| 881 | def __init__( |
| 882 | self, |
| 883 | config: KimiK25VisionConfig, |
| 884 | use_data_parallel: bool = False, |
| 885 | quant_config: QuantizationConfig | None = None, |
| 886 | prefix: str = "", |
| 887 | ): |
| 888 | super().__init__() |
| 889 | self.use_data_parallel = use_data_parallel |
| 890 | self.mm_projector_type = getattr(config, "mm_projector_type", "patchmerger") |
| 891 | |
| 892 | # Hidden size after patch merging |
| 893 | merge_h, merge_w = config.merge_kernel_size |
| 894 | self.hidden_size = config.hidden_size * merge_h * merge_w |
| 895 | |
| 896 | if self.mm_projector_type == "patchmergerv2": |
| 897 | self.linear_1 = ReplicatedLinear( |
| 898 | self.hidden_size, |
| 899 | self.hidden_size, |
| 900 | bias=False, |
| 901 | quant_config=quant_config, |
| 902 | prefix=f"{prefix}.linear_1", |
| 903 | ) |
| 904 | self.linear_2 = ReplicatedLinear( |
| 905 | self.hidden_size, |
| 906 | getattr(config, "text_hidden_size", config.mm_hidden_size), |
| 907 | bias=False, |
| 908 | quant_config=quant_config, |
| 909 | prefix=f"{prefix}.linear_2", |
| 910 | ) |
| 911 | self.post_norm = torch.nn.RMSNorm( |
| 912 | getattr(config, "text_hidden_size", config.mm_hidden_size), |
| 913 | eps=config.projector_ln_eps, |
| 914 | ) |
| 915 | self.act = GELUActivation() |
| 916 | return |
| 917 | |
| 918 | self.pre_norm = torch.nn.LayerNorm(config.hidden_size, eps=1e-5) |
| 919 | self.linear_1 = ReplicatedLinear( |
| 920 | self.hidden_size, |
| 921 | self.hidden_size, |
| 922 | bias=True, |
| 923 | quant_config=quant_config, |
| 924 | prefix=f"{prefix}.linear_1", |
| 925 | ) |
| 926 | self.linear_2 = ReplicatedLinear( |
| 927 | self.hidden_size, |
| 928 | config.mm_hidden_size, |
| 929 | bias=True, |
| 930 | quant_config=quant_config, |
| 931 | prefix=f"{prefix}.linear_2", |
| 932 | ) |
| 933 | self.act = GELUActivation() |
| 934 | |
| 935 | def forward(self, image_features: torch.Tensor) -> torch.Tensor: |
| 936 | if self.mm_projector_type == "patchmergerv2": |
| 937 | hidden_states = image_features.view(image_features.shape[0], -1) |
| 938 | hidden_states, _ = self.linear_1(hidden_states) |
| 939 | hidden_states = self.act(hidden_states) |
| 940 | hidden_states, _ = self.linear_2(hidden_states) |
| 941 | return self.post_norm(hidden_states) |
| 942 | |
| 943 | hidden_states = self.pre_norm(image_features).view(-1, self.hidden_size) |
| 944 | hidden_states, _ = self.linear_1(hidden_states) |
| 945 | hidden_states = self.act(hidden_states) |
| 946 | hidden_states, _ = self.linear_2(hidden_states) |
| 947 | return hidden_states |