"""Unified HF processor for the WeMM Video 1.5 Qwen3.5 audio checkpoints.""" import copy import json import shutil from pathlib import Path import torch from transformers import BatchFeature, ProcessorMixin, Qwen3VLProcessor, WhisperFeatureExtractor from transformers.utils.hub import cached_file from transformers.video_utils import VideoMetadata from .wemm_media import AUDIO, IMAGE, VIDEO, audio_features, audio_ids, decode_audio from .wemm_media import decode_video, load_image, resolve_media, tokens_per_segment, video_ids from .wemm_media import audio_timestamp_groups, resolve_timestamp_interval, timestamp_interval_option from .wemm_video_policy import resolve_video_options, video_environment_defaults, video_resize_size # Transformers 5.2 must see these dependencies directly when caching remote # processor code; an import through wemm_media alone misses the nested file. from .wemm_zip import load_zip_video as _zip_decoder_dependency from .wemm_zip_policy import positive_frame_count class WeMMProcessor(ProcessorMixin): """Single-conversation media + question API, including ordered chat messages. ``processor(question='描述视频', videos='clip.mp4', use_audio_in_video=True)`` returns a BatchFeature directly accepted by ``model.generate(**inputs)``. Lists of media belong to the same conversation, not a batch of requests. """ image_processor_class = 'AutoImageProcessor' video_processor_class = 'AutoVideoProcessor' tokenizer_class = 'AutoTokenizer' feature_extractor_class = 'WhisperFeatureExtractor' @classmethod def get_attributes(cls): return ['image_processor', 'tokenizer', 'video_processor', 'feature_extractor'] def __init__(self, image_processor, tokenizer, video_processor, feature_extractor, audio_config, fps=2.0, min_frames=4, max_frames=128, total_pixels=None, min_pixels=4096, system_prompt='You are a helpful assistant.', chat_template=None, timestamp_interval_sec=1.0, max_allowed_frames=None, max_allowed_total_pixels=None): super().__init__(image_processor=image_processor, tokenizer=tokenizer, video_processor=video_processor, feature_extractor=feature_extractor, chat_template=chat_template or tokenizer.chat_template) self.audio_config = copy.deepcopy(audio_config) self.fps, self.min_frames, self.max_frames = fps, min_frames, max_frames self.total_pixels = total_pixels or video_processor.size['longest_edge'] self.min_pixels, self.system_prompt = min_pixels, system_prompt self.timestamp_interval_sec = resolve_timestamp_interval(timestamp_interval_sec) self.max_allowed_frames = max_allowed_frames self.max_allowed_total_pixels = max_allowed_total_pixels self._native = Qwen3VLProcessor(image_processor=image_processor, tokenizer=tokenizer, video_processor=video_processor) @classmethod def from_pretrained(cls, pretrained_model_name_or_path, **kwargs): names = ('fps', 'min_frames', 'max_frames', 'total_pixels', 'min_pixels', 'system_prompt', 'timestamp_interval_sec', 'max_allowed_frames', 'max_allowed_total_pixels') overrides = {key: kwargs.pop(key) for key in names if key in kwargs} if 'TIMESTAMP_INTERVAL_SEC' in kwargs: raise ValueError('Use lowercase timestamp_interval_sec in processor arguments') trust_remote_code = kwargs.pop('trust_remote_code', True) kwargs.pop('_from_auto', None) hub = {key: kwargs[key] for key in ('cache_dir', 'force_download', 'local_files_only', 'token', 'revision', 'subfolder') if key in kwargs} proc_path = cached_file(pretrained_model_name_or_path, 'processor_config.json', _raise_exceptions_for_missing_entries=False, **hub) processor_config = json.loads(Path(proc_path).read_text()) if proc_path else {} config_path = cached_file(pretrained_model_name_or_path, 'config.json', _raise_exceptions_for_missing_entries=False, **hub) config = json.loads(Path(config_path).read_text()) if config_path else processor_config defaults = processor_config.get('wemm_defaults', {}) native = Qwen3VLProcessor.from_pretrained( pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs) audio = config['audio_config'] extractor = WhisperFeatureExtractor(feature_size=audio['num_mel_bins'], sampling_rate=16000, hop_length=160, chunk_length=30) environment = video_environment_defaults() effective = resolve_video_options({"total_pixels": native.video_processor.size['longest_edge'], **defaults, **environment}, overrides) parameters = {**defaults, **overrides, **{key: effective[key] for key in names if key in effective}} return cls(native.image_processor, native.tokenizer, native.video_processor, extractor, audio, chat_template=native.chat_template, **parameters) def save_pretrained(self, save_directory, **kwargs): """Save processor metadata/code beside a config containing audio_config.""" destination = Path(save_directory) destination.mkdir(parents=True, exist_ok=True) self._native.save_pretrained(destination, **kwargs) # Transformers 5.2 stores nested components in processor_config.json. # Also write standalone component configs before our custom metadata. self.image_processor.save_pretrained(destination) self.video_processor.save_pretrained(destination) defaults = {key: getattr(self, key) for key in ('fps', 'min_frames', 'max_frames', 'total_pixels', 'min_pixels', 'system_prompt', 'timestamp_interval_sec', 'max_allowed_frames', 'max_allowed_total_pixels')} mapping = {'AutoProcessor': 'processing_wemm.WeMMProcessor'} (destination / 'processor_config.json').write_text(json.dumps({ 'processor_class': 'WeMMProcessor', 'auto_map': mapping, 'wemm_defaults': defaults, 'audio_config': self.audio_config}, indent=2) + '\n') for name in ('processing_wemm.py', 'wemm_media.py', 'wemm_video_policy.py', 'wemm_zip.py', 'wemm_zip_policy.py'): source = Path(__file__).parent / name if source.resolve() != (destination / name).resolve(): shutil.copy2(source, destination / name) config_path = destination / 'config.json' if config_path.exists(): config = json.loads(config_path.read_text()) config.setdefault('auto_map', {}).update(mapping) config_path.write_text(json.dumps(config, indent=2) + '\n') return [str(destination / 'processor_config.json')] @property def model_input_names(self): return ['input_ids', 'attention_mask', 'pixel_values', 'image_grid_thw', 'pixel_values_videos', 'video_grid_thw', 'audio_feature'] def decode(self, *args, **kwargs): return self.tokenizer.decode(*args, **kwargs) def batch_decode(self, *args, **kwargs): return self.tokenizer.batch_decode(*args, **kwargs) @property def _calculate_timestamps(self): # Existing project helpers override native timestamp arithmetic for AV. return self._native._calculate_timestamps @_calculate_timestamps.setter def _calculate_timestamps(self, function): self._native._calculate_timestamps = function @staticmethod def _items(value, modality): if value is None: return [] if modality == 'audio' and isinstance(value, tuple) and len(value) == 2: return [value] return value if isinstance(value, list) else [value] def _messages(self, question, media, images, videos, audios, video_audio, use_audio_in_video): items = [] for kind, value in (('image', images), ('video', videos), ('audio', audios), ('video_audio', video_audio)): for item in self._items(value, kind): items.append({'type': 'video_audio' if kind == 'video' and use_audio_in_video else kind, kind if kind != 'video_audio' else 'video': item}) for item in self._items(media, 'media'): if isinstance(item, dict): items.append(item) continue suffix = Path(str(item)).suffix.lower() if suffix in ('.jpg', '.jpeg', '.png', '.webp', '.bmp', '.tiff'): kind = 'image' elif suffix in ('.wav', '.flac', '.mp3', '.ogg', '.m4a', '.aac'): kind = 'audio' elif suffix in ('.mp4', '.mov', '.mkv', '.webm', '.avi', '.zip'): kind = 'video_audio' if use_audio_in_video else 'video' else: raise ValueError('Specify the media type explicitly for this input') items.append({'type': kind, 'video' if kind == 'video_audio' else kind: item}) content = [] for item in items: content.extend([item, {'type': 'text', 'text': '\n'}]) content.append({'type': 'text', 'text': question}) return [{'role': 'system', 'content': self.system_prompt}, {'role': 'user', 'content': content}] def _normalize_messages(self, messages, use_audio_in_video=False): normalized, routes = [], [] for message in messages: row = {key: value for key, value in message.items() if key != 'content'} if isinstance(message['content'], str): row['content'] = message['content'] else: parts = [] for item in message['content']: kind = item['type'] if kind in ('video_url', 'video_audio_url'): payload = item.get('video_url') nested = payload if isinstance(payload, dict) else {} video_source = nested.get('url') if nested else payload audio = item.get('audio_url') or nested.get('audio') audio_source = audio.get('url') if isinstance(audio, dict) else audio if video_source: kind = 'video_audio' if audio_source or nested.get('use_audio') else 'video' item = {**item, **nested, 'type': kind, 'video': video_source} if audio_source: item['audio'] = audio_source elif item['type'] == 'video_audio_url' and audio_source: kind, item = 'audio', {'type': 'audio', 'audio': audio_source} else: raise ValueError('Missing video URL in media item') if kind in ('audio_url', 'input_audio'): payload = item[kind] if kind == 'audio_url': source = payload['url'] if isinstance(payload, dict) else payload else: source = f"data:audio/{payload['format']};base64,{payload['data']}" nested = payload if isinstance(payload, dict) else {} interval = timestamp_interval_option(item, timestamp_interval_option(nested)) item = {'type': 'audio', 'audio': source, 'timestamp_interval_sec': interval} kind = 'audio' if kind == 'video' and item.get('use_audio_in_video', use_audio_in_video): kind = 'video_audio' if kind == 'text': parts.append(item['text']) elif kind in ('image', 'video', 'audio', 'video_audio'): key = 'video' if kind == 'video_audio' else kind if key not in item and 'path' not in item: raise ValueError(f'Missing {key} or path in media item') routes.append({**item, 'type': kind, 'source': item.get(key, item.get('path'))}) parts.append({'image': IMAGE, 'video': VIDEO, 'video_audio': VIDEO, 'audio': AUDIO}[kind]) else: raise ValueError(f'Unsupported content type: {kind}') row['content'] = ''.join(parts) normalized.append(row) return normalized, routes def apply_chat_template(self, conversation, *, tokenize=False, return_dict=False, return_tensors=None, add_generation_prompt=True, enable_thinking=False, **kwargs): if tokenize: result = self(messages=conversation, return_tensors=return_tensors or 'pt', add_generation_prompt=add_generation_prompt, enable_thinking=enable_thinking, **kwargs) return result if return_dict else result['input_ids'] messages, _ = self._normalize_messages(conversation) return self.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=add_generation_prompt, enable_thinking=enable_thinking, **kwargs) def __call__(self, text=None, *, question=None, media=None, images=None, videos=None, audio=None, audios=None, video_audio=None, use_audio_in_video=False, messages=None, text_is_prompt=False, fps=None, num_frames=None, nframes=None, min_frames=None, max_frames=None, total_pixels=None, min_pixels=None, max_pixels=None, images_kwargs=None, videos_kwargs=None, return_tensors='pt', add_generation_prompt=True, enable_thinking=False, timestamp_interval_sec=None): timestamp_interval_sec = resolve_timestamp_interval( timestamp_interval_sec, default=self.timestamp_interval_sec) if return_tensors != 'pt': raise ValueError('WeMMProcessor currently returns PyTorch tensors; use return_tensors="pt"') if audio is not None and audios is not None: raise ValueError('Use either audio or audios') audios = audio if audio is not None else audios if messages is not None: if any(x is not None for x in (text, question, media, images, videos, audios, video_audio)): raise ValueError('Put media and text inside messages, or use the question/media interface') normalized, routes = self._normalize_messages(messages, use_audio_in_video=use_audio_in_video) prompt = self.tokenizer.apply_chat_template(normalized, tokenize=False, add_generation_prompt=add_generation_prompt, enable_thinking=enable_thinking) else: if text is not None and question is not None: raise ValueError('Use either text or question') question = question if question is not None else text if not isinstance(question, str): raise ValueError('Pass one question string or one messages conversation') normalized, routes = self._normalize_messages(self._messages( question, media, images, videos, audios, video_audio, use_audio_in_video), use_audio_in_video=use_audio_in_video) prompt = question if text_is_prompt else self.tokenizer.apply_chat_template( normalized, tokenize=False, add_generation_prompt=add_generation_prompt, enable_thinking=enable_thinking) image_pixels, image_grids, video_pixels, video_grids, features = [], [], [], [], [] blocks = {'image': [], 'video': [], 'audio': []} block_features = {kind: [] for kind in blocks} block_metadata = {kind: [] for kind in blocks} metadata = [] for route in routes: kind, source = route['type'], resolve_media(route['source']) item_info = {'type': kind} feature = None if kind == 'image': image = self.image_processor(images=[load_image(source)], return_tensors='pt', **(images_kwargs or {})) grid = image['image_grid_thw'][0] image_pixels.append(image['pixel_values']) image_grids.append(image['image_grid_thw']) blocks['image'].append(self.tokenizer.encode('<|vision_start|>', add_special_tokens=False) + [self.tokenizer.convert_tokens_to_ids('<|image_pad|>')] * (int(grid.prod()) // self.image_processor.merge_size**2) + self.tokenizer.encode('<|vision_end|>', add_special_tokens=False)) item_info['grid'] = grid.tolist() elif kind == 'audio': interval = resolve_timestamp_interval(timestamp_interval_option(route, timestamp_interval_sec)) samples = decode_audio(source) duration = len(samples) / 16000 feature = audio_features(samples, self.feature_extractor) count = feature.shape[0] * tokens_per_segment(self.audio_config) blocks['audio'].append(audio_ids(self.tokenizer, count, self.audio_config['second_per_window'], duration, interval)) groups, times, effective = audio_timestamp_groups( count, self.audio_config['second_per_window'], duration, interval) item_info.update(duration=duration, segments=len(feature), audio_tokens=count, timestamp_interval_sec=interval, audio_groups=groups, timestamps=times, effective_audio_tokens=effective, padding_audio_tokens=count-effective) else: options = resolve_video_options(dict( fps=self.fps if fps is None else fps, num_frames=num_frames, nframes=nframes, min_frames=self.min_frames if min_frames is None else min_frames, max_frames=self.max_frames if max_frames is None else max_frames, total_pixels=self.total_pixels if total_pixels is None else total_pixels, min_pixels=self.min_pixels if min_pixels is None else min_pixels, max_pixels=max_pixels, max_allowed_frames=self.max_allowed_frames, max_allowed_total_pixels=self.max_allowed_total_pixels), route) # SGLang supplies this internal cap from server-side defaults; # the HTTP parser never accepts it from a request. deployment_cap = positive_frame_count( route.get('wemm_zip_max_frames', self.max_frames), 'ZIP deployment max_frames') options['wemm_zip_max_frames'] = min( deployment_cap, int(options.get('max_allowed_frames') or deployment_cap)) decode_keys = ('fps', 'num_frames', 'min_frames', 'max_frames', 'min_pixels', 'max_pixels', 'total_pixels', 'max_allowed_frames', 'max_allowed_total_pixels', 'wemm_zip_max_frames', 'resized_height', 'resized_width') frames, info = decode_video(source, **{key: options[key] for key in decode_keys if key in options}) bundled_audio = info.pop('audio_bytes', None) pre_resized = bool(info.get('wemm_pre_resized', False)) area = int(frames.shape[1] * frames.shape[2] * len(frames)) size = ({'longest_edge': area, 'shortest_edge': area} if pre_resized else video_resize_size(len(frames), options)) kwargs = dict(videos_kwargs or {}) kwargs.setdefault('size', size) if pre_resized: kwargs.setdefault('do_resize', False) video_metadata = {key: info[key] for key in ('total_num_frames', 'fps', 'frames_indices', 'duration', 'width', 'height', 'video_backend') if key in info} kwargs.update(do_sample_frames=False, video_metadata=[VideoMetadata(**video_metadata)]) video = self.video_processor(videos=[frames], return_tensors='pt', **kwargs) grid = video['video_grid_thw'][0] video_pixels.append(video['pixel_values_videos']) video_grids.append(video['video_grid_thw']) count, duration = 0, None if kind == 'video_audio' or bundled_audio is not None: try: samples = decode_audio(route.get('audio', bundled_audio if bundled_audio is not None else source)) except (RuntimeError, ValueError) as exc: if 'audio' not in str(exc).lower() or not any(word in str(exc).lower() for word in ('stream', 'best')): raise item_info['audio_fallback'] = 'no audio stream; video only' else: duration = len(samples) / 16000 feature = audio_features(samples, self.feature_extractor, pcm16=False) count = len(feature) * tokens_per_segment(self.audio_config) item_info['segments'] = len(feature) ids, groups, times = video_ids(self.tokenizer, grid, info, self.video_processor.temporal_patch_size, self.video_processor.merge_size, count, duration, self.audio_config['second_per_window']) blocks['video'].append(ids) item_info.update(info, grid=grid.tolist(), size=kwargs['size'], audio_tokens=count, audio_duration=duration, audio_groups=groups, timestamps=times, sampled_frames=len(frames)) metadata.append(item_info) block_kind = 'video' if kind == 'video_audio' else kind block_features[block_kind].append(feature) block_metadata[block_kind].append(item_info) if blocks[block_kind][-1].count(self.tokenizer.convert_tokens_to_ids('<|audio_pad|>')) != item_info.get('audio_tokens', 0): raise ValueError('Audio features and placeholders differ for one media item') source_ids = self.tokenizer.encode(prompt, add_special_tokens=False) patterns = {kind: self.tokenizer.encode(pattern, add_special_tokens=False) for kind, pattern in (('image', IMAGE), ('video', VIDEO), ('audio', AUDIO))} offsets = dict.fromkeys(patterns, 0) ids, index, ordered_metadata = [], 0, [] while index < len(source_ids): for kind, pattern in patterns.items(): if source_ids[index:index + len(pattern)] == pattern: if offsets[kind] >= len(blocks[kind]): raise ValueError(f'Extra {kind} placeholder in the question/prompt') ids.extend(blocks[kind][offsets[kind]]) feature = block_features[kind][offsets[kind]] if feature is not None: features.append(feature) ordered_metadata.append(block_metadata[kind][offsets[kind]]) offsets[kind] += 1 index += len(pattern) break else: ids.append(source_ids[index]) index += 1 if any(offsets[kind] != len(blocks[kind]) for kind in blocks): raise ValueError('Media items and prompt placeholders differ') data = {'input_ids': torch.tensor([ids], dtype=torch.long), 'attention_mask': torch.ones(1, len(ids), dtype=torch.long)} for pixels, grids, pkey, gkey in ((image_pixels, image_grids, 'pixel_values', 'image_grid_thw'), (video_pixels, video_grids, 'pixel_values_videos', 'video_grid_thw')): if pixels: data[pkey], data[gkey] = torch.cat(pixels), torch.cat(grids) if features: data['audio_feature'] = torch.cat(features) expected = sum(item.get('audio_tokens', 0) for item in metadata) if ids.count(self.tokenizer.convert_tokens_to_ids('<|audio_pad|>')) != expected: raise ValueError('Audio features and placeholder counts differ') result = BatchFeature(data=data) # These attributes are diagnostic metadata, never forwarded to generate. result.media_metadata = ordered_metadata result.prompt_text = prompt return result