| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Train-consistent video preprocessing for Dots3Note.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | import hashlib |
| 8 | import io |
| 9 | import math |
| 10 | import os |
| 11 | import random |
| 12 | from dataclasses import dataclass |
| 13 | from typing import Literal |
| 14 | |
| 15 | import numpy as np |
| 16 | from PIL import Image |
| 17 | |
| 18 | _ALIGN = 28 |
| 19 | _MIN_FRAMES = 4 |
| 20 | _PF_FLOOR = 128 |
| 21 | _PF_CEIL = 1024 |
| 22 | _FPS_CAP = 1.0 |
| 23 | _FPS_MIN = 0.2 |
| 24 | _FRAME_OVERHEAD = 15 |
| 25 | _BUDGET_OVERHEAD = 2240 |
| 26 | _INTERLEAVE_MIN_SECONDS = 1.0 |
| 27 | _AUDIO_SAMPLE_RATE = 16000 |
| 28 | _AUDIO_SAMPLES_PER_TOKEN = 1280 |
| 29 | _AUDIO_CHUNK_SECONDS = 30 |
| 30 | |
| 31 | |
| 32 | @dataclass(frozen=True) |
| 33 | class Dots3NoteVideoPart: |
| 34 | kind: Literal["text", "image", "audio"] |
| 35 | value: str | Image.Image | np.ndarray |
| 36 | |
| 37 | |
| 38 | def _token_len(tokenizer, text: str) -> int: |
| 39 | if not text: |
| 40 | return 0 |
| 41 | if hasattr(tokenizer, "encode"): |
| 42 | return len(tokenizer.encode(text, add_special_tokens=False)) |
| 43 | return len(tokenizer(text, add_special_tokens=False)["input_ids"]) |
| 44 | |
| 45 | |
| 46 | def _compute_target_size( |
| 47 | orig_h: int, |
| 48 | orig_w: int, |
| 49 | min_pixels: int, |
| 50 | max_pixels: int, |
| 51 | ) -> tuple[int, int]: |
| 52 | height = max(_ALIGN, round(orig_h / _ALIGN) * _ALIGN) |
| 53 | width = max(_ALIGN, round(orig_w / _ALIGN) * _ALIGN) |
| 54 | if height * width > max_pixels: |
| 55 | beta = math.sqrt(orig_h * orig_w / max_pixels) |
| 56 | height = max(_ALIGN, math.floor(orig_h / beta / _ALIGN) * _ALIGN) |
| 57 | width = max(_ALIGN, math.floor(orig_w / beta / _ALIGN) * _ALIGN) |
| 58 | elif height * width < min_pixels: |
| 59 | beta = math.sqrt(min_pixels / max(1, orig_h * orig_w)) |
| 60 | height = math.ceil(orig_h * beta / _ALIGN) * _ALIGN |
| 61 | width = math.ceil(orig_w * beta / _ALIGN) * _ALIGN |
| 62 | if height * width > max_pixels: |
| 63 | beta = math.sqrt(height * width / max_pixels) |
| 64 | height = max(_ALIGN, math.floor(height / beta / _ALIGN) * _ALIGN) |
| 65 | width = max(_ALIGN, math.floor(width / beta / _ALIGN) * _ALIGN) |
| 66 | return int(height), int(width) |
| 67 | |
| 68 | |
| 69 | def _real_patches_at( |
| 70 | orig_h: int, |
| 71 | orig_w: int, |
| 72 | patch_cap: int, |
| 73 | ) -> int: |
| 74 | target_h, target_w = _compute_target_size( |
| 75 | orig_h, |
| 76 | orig_w, |
| 77 | _PF_FLOOR * _ALIGN * _ALIGN, |
| 78 | max(_PF_FLOOR, patch_cap) * _ALIGN * _ALIGN, |
| 79 | ) |
| 80 | return (target_h // _ALIGN) * (target_w // _ALIGN) |
| 81 | |
| 82 | |
| 83 | def _frame_hard_cap(seq_length: int) -> int: |
| 84 | need = max(1, (seq_length - _BUDGET_OVERHEAD) // (_PF_FLOOR + 15)) |
| 85 | if need <= 1024: |
| 86 | return 1024 |
| 87 | cap = 1 |
| 88 | while cap < need: |
| 89 | cap <<= 1 |
| 90 | return cap |
| 91 | |
| 92 | |
| 93 | def _solve_degrade( |
| 94 | visual_budget: int, |
| 95 | duration: float, |
| 96 | orig_h: int, |
| 97 | orig_w: int, |
| 98 | orig_fps: float, |
| 99 | seq_length: int, |
| 100 | ) -> tuple[int, int]: |
| 101 | aligned_h = max(_ALIGN, round(orig_h / _ALIGN) * _ALIGN) |
| 102 | aligned_w = max(_ALIGN, round(orig_w / _ALIGN) * _ALIGN) |
| 103 | orig_max_pf = (aligned_h // _ALIGN) * (aligned_w // _ALIGN) |
| 104 | fps_cap = min(_FPS_CAP, max(orig_fps, 1e-6)) |
| 105 | pf_cap = min(_PF_CEIL, max(orig_max_pf, _PF_FLOOR)) |
| 106 | frame_cap = _frame_hard_cap(seq_length) |
| 107 | |
| 108 | def usage(scale: float) -> tuple[int, float, int, int]: |
| 109 | fps = _FPS_MIN + scale * (fps_cap - _FPS_MIN) |
| 110 | patch_cap = _PF_FLOOR + scale * (pf_cap - _PF_FLOOR) |
| 111 | num_frames = max( |
| 112 | _MIN_FRAMES, |
| 113 | min(int(round(duration * fps)), frame_cap), |
| 114 | ) |
| 115 | patches = _real_patches_at( |
| 116 | orig_h, |
| 117 | orig_w, |
| 118 | int(round(patch_cap)), |
| 119 | ) |
| 120 | return ( |
| 121 | num_frames * (patches + _FRAME_OVERHEAD), |
| 122 | fps, |
| 123 | int(round(patch_cap)), |
| 124 | num_frames, |
| 125 | ) |
| 126 | |
| 127 | if usage(1.0)[0] <= visual_budget: |
| 128 | _, _, patch_cap, num_frames = usage(1.0) |
| 129 | return num_frames, patch_cap |
| 130 | |
| 131 | floor_cost = _real_patches_at(orig_h, orig_w, _PF_FLOOR) + _FRAME_OVERHEAD |
| 132 | if usage(0.0)[0] > visual_budget: |
| 133 | return max( |
| 134 | _MIN_FRAMES, |
| 135 | min(visual_budget // floor_cost, frame_cap), |
| 136 | ), _PF_FLOOR |
| 137 | |
| 138 | low, high = 0.0, 1.0 |
| 139 | for _ in range(50): |
| 140 | mid = (low + high) / 2 |
| 141 | if usage(mid)[0] <= visual_budget: |
| 142 | low = mid |
| 143 | else: |
| 144 | high = mid |
| 145 | _, _, patch_cap, num_frames = usage(low) |
| 146 | return num_frames, patch_cap |
| 147 | |
| 148 | |
| 149 | def _audio_tokens(duration: float, sample_rate: int) -> int: |
| 150 | if duration <= 0: |
| 151 | return 0 |
| 152 | total_samples = int(duration * sample_rate) |
| 153 | chunk_samples = _AUDIO_CHUNK_SECONDS * sample_rate |
| 154 | num_tokens = 0 |
| 155 | position = 0 |
| 156 | while position < total_samples: |
| 157 | length = min(chunk_samples, total_samples - position) |
| 158 | num_tokens += math.ceil(length / _AUDIO_SAMPLES_PER_TOKEN) |
| 159 | position += chunk_samples |
| 160 | return num_tokens + 2 |
| 161 | |
| 162 | |
| 163 | def _decode_audio( |
| 164 | video_bytes: bytes, |
| 165 | sample_rate: int, |
| 166 | ) -> tuple[np.ndarray | None, float]: |
| 167 | from torchcodec.decoders import AudioDecoder |
| 168 | |
| 169 | try: |
| 170 | decoder = AudioDecoder( |
| 171 | video_bytes, |
| 172 | sample_rate=sample_rate, |
| 173 | num_channels=1, |
| 174 | ) |
| 175 | samples = decoder.get_all_samples() |
| 176 | except Exception: |
| 177 | return None, 0.0 |
| 178 | |
| 179 | waveform = samples.data |
| 180 | if waveform is None or waveform.numel() == 0: |
| 181 | return None, 0.0 |
| 182 | if waveform.ndim == 2: |
| 183 | waveform = waveform.mean(dim=0) if waveform.shape[0] > 1 else waveform[0] |
| 184 | array = waveform.cpu().numpy() |
| 185 | pcm = (np.clip(array, -1.0, 1.0) * 32767.0).astype(np.int16) |
| 186 | return pcm, float(pcm.shape[0]) / sample_rate |
| 187 | |
| 188 | |
| 189 | def _open_video(video_bytes: bytes): |
| 190 | from torchcodec.decoders import VideoDecoder |
| 191 | |
| 192 | try: |
| 193 | return VideoDecoder( |
| 194 | video_bytes, |
| 195 | dimension_order="NHWC", |
| 196 | num_ffmpeg_threads=1, |
| 197 | seek_mode="approximate", |
| 198 | ) |
| 199 | except TypeError: |
| 200 | return VideoDecoder( |
| 201 | video_bytes, |
| 202 | dimension_order="NHWC", |
| 203 | num_ffmpeg_threads=1, |
| 204 | ) |
| 205 | |
| 206 | |
| 207 | def _jpeg_roundtrip(image: Image.Image, quality: int) -> Image.Image: |
| 208 | buffer = io.BytesIO() |
| 209 | image.save(buffer, format="JPEG", quality=quality) |
| 210 | buffer.seek(0) |
| 211 | with Image.open(buffer) as decoded: |
| 212 | return decoded.convert("RGB").copy() |
| 213 | |
| 214 | |
| 215 | def _decode_frames( |
| 216 | video_bytes: bytes, |
| 217 | visual_budget: int, |
| 218 | seq_length: int, |
| 219 | jpeg_quality: int, |
| 220 | ) -> tuple[list[tuple[float, Image.Image]], float]: |
| 221 | decoder = _open_video(video_bytes) |
| 222 | metadata = decoder.metadata |
| 223 | duration = float(metadata.duration_seconds or 0) |
| 224 | orig_h = int(metadata.height) |
| 225 | orig_w = int(metadata.width) |
| 226 | total_frames = int(getattr(metadata, "num_frames", 0) or 0) |
| 227 | orig_fps = float(getattr(metadata, "average_fps", 0) or 0) or 25.0 |
| 228 | if duration <= 0 or orig_h <= 0 or orig_w <= 0: |
| 229 | raise ValueError( |
| 230 | "Invalid video metadata: " |
| 231 | f"duration={duration}, height={orig_h}, width={orig_w}" |
| 232 | ) |
| 233 | if total_frames <= 0: |
| 234 | total_frames = max(1, int(duration * orig_fps)) |
| 235 | |
| 236 | num_frames, patch_cap = _solve_degrade( |
| 237 | visual_budget, |
| 238 | duration, |
| 239 | orig_h, |
| 240 | orig_w, |
| 241 | orig_fps, |
| 242 | seq_length, |
| 243 | ) |
| 244 | aligned_h = max(_ALIGN, round(orig_h / _ALIGN) * _ALIGN) |
| 245 | aligned_w = max(_ALIGN, round(orig_w / _ALIGN) * _ALIGN) |
| 246 | orig_max_pf = (aligned_h // _ALIGN) * (aligned_w // _ALIGN) |
| 247 | max_pixels = min(patch_cap, orig_max_pf) * _ALIGN * _ALIGN |
| 248 | target_h, target_w = _compute_target_size( |
| 249 | orig_h, |
| 250 | orig_w, |
| 251 | _PF_FLOOR * _ALIGN * _ALIGN, |
| 252 | max_pixels, |
| 253 | ) |
| 254 | num_frames = max(_MIN_FRAMES, min(num_frames, total_frames)) |
| 255 | if num_frames == 1: |
| 256 | indices = [0] |
| 257 | else: |
| 258 | step = (total_frames - 1) / (num_frames - 1) |
| 259 | indices = [int(round(i * step)) for i in range(num_frames)] |
| 260 | indices = sorted({max(0, min(index, total_frames - 1)) for index in indices}) |
| 261 | |
| 262 | try: |
| 263 | decoded = decoder.get_frames_at(indices=indices).data |
| 264 | except (IndexError, RuntimeError): |
| 265 | safe = list(indices) |
| 266 | while safe and safe[-1] > 0: |
| 267 | safe.pop() |
| 268 | try: |
| 269 | decoded = decoder.get_frames_at(indices=safe).data |
| 270 | indices = safe |
| 271 | break |
| 272 | except (IndexError, RuntimeError): |
| 273 | continue |
| 274 | else: |
| 275 | raise |
| 276 | |
| 277 | frames: list[tuple[float, Image.Image]] = [] |
| 278 | for index, frame in zip(indices, decoded): |
| 279 | array = frame.cpu().numpy() if hasattr(frame, "cpu") else np.asarray(frame) |
| 280 | image = Image.fromarray(array) |
| 281 | if image.size != (target_w, target_h): |
| 282 | image = image.resize((target_w, target_h), Image.Resampling.BICUBIC) |
| 283 | frames.append( |
| 284 | (round(index / orig_fps, 3), _jpeg_roundtrip(image, jpeg_quality)) |
| 285 | ) |
| 286 | return frames, duration |
| 287 | |
| 288 | |
| 289 | def _decoded_frames(video) -> tuple[np.ndarray, float]: |
| 290 | metadata = None |
| 291 | if isinstance(video, tuple): |
| 292 | video, metadata = video |
| 293 | frames = np.asarray(video) |
| 294 | if frames.ndim != 4 or frames.shape[-1] not in (3, 4): |
| 295 | raise TypeError( |
| 296 | "Decoded NOTE video must have shape (frames, height, width, channels)" |
| 297 | ) |
| 298 | fps = float((metadata or {}).get("fps", 1.0)) if metadata else 1.0 |
| 299 | return frames, max(fps, 1e-6) |
| 300 | |
| 301 | |
| 302 | def _prepare_decoded_frames( |
| 303 | video, |
| 304 | visual_budget: int, |
| 305 | seq_length: int, |
| 306 | jpeg_quality: int, |
| 307 | ) -> tuple[list[tuple[float, Image.Image]], float]: |
| 308 | frames, fps = _decoded_frames(video) |
| 309 | duration = len(frames) / fps |
| 310 | orig_h, orig_w = frames.shape[1:3] |
| 311 | num_frames, patch_cap = _solve_degrade( |
| 312 | visual_budget, |
| 313 | duration, |
| 314 | orig_h, |
| 315 | orig_w, |
| 316 | fps, |
| 317 | seq_length, |
| 318 | ) |
| 319 | num_frames = min(max(1, num_frames), len(frames)) |
| 320 | indices = np.linspace(0, len(frames) - 1, num_frames).round().astype(int) |
| 321 | target_h, target_w = _compute_target_size( |
| 322 | orig_h, |
| 323 | orig_w, |
| 324 | _PF_FLOOR * _ALIGN * _ALIGN, |
| 325 | patch_cap * _ALIGN * _ALIGN, |
| 326 | ) |
| 327 | output = [] |
| 328 | for index in sorted(set(indices.tolist())): |
| 329 | image = Image.fromarray(frames[index]).convert("RGB") |
| 330 | if image.size != (target_w, target_h): |
| 331 | image = image.resize((target_w, target_h), Image.Resampling.BICUBIC) |
| 332 | output.append((round(index / fps, 3), _jpeg_roundtrip(image, jpeg_quality))) |
| 333 | return output, duration |
| 334 | |
| 335 | |
| 336 | def _format_timestamp(seconds: float) -> str: |
| 337 | total_centiseconds = int(round(max(seconds, 0.0) * 100)) |
| 338 | hours = total_centiseconds // 360000 |
| 339 | minutes = (total_centiseconds // 6000) % 60 |
| 340 | secs = (total_centiseconds // 100) % 60 |
| 341 | centiseconds = total_centiseconds % 100 |
| 342 | return f"{hours:02d}:{minutes:02d}:{secs:02d}.{centiseconds:02d}" |
| 343 | |
| 344 | |
| 345 | def _group_bounds( |
| 346 | num_frames: int, |
| 347 | duration: float, |
| 348 | mode: str, |
| 349 | rng: random.Random, |
| 350 | ) -> list[int]: |
| 351 | if num_frames <= 1 or duration <= 0: |
| 352 | return [0, num_frames] |
| 353 | max_groups = min( |
| 354 | num_frames, |
| 355 | max(1, int(duration // _INTERLEAVE_MIN_SECONDS)), |
| 356 | ) |
| 357 | if mode == "whole" or max_groups <= 1: |
| 358 | groups = 1 |
| 359 | elif mode == "eval30": |
| 360 | groups = round(math.sqrt(max_groups)) |
| 361 | elif mode == "eval_ek": |
| 362 | groups = round((max_groups - 1) / math.log(max_groups)) |
| 363 | elif mode == "logk": |
| 364 | groups = round(math.exp(rng.uniform(0.0, math.log(max_groups)))) |
| 365 | else: |
| 366 | raise ValueError(f"Unsupported video k_mode: {mode}") |
| 367 | groups = max(1, min(max_groups, groups)) |
| 368 | if groups == 1: |
| 369 | return [0, num_frames] |
| 370 | if mode == "logk": |
| 371 | cuts = sorted(rng.sample(range(1, num_frames), groups - 1)) |
| 372 | else: |
| 373 | cuts = sorted( |
| 374 | { |
| 375 | round(index * num_frames / groups) |
| 376 | for index in range(1, groups) |
| 377 | if 0 < round(index * num_frames / groups) < num_frames |
| 378 | } |
| 379 | ) |
| 380 | return [0, *cuts, num_frames] |
| 381 | |
| 382 | |
| 383 | def preprocess_dots3_note_video( |
| 384 | video, |
| 385 | *, |
| 386 | tokenizer, |
| 387 | question: str, |
| 388 | seq: int, |
| 389 | output_reserve: int | None = None, |
| 390 | audio_cap: float = 1.0, |
| 391 | audio_sample_rate: int = _AUDIO_SAMPLE_RATE, |
| 392 | k_mode: str = "eval_ek", |
| 393 | max_new_tokens: int = 0, |
| 394 | ) -> list[Dots3NoteVideoPart]: |
| 395 | """Expand one video into timestamped, interleaved image/audio parts.""" |
| 396 | if seq <= 0: |
| 397 | raise ValueError(f"seq must be positive, got {seq}") |
| 398 | configured_reserve = seq // 4 if output_reserve is None else output_reserve |
| 399 | effective_reserve = max(configured_reserve, max_new_tokens) |
| 400 | if effective_reserve >= seq: |
| 401 | raise ValueError( |
| 402 | "output_reserve/max_new_tokens must leave room for video input" |
| 403 | ) |
| 404 | if audio_cap < 0: |
| 405 | raise ValueError(f"audio_cap must be non-negative, got {audio_cap}") |
| 406 | if audio_sample_rate <= 0: |
| 407 | raise ValueError(f"audio_sample_rate must be positive, got {audio_sample_rate}") |
| 408 | |
| 409 | seq_length = seq - effective_reserve |
| 410 | jpeg_quality = int(os.environ.get("XHS_VIDEO_JPEG_QUALITY", "85")) |
| 411 | pcm: np.ndarray | None = None |
| 412 | audio_duration = 0.0 |
| 413 | if isinstance(video, bytes) and audio_cap > 0: |
| 414 | pcm, audio_duration = _decode_audio(video, audio_sample_rate) |
| 415 | |
| 416 | audio_token_count = ( |
| 417 | _audio_tokens(audio_duration, audio_sample_rate) if pcm is not None else 0 |
| 418 | ) |
| 419 | estimated_frames = max(1, int(audio_duration * _FPS_CAP)) |
| 420 | max_groups = min( |
| 421 | estimated_frames, |
| 422 | max(1, int(audio_duration // _INTERLEAVE_MIN_SECONDS)), |
| 423 | ) |
| 424 | reserved_audio_tokens = audio_token_count + 3 * max_groups |
| 425 | min_visual_tokens = _MIN_FRAMES * (_PF_FLOOR + _FRAME_OVERHEAD) |
| 426 | if ( |
| 427 | audio_token_count > audio_cap * seq_length |
| 428 | or reserved_audio_tokens + min_visual_tokens + _BUDGET_OVERHEAD > seq_length |
| 429 | ): |
| 430 | pcm = None |
| 431 | audio_duration = 0.0 |
| 432 | reserved_audio_tokens = 0 |
| 433 | |
| 434 | overhead = ( |
| 435 | _token_len( |
| 436 | tokenizer, |
| 437 | "<|system|>You are a helpful assistant.<|endofsystem|>\n", |
| 438 | ) |
| 439 | + 2 |
| 440 | + _token_len(tokenizer, "<video_0>") |
| 441 | + 64 |
| 442 | ) |
| 443 | visual_budget = max( |
| 444 | _PF_FLOOR + _FRAME_OVERHEAD, |
| 445 | seq_length - overhead - reserved_audio_tokens, |
| 446 | ) |
| 447 | if isinstance(video, bytes): |
| 448 | frames, video_duration = _decode_frames( |
| 449 | video, |
| 450 | visual_budget, |
| 451 | seq_length, |
| 452 | jpeg_quality, |
| 453 | ) |
| 454 | else: |
| 455 | frames, video_duration = _prepare_decoded_frames( |
| 456 | video, |
| 457 | visual_budget, |
| 458 | seq_length, |
| 459 | jpeg_quality, |
| 460 | ) |
| 461 | |
| 462 | if pcm is None: |
| 463 | parts: list[Dots3NoteVideoPart] = [] |
| 464 | for timestamp, image in frames: |
| 465 | parts.append( |
| 466 | Dots3NoteVideoPart("text", f"<{_format_timestamp(timestamp)}>") |
| 467 | ) |
| 468 | parts.append(Dots3NoteVideoPart("image", image)) |
| 469 | return parts |
| 470 | |
| 471 | video_id = hashlib.sha1(video).hexdigest() |
| 472 | record_key = hashlib.sha1(f"{video_id}|{question}".encode()).hexdigest() |
| 473 | seed_hex = hashlib.sha1(f"42|flatten|{record_key}".encode()).hexdigest() |
| 474 | rng = random.Random(int(seed_hex[:8], 16)) |
| 475 | bounds = _group_bounds(len(frames), audio_duration, k_mode, rng) |
| 476 | parts = [] |
| 477 | for group in range(len(bounds) - 1): |
| 478 | start, end = bounds[group], bounds[group + 1] |
| 479 | if end <= start: |
| 480 | continue |
| 481 | start_time = 0.0 if group == 0 else frames[start][0] |
| 482 | end_time = audio_duration if group == len(bounds) - 2 else frames[end][0] |
| 483 | if end_time <= start_time: |
| 484 | end_time = start_time + audio_duration / max(1, len(bounds) - 1) |
| 485 | for timestamp, image in frames[start:end]: |
| 486 | parts.append( |
| 487 | Dots3NoteVideoPart("text", f"<{_format_timestamp(timestamp)}>") |
| 488 | ) |
| 489 | parts.append(Dots3NoteVideoPart("image", image)) |
| 490 | sample_start = max(0, int(round(start_time * audio_sample_rate))) |
| 491 | sample_end = min(len(pcm), int(round(end_time * audio_sample_rate))) |
| 492 | if sample_end > sample_start: |
| 493 | segment = np.ascontiguousarray( |
| 494 | pcm[sample_start:sample_end].astype(np.float32) / 32768.0 |
| 495 | ) |
| 496 | parts.append(Dots3NoteVideoPart("audio", segment)) |
| 497 | return parts |