import asyncio import concurrent import concurrent.futures import dataclasses import multiprocessing as mp import os import re from abc import ABC, abstractmethod from typing import Any, Dict, Iterator, List, Optional, Tuple, Union import numpy as np import torch from PIL import Image from transformers import BaseImageProcessor from sglang.srt.managers.schedule_batch import ( Modality, MultimodalDataItem, MultimodalInputFormat, MultimodalProcessorOutput, ) from sglang.srt.runtime_context import get_server_args from sglang.srt.utils import ( envs, is_cpu, is_npu, is_xpu, load_audio, load_image, load_video, logger, ) from sglang.srt.utils.cuda_ipc_transport_utils import ( MM_FEATURE_CACHE_SIZE, MM_ITEM_MEMORY_POOL_RECYCLE_INTERVAL, CudaIpcTensorTransportProxy, MmItemMemoryPool, ) _is_cpu = is_cpu() _is_npu = is_npu() _is_xpu = is_xpu() SGL_USE_CUDA_IPC = envs.SGLANG_USE_CUDA_IPC_TRANSPORT.get() _IPC_POOL_HANDLE_CACHE = envs.SGLANG_USE_IPC_POOL_HANDLE_CACHE.get() @dataclasses.dataclass class BaseMultiModalProcessorOutput: # input_text with all multimodality placeholder token expanded input_text: str # original pre-tokenized ids, useful for processor_output/precomputed inputs, # when they already carry the input ids input_ids: Optional[Union[List[int], torch.Tensor]] = None # frames loaded from image, in given order images: Optional[list[Union[Image.Image, dict]]] = dataclasses.field( default_factory=list ) # videos videos: Optional[list[Union[torch.Tensor, dict]]] = dataclasses.field( default_factory=list ) # audios audios: Optional[list[Union[np.ndarray, dict]]] = dataclasses.field( default_factory=list ) def organize_results(self) -> List[Tuple[Modality, Any]]: """ :return: a list of results, with their corresponding modalities """ return ( [(Modality.IMAGE, data) for data in self.images] + [(Modality.VIDEO, data) for data in self.videos] + [(Modality.AUDIO, data) for data in self.audios] ) @dataclasses.dataclass class MultimodalSpecialTokens: image_token: Optional[Union[str, List[str]]] = None video_token: Optional[Union[str, List[str]]] = None audio_token: Optional[Union[str, List[str]]] = None image_token_id: Optional[int] = None video_token_id: Optional[int] = None audio_token_id: Optional[int] = None image_token_regex: Optional[re.Pattern] = None video_token_regex: Optional[re.Pattern] = None audio_token_regex: Optional[re.Pattern] = None combined_regex: Optional[re.Pattern] = None def build(self, processor): self.convert_to_strs(processor) self.parse_regex() self.get_combined_regex() return self def convert_to_str(self, token: Union[str, int], processor) -> str: if token is None: return token if isinstance(token, str): return token return processor.tokenizer.convert_ids_to_tokens([token])[0] def convert_to_strs(self, processor): if not self.image_token: self.image_token = self.convert_to_str(self.image_token_id, processor) if not self.video_token: self.video_token = self.convert_to_str(self.video_token_id, processor) if not self.audio_token: self.audio_token = self.convert_to_str(self.audio_token_id, processor) def get_modality_of_token(self, token: str) -> Optional[Modality]: """ :return: the modality associated with the given token, if the token is a special_token or matches with the multimodal token regex """ modality = { self.image_token: Modality.IMAGE, self.video_token: Modality.VIDEO, self.audio_token: Modality.AUDIO, }.get(token) if modality: return modality for regex, modality in [ (self.image_token_regex, Modality.IMAGE), (self.video_token_regex, Modality.VIDEO), (self.audio_token_regex, Modality.AUDIO), ]: if regex and regex.match(token): return modality return None def get_token_id_by_modality(self, modality: Modality) -> Optional[int]: return { Modality.IMAGE: self.image_token_id, Modality.VIDEO: self.video_token_id, Modality.AUDIO: self.audio_token_id, }.get(modality) def parse_regex(self): if self.image_token_regex is None and self.image_token is not None: self.image_token_regex = re.compile(re.escape(self.image_token)) if self.video_token_regex is None and self.video_token is not None: self.video_token_regex = re.compile(re.escape(self.video_token)) if self.audio_token_regex is None and self.audio_token is not None: self.audio_token_regex = re.compile(re.escape(self.audio_token)) def get_combined_regex(self) -> re.Pattern: """ Builds and returns a regex, used to split input str into tokens (with mm special tokens) """ if self.combined_regex: return self.combined_regex tokens = [ self.image_token_regex, self.video_token_regex, self.audio_token_regex, ] patterns = [] flags = 0 for t in tokens: if t is not None: patterns.append(t.pattern) flags |= t.flags combined = "(" + "|".join(f"(?:{p})" for p in patterns) + ")" self.combined_regex = re.compile(combined, flags) return self.combined_regex class BaseMultimodalProcessor(ABC): models = [] gpu_image_decode = True # Enable GPU decoding by default def __init__( self, hf_config, server_args, _processor, transport_mode, *args, **kwargs ): self.hf_config = hf_config self._processor = _processor self.server_args = server_args self.transport_mode = transport_mode self.keep_mm_feature_on_device = server_args.keep_mm_feature_on_device self.disable_fast_image_processor = server_args.disable_fast_image_processor self.skip_tokenizer_init = server_args.skip_tokenizer_init mm_process_config = self.server_args.mm_process_config self.image_config = mm_process_config.get("image", {}) self.video_config = mm_process_config.get("video", {}) self.audio_config = mm_process_config.get("audio", {}) # Resolve tokenizer: some processors (e.g. InternVL) pass a tokenizer # directly as _processor rather than a processor that wraps a tokenizer. if hasattr(self._processor, "tokenizer"): self._tokenizer = self._processor.tokenizer else: self._tokenizer = self._processor # Same guard as in serving_chat.py against double BOS. try: self._tokenizer_auto_adds_specials = len(self._tokenizer.encode("")) > 0 except Exception: self._tokenizer_auto_adds_specials = False # FIXME: not accurate, model and image specific self.NUM_TOKEN_PER_FRAME = 330 self.io_executor = concurrent.futures.ThreadPoolExecutor( max_workers=int(os.environ.get("SGLANG_IO_WORKERS", 4)) ) self.cpu_executor = concurrent.futures.ProcessPoolExecutor( mp_context=mp.get_context("fork"), max_workers=int(os.environ.get("SGLANG_CPU_WORKERS", os.cpu_count())), ) # Mapping from attribute names to modality types self.ATTR_NAME_TO_MODALITY = { # Image-related attributes "pixel_values": Modality.IMAGE, "image_sizes": Modality.IMAGE, "image_grid_thw": Modality.IMAGE, "image_attention_mask": Modality.IMAGE, "image_emb_mask": Modality.IMAGE, "images_spatial_crop": Modality.IMAGE, "images_crop": Modality.IMAGE, "has_local_crops": Modality.IMAGE, "has_images": Modality.IMAGE, "tgt_size": Modality.IMAGE, "image_grid_hws": Modality.IMAGE, "aspect_ratio_ids": Modality.IMAGE, "aspect_ratio_mask": Modality.IMAGE, "num_patches": Modality.IMAGE, "patch_pixel_values": Modality.IMAGE, "block_sizes": Modality.IMAGE, "grid_thws": Modality.IMAGE, # for kimi k2.5 # Audio-related attributes "audio_features": Modality.AUDIO, "audio_feature_lens": Modality.AUDIO, "input_features": Modality.AUDIO, "input_features_mask": Modality.AUDIO, "audio_attention_mask": Modality.AUDIO, "feature_attention_mask": Modality.AUDIO, # Video-related attributes "pixel_values_videos": Modality.VIDEO, "second_per_grid_ts": Modality.VIDEO, "video_grid_thw": Modality.VIDEO, # Generic attributes that could apply to multiple modalities # "precomputed_embeddings" - handled specially as it can be any modality } # name of the feature filed # TODO: pass from processors self.FEATURE_NAMES = [ "pixel_values", "pixel_values_videos", "audio_features", "input_features", ] skip_mm_pool = kwargs.get("skip_mm_pool", False) if SGL_USE_CUDA_IPC and not skip_mm_pool: # SGLANG_MM_FEATURE_CACHE_MB is the total pool budget across all # tokenizer workers. Each worker gets an equal share so that adding # workers doesn't multiply the GPU-side footprint. worker_num = self.server_args.tokenizer_worker_num per_worker_pool_size = max( MM_FEATURE_CACHE_SIZE // worker_num, 128 * 1024 * 1024, ) logger.info( "MmItemMemoryPool size per tokenizer worker: %.0f MiB " "(budget %.0f MiB / %d worker(s))", per_worker_pool_size / (1024 * 1024), MM_FEATURE_CACHE_SIZE / (1024 * 1024), worker_num, ) self.cudaipc_mmfeature_pool = MmItemMemoryPool( per_worker_pool_size, MM_ITEM_MEMORY_POOL_RECYCLE_INTERVAL, self.server_args.base_gpu_id, ) def compute_mrope_positions(self, input_ids, mm_items): """Compute M-RoPE positions from expanded input_ids and multimodal items. Returns (mrope_positions, mrope_position_delta) or (None, None) if the model does not use M-RoPE. """ return None, None @property def spatial_merge_size(self): return self.hf_config.vision_config.spatial_merge_size def build_input_ids( self, prompt, img_grid_thw=None, video_grid_thw=None, audio_seq_lens=None ): """ Use prompt, img_grid_thw, video_grid_thw, and audio_seq_lens to build input_ids. Supports image, video, and audio tokens. """ if not isinstance(prompt, list): prompt = self._tokenizer.encode(prompt) img_token_id = getattr(self, "IM_TOKEN_ID", None) video_token_id = getattr(self, "VIDEO_TOKEN_ID", None) audio_token_id = getattr(self, "audio_token_id", None) spatial_merge_size = getattr(self, "spatial_merge_size", 1) input_ids = [] offsets = [] cur_idx = 0 # Use img_token_id instead of im_start_id, because a dummy im_start_id # may be generated by the tokenizer. vision_start_indices = [] for i in range(len(prompt) - 1): if img_token_id is not None and prompt[i + 1] == img_token_id: vision_start_indices.append((i, Modality.IMAGE)) elif video_token_id is not None and prompt[i + 1] == video_token_id: vision_start_indices.append((i, Modality.VIDEO)) elif audio_token_id is not None and prompt[i + 1] == audio_token_id: vision_start_indices.append((i, Modality.AUDIO)) # get modality list with order preserved modality_list = [modality for _, modality in vision_start_indices] img_idx = 0 video_idx = 0 audio_idx = 0 for mm_start_idx, modality in vision_start_indices: if modality == Modality.IMAGE: mm_token_num = img_grid_thw[img_idx].prod() // (spatial_merge_size**2) mm_token_id = img_token_id img_idx += 1 elif modality == Modality.VIDEO: mm_token_num = video_grid_thw[video_idx].prod() // ( spatial_merge_size**2 ) mm_token_id = video_token_id video_idx += 1 elif modality == Modality.AUDIO: mm_token_num = int(audio_seq_lens[audio_idx].item()) mm_token_id = audio_token_id audio_idx += 1 else: raise ValueError(f"Invalid modality: {modality}") assert cur_idx <= mm_start_idx input_ids.extend(prompt[cur_idx : mm_start_idx + 1]) mm_offset_start = len(input_ids) input_ids.extend([mm_token_id] * mm_token_num) cur_idx = ( mm_start_idx + 2 ) # jump to img_end_id, video_end_id, or audio_end_id offsets.append((mm_offset_start, len(input_ids) - 1)) else: input_ids.extend(prompt[cur_idx:]) return input_ids, offsets, modality_list def get_mm_data(self, prompt, embeddings, **kwargs): img_grid_thw = kwargs.get("img_grid_thw", None) video_grid_thw = kwargs.get("video_grid_thw", None) audio_feature_lens = kwargs.get("audio_feature_lens", None) input_ids, offsets, modality_list = self.build_input_ids( prompt, img_grid_thw=img_grid_thw, video_grid_thw=video_grid_thw, audio_seq_lens=audio_feature_lens, ) assert all(isinstance(modality, Modality) for modality in modality_list) mm_items = [] consumed_per_modality = {} for modality, offset in zip(modality_list, offsets): num_tokens = offset[1] - offset[0] + 1 embedding_start = consumed_per_modality.get(modality, 0) embedding_slice = embeddings[modality][ embedding_start : embedding_start + num_tokens ] consumed_per_modality[modality] = embedding_start + num_tokens mm_items.append( MultimodalDataItem( modality=modality, offsets=[offset], precomputed_embeddings=embedding_slice, ) ) return MultimodalProcessorOutput( input_ids=input_ids, mm_items=mm_items, im_start_id=self.IM_START_TOKEN_ID, im_end_id=self.IM_END_TOKEN_ID, im_token_id=self.IM_TOKEN_ID, video_token_id=getattr(self, "VIDEO_TOKEN_ID", None), ) def process_mm_data( self, input_text, images=None, videos=None, audios=None, **kwargs ) -> dict: """ process multimodal data with transformers AutoProcessor """ if images: kwargs["images"] = images if self.image_config: kwargs.setdefault("images_kwargs", {}).update(self.image_config) if videos: kwargs["videos"] = videos if self.video_config: kwargs.setdefault("videos_kwargs", {}).update(self.video_config) if audios: if self._processor.__class__.__name__ in { "Gemma3nProcessor", "Gemma4Processor", "Gemma4UnifiedProcessor", "GlmAsrProcessor", "Qwen2AudioProcessor", "Qwen3ASRProcessor", "Qwen3OmniMoeProcessor", }: # Note(Xinyuan): for gemma3n, ref: https://github.com/huggingface/transformers/blob/ccf2ca162e33f381e454cdb74bf4b41a51ab976d/src/transformers/models/gemma3n/processing_gemma3n.py#L107 kwargs["audio"] = audios kwargs.setdefault("audio_kwargs", {}) kwargs["audio_kwargs"].setdefault("truncation", False) else: kwargs["audios"] = audios if self.audio_config: kwargs.setdefault("audio_kwargs", {}).update(self.audio_config) processor = self._processor if ( hasattr(processor, "image_processor") and isinstance(processor.image_processor, BaseImageProcessor) and not self.disable_fast_image_processor ): if _is_cpu or get_server_args().rl_on_policy_target is not None: kwargs["device"] = "cpu" elif _is_xpu: kwargs["device"] = "xpu" elif not _is_npu: base_gpu_id = get_server_args().base_gpu_id kwargs["device"] = f"cuda:{base_gpu_id}" elif processor.__class__.__name__ not in { "Glm4vProcessor", "Glm46VProcessor", }: # Note: for qwen-vl, processor has some reshape issue because of dims restriction on Ascend. from sglang.srt.hardware_backend.npu.modules.qwen_vl_processor import ( npu_apply_qwen_image_preprocess_patch, ) npu_apply_qwen_image_preprocess_patch() kwargs["device"] = "npu" elif processor.__class__.__name__ == "Glm46VProcessor": from sglang.srt.hardware_backend.npu.modules.glm46v_processor import ( npu_apply_glm46v_image_preprocess_patch, ) npu_apply_glm46v_image_preprocess_patch() kwargs["device"] = "npu" # Avoid double BOS when the chat template already wrote one. if self._tokenizer_auto_adds_specials and isinstance(input_text, str): bos = getattr(self._tokenizer, "bos_token", None) if bos and input_text.startswith(bos): kwargs.setdefault("add_special_tokens", False) result = processor.__call__( text=[input_text], padding=True, return_tensors="pt", **kwargs, ) if not self.keep_mm_feature_on_device: # move feature tensors to cpu for feature_name in self.FEATURE_NAMES: if SGL_USE_CUDA_IPC: pass else: if feature_name in result and isinstance( result[feature_name], torch.Tensor ): result[feature_name] = result[feature_name].to("cpu") return result @abstractmethod async def process_mm_data_async( self, image_data, audio_data, input_text, request_obj, **kwargs, ) -> Optional[Dict[str, Any]]: pass def get_estimated_frames_list(self, image_data): """ estimate the total frame count from all visual input """ from sglang.srt.utils.video_decoder import VideoDecoderWrapper # Before processing inputs if not image_data or len(image_data) == 0: return [] estimated_frames_list = [] for image in image_data: if isinstance(image, str) and image.startswith("video:"): path = image[len("video:") :] decoder = VideoDecoderWrapper(path) num_frames = len(decoder) else: # For images, each contributes one frame num_frames = 1 estimated_frames_list.append(num_frames) return estimated_frames_list @classmethod def _load_single_item( cls, data, modality: Modality, frame_count_limit=None, audio_sample_rate: Optional[int] = None, discard_alpha_channel=True, ): """ Load a single multimodal data. If data is processor_output or precomputed embedding, return directly. Class method that can be pickled for multiprocessing """ if cls._is_preprocessed_input(data): return data try: if modality == Modality.IMAGE: img, _ = load_image(data, cls.gpu_image_decode) if isinstance(img, torch.Tensor): return img # JPEG already decoded on GPU by nvJPEG # PIL decodes lazily; do it here in the io worker so the decode # doesn't run later on the event-loop thread. if discard_alpha_channel and img.mode != "RGB": return img.convert("RGB") img.load() return img elif modality == Modality.VIDEO: return load_video(data, frame_count_limit) elif modality == Modality.AUDIO: return load_audio(data, audio_sample_rate) except ValueError as e: # Bad input (e.g. invalid base64) -> 400, not 500. data_str = str(data) if len(data_str) > 100: data_str = data_str[:100] + "..." raise ValueError(f"Error while loading data {data_str}: {e}") from e except Exception as e: data_str = str(data) if len(data_str) > 100: data_str = data_str[:100] + "..." raise RuntimeError(f"Error while loading data {data_str}: {e}") from e @staticmethod def _get_preprocessed_input_format(data): """returns the detailed format if the provided data is already preprocessed. returns none if the provided data is not preprocessed """ if not isinstance(data, dict): return None data_format = data.get("format") if isinstance(data_format, MultimodalInputFormat): return data_format if data_format in ( MultimodalInputFormat.PROCESSOR_OUTPUT.name, "processor_output", ): return MultimodalInputFormat.PROCESSOR_OUTPUT if data_format in ( MultimodalInputFormat.PRECOMPUTED_EMBEDDING.name, "precomputed_embedding", ): return MultimodalInputFormat.PRECOMPUTED_EMBEDDING return None @classmethod def _is_preprocessed_input(cls, data): """returns if the data is already preprocessed (by the vlm processor)""" return cls._get_preprocessed_input_format(data) is not None @classmethod def _all_mm_data_is_preprocessed(cls, *data_lists): has_mm_data = False for data_list in data_lists: if not data_list: continue if not isinstance(data_list, list): data_list = [data_list] for item in data_list: if item is None: continue has_mm_data = True if not cls._is_preprocessed_input(item): return False return has_mm_data def _submit_mm_data_loading_tasks_simple( self, data_list: Optional[list], modality: Modality, audio_sample_rate: Optional[int], discard_alpha_channel: bool, ) -> List[Tuple[Modality, int, concurrent.futures.Future]]: """ Simple version: For one modal data submit IO load task. Return: List[(modality, index_in_that_modality, future)] """ futures: List[Tuple[Modality, int, concurrent.futures.Future]] = [] if not data_list: logger.debug( "[_submit_mm_data_loading_tasks_simple] no data for modality=%s", modality.name, ) return futures for idx, data in enumerate(data_list): logger.debug( "[_submit_mm_data_loading_tasks_simple] submit load task: " "modality=%s, index=%d, data_type=%s", modality.name, idx, type(data), ) future = self.io_executor.submit( self.__class__._load_single_item, data, modality, None, # frame_count_limit: no consider for fast path audio_sample_rate, discard_alpha_channel, ) futures.append((modality, idx, future)) return futures def submit_data_loading_tasks( self, text_parts: List[str], multimodal_tokens: MultimodalSpecialTokens, data_iterators: dict[Modality, Iterator[Any]], discard_alpha_channel: bool = True, image_estimated_frames_iter: Optional[iter] = None, image_scaling_factor: float = 1.0, max_image_frames: int = 30, audio_sample_rate: Optional[int] = None, ) -> Tuple[List, List]: """ load multimodal data parallelly using iterators. """ futures = [] task_info = [] for text_part in text_parts: modality = multimodal_tokens.get_modality_of_token(text_part) if modality is not None: data_iterator = data_iterators.get(modality) if data_iterator is None: raise ValueError(f"No data iterator found for token: {text_part}") try: data = next(data_iterator) except StopIteration: logger.warning( f"Mismatch: More '{modality.name}' tokens found than corresponding data provided." ) return futures, task_info frame_count_limit = None if modality == Modality.IMAGE and image_estimated_frames_iter: try: estimated_frames = next(image_estimated_frames_iter) # Use the pre-calculated scaling factor and max frames frame_count_limit = max( 1, int(estimated_frames * image_scaling_factor) ) # Ensure we don't exceed the absolute max (redundant if scaling_factor handles it) # frame_count_limit = min(frame_count_limit, max_image_frames) except StopIteration: raise ValueError( "Mismatch between image tokens and estimated frame counts." ) futures.append( self.io_executor.submit( self.__class__._load_single_item, data, modality, frame_count_limit, audio_sample_rate, discard_alpha_channel, ) ) task_info.append((modality, data, frame_count_limit)) for modality, iterator in data_iterators.items(): try: next(iterator) logger.warning( f"Warning: More {modality.name.lower()} data items provided than corresponding tokens found in the prompt." ) except StopIteration: pass except Exception: pass return futures, task_info @staticmethod def _validate_one_modality(modality: Modality, data_list: Optional[list]): if data_list is None: return if not isinstance(data_list, list): raise TypeError( f"{modality.name} must be a list or None, got {type(data_list)}" ) formatted_indices = [] for idx, item in enumerate(data_list): if BaseMultimodalProcessor._is_preprocessed_input(item): formatted_indices.append(idx) if formatted_indices: if len(data_list) != 1: raise ValueError( f"For {modality}, when providing a 'processor_output' or " f"'precomputed_embedding', you must pass exactly one item; " f"received {len(data_list)} items (formatted at indices {formatted_indices})." ) @staticmethod def validate_mm_data( image_data: Optional[list] = None, video_data: Optional[list] = None, audio_data: Optional[list] = None, ): """ Validate multimodal input lists per modality. Rule per modality (image/video/audio): - Either the list has exactly one item and that single item is a dict with format in {"processor_output", "precomputed_embedding"}; - Or, the list contains only "normal" items (i.e., does not include any item whose format is one of the two above). Empty or None lists are considered valid. """ BaseMultimodalProcessor._validate_one_modality(Modality.IMAGE, image_data) BaseMultimodalProcessor._validate_one_modality(Modality.VIDEO, video_data) BaseMultimodalProcessor._validate_one_modality(Modality.AUDIO, audio_data) def _process_loaded_mm_data(self, modality, raw_data, result): images, videos, audios = [], [], [] is_precomputed = self._is_preprocessed_input(raw_data) if modality == Modality.IMAGE: if is_precomputed: images.append(result) else: if isinstance(result, list): images.extend(result) else: images.append(result) elif modality == Modality.VIDEO: videos.append(result) elif modality == Modality.AUDIO: audios.append(result) return is_precomputed, images, videos, audios async def load_mm_data( self, prompt: str, multimodal_tokens: MultimodalSpecialTokens, image_data: Optional[list] = None, video_data: Optional[list] = None, audio_data: Optional[list] = None, return_text: Optional[bool] = True, discard_alpha_channel: bool = True, audio_sample_rate: Optional[int] = None, ) -> BaseMultiModalProcessorOutput: BaseMultimodalProcessor.validate_mm_data(image_data, video_data, audio_data) input_ids = prompt if isinstance(prompt, list) else None if input_ids is not None and self._all_mm_data_is_preprocessed( image_data, video_data, audio_data ): # fast path for preprocessed data: early return return BaseMultiModalProcessorOutput( input_text="", input_ids=input_ids, images=list(image_data or []), videos=list(video_data or []), audios=list(audio_data or []), ) multimodal_tokens_pattern = multimodal_tokens.get_combined_regex() if isinstance(prompt, list) and return_text: assert len(prompt) and isinstance(prompt[0], int) prompt = self._tokenizer.decode(prompt) else: prompt = prompt assert isinstance(prompt, str) # split text into list of normal text and special tokens text_parts = re.split(multimodal_tokens_pattern, prompt) cnt = {Modality.IMAGE: 0, Modality.VIDEO: 0, Modality.AUDIO: 0} for text_part in text_parts: modality = multimodal_tokens.get_modality_of_token(text_part) if modality is not None: cnt[modality] += 1 n_image = len(image_data) if image_data else 0 n_video = len(video_data) if video_data else 0 n_audio = len(audio_data) if audio_data else 0 # For MiniCPMO and MiniCPMV or multimodal_tokens not totally align, legacy show path if ( self.skip_tokenizer_init or cnt[Modality.IMAGE] != n_image or cnt[Modality.VIDEO] != n_video or cnt[Modality.AUDIO] != n_audio or getattr(self, "support_dynamic_frame_expansion", False) ): return await self.legacy_load_mm_data( prompt=prompt, multimodal_tokens=multimodal_tokens, image_data=image_data, video_data=video_data, audio_data=audio_data, return_text=return_text, discard_alpha_channel=discard_alpha_channel, audio_sample_rate=audio_sample_rate, input_ids=input_ids, ) # For models other than MiniCPMO and MiniCPMV, # totally align multimodal_tokens, fast path return await self.fast_load_mm_data( prompt=prompt, multimodal_tokens=multimodal_tokens, image_data=image_data, video_data=video_data, audio_data=audio_data, return_text=return_text, discard_alpha_channel=discard_alpha_channel, audio_sample_rate=audio_sample_rate, input_ids=input_ids, ) async def fast_load_mm_data( self, prompt: str, multimodal_tokens: MultimodalSpecialTokens, image_data: Optional[list] = None, video_data: Optional[list] = None, audio_data: Optional[list] = None, return_text: Optional[bool] = True, discard_alpha_channel: bool = True, audio_sample_rate: Optional[int] = None, input_ids: Optional[Union[List[int], torch.Tensor]] = None, ) -> BaseMultiModalProcessorOutput: """ A fast version of `load_mm_data` that loads multimodal data directly. This version does not scan the prompt to recognize tokens. It assumes that the caller has already aligned the tokens and data in a 1:1 manner. The behavior is as follows: 1. It runs `_load_single_item` for all input data concurrently. 2. It returns the loaded images, videos, and audios in their original order. 3. It returns the input prompt as a string. """ # Convert prompt into str if isinstance(prompt, list) and return_text: assert len(prompt) and isinstance(prompt[0], int) prompt_str = self._tokenizer.decode(prompt) else: assert isinstance(prompt, str) prompt_str = prompt futures: List[Tuple[Modality, int, concurrent.futures.Future]] = [] modalities_data = [ (image_data, Modality.IMAGE), (video_data, Modality.VIDEO), (audio_data, Modality.AUDIO), ] for data_list, modality in modalities_data: futures.extend( self._submit_mm_data_loading_tasks_simple( data_list, modality, audio_sample_rate, discard_alpha_channel ) ) logger.debug("[load_mm_data(simple)] total futures submitted: %d", len(futures)) images: List[Any] = [None] * len(image_data) if image_data else [] videos: List[Any] = [None] * len(video_data) if video_data else [] audios: List[Any] = [None] * len(audio_data) if audio_data else [] for modality, idx, future in futures: try: result = await asyncio.wrap_future(future) except Exception as e: logger.exception( "[load_mm_data(simple)] error loading %s data at index=%d", modality.name, idx, ) raise RuntimeError( f"An exception occurred while loading {modality.name} data at index {idx}: {e}" ) if modality == Modality.IMAGE: images[idx] = result elif modality == Modality.VIDEO: videos[idx] = result elif modality == Modality.AUDIO: audios[idx] = result logger.debug( "[load_mm_data(simple)] loaded counts: images=%d, videos=%d, audios=%d", len(images), len(videos), len(audios), ) return BaseMultiModalProcessorOutput( images=images, audios=audios, videos=videos, input_text=prompt_str, input_ids=input_ids, ) async def legacy_load_mm_data( self, prompt: str, multimodal_tokens: MultimodalSpecialTokens, image_data: Optional[list] = None, video_data: Optional[list] = None, audio_data: Optional[list] = None, return_text: Optional[bool] = True, discard_alpha_channel: bool = True, audio_sample_rate: Optional[int] = None, input_ids: Optional[Union[List[int], torch.Tensor]] = None, ) -> BaseMultiModalProcessorOutput: """ Each frame of video/image will be replaced by a single image token Args: multimodal_tokens (list[str]): list of special token which denoting a single multimodal data e.g. image token or audio token discard_alpha_channel: if True, discards the alpha channel in the returned images """ multimodal_tokens_pattern = multimodal_tokens.get_combined_regex() if isinstance(prompt, list) and return_text: assert len(prompt) and isinstance(prompt[0], int) prompt = self._tokenizer.decode(prompt) else: prompt = prompt assert isinstance(prompt, str) # split text into list of normal text and special tokens text_parts = re.split(multimodal_tokens_pattern, prompt) # collect all data data_iterators = {} if multimodal_tokens.image_token and image_data: data_iterators[Modality.IMAGE] = iter(image_data) if multimodal_tokens.video_token and video_data: data_iterators[Modality.VIDEO] = iter(video_data) if multimodal_tokens.audio_token and audio_data: data_iterators[Modality.AUDIO] = iter(audio_data) # futures: the futures of loaded data # task_info: modality, raw_data, and other metadata of each data futures, task_info = self.submit_data_loading_tasks( text_parts=text_parts, multimodal_tokens=multimodal_tokens, data_iterators=data_iterators, discard_alpha_channel=discard_alpha_channel, audio_sample_rate=audio_sample_rate, ) task_info_iter = iter(task_info) futures_iter = iter(futures) # Process results images, videos, audios = [], [], [] new_text_parts = [] has_precomputed_input = False for text_part in text_parts: try: if multimodal_tokens_pattern.match(text_part): modality, raw_data, frame_limit = next(task_info_iter) result = await asyncio.wrap_future(next(futures_iter)) is_precomputed, new_imgs, new_vids, new_auds = ( self._process_loaded_mm_data(modality, raw_data, result) ) has_precomputed_input |= is_precomputed images.extend(new_imgs) videos.extend(new_vids) audios.extend(new_auds) if modality == Modality.IMAGE: if is_precomputed: new_text_parts += [text_part] else: count = len(new_imgs) if count > 0: new_text_parts += [ multimodal_tokens.image_token ] * count elif modality == Modality.VIDEO: # load as video mm_tokens = ( text_part if is_precomputed else multimodal_tokens.video_token ) new_text_parts += mm_tokens elif modality == Modality.AUDIO: # audio mm_tokens = ( text_part if is_precomputed else multimodal_tokens.audio_token ) new_text_parts += mm_tokens else: # normal text new_text_parts += [text_part] except StopIteration as e: # when precomputed_input is presented with multi-images, StopIteration is expected if has_precomputed_input: new_text_parts += [text_part] continue raise RuntimeError( f"An exception occurred while loading multimodal data: {e}" ) except Exception as e: raise RuntimeError( f"An exception occurred while loading multimodal data: {e}" ) return BaseMultiModalProcessorOutput( images=images, audios=audios, videos=videos, input_text="".join(new_text_parts), input_ids=input_ids, ) @staticmethod def get_mm_items_offset( input_ids: torch.Tensor, mm_token_id: int ) -> List[Tuple[int, int]]: """ Get a set of range for mm_items from input_ids Example: input_ids = [1, 2, 3, 3, 3, 4, 3, 3] mm_token_id = 3 return result = [(2,4),(6,7)] """ mask = input_ids == mm_token_id start_positions = (mask & ~torch.roll(mask, 1)).nonzero(as_tuple=True)[0] end_positions = (mask & ~torch.roll(mask, -1)).nonzero(as_tuple=True)[0] return list(zip(start_positions.tolist(), end_positions.tolist())) @staticmethod def get_mm_items_offset_by_pair( input_ids: torch.Tensor, mm_start_id: int, mm_end_id: int ) -> List[Tuple[int, int]]: indices_start = (input_ids == mm_start_id).nonzero(as_tuple=True)[0] + 1 indices_end = (input_ids == mm_end_id).nonzero(as_tuple=True)[0] - 1 return list(zip(indices_start.tolist(), indices_end.tolist())) def collect_mm_items_from_processor_output( self, data_dict: dict, modality: Modality = None ) -> List[MultimodalDataItem]: """ Create mm_items from processor output. Initially creates one item per modality; these are later split into per-image/video items by get_new_expanded_mm_items. Note that the data_dict can be hf processor output, or passed via offline engine api Args: modality: if provided, force the data into a single MultimodalDataItem of that modality """ # universal getter for data_dict get_data_value = ( data_dict.get if hasattr(data_dict, "get") else lambda name, default=None: getattr(data_dict, name, default) ) # decide explicitly-set modality explicit_modality = modality modality_value = get_data_value("modality") if explicit_modality is None and modality_value is not None: explicit_modality = ( modality_value if isinstance(modality_value, Modality) else Modality.from_str(str(modality_value)) ) items: dict[Modality, MultimodalDataItem] = {} for attr_name, value in data_dict.items(): if attr_name in ( "input_ids", "format", "modality", "hash", "pad_value", "offsets", ): # metadata fields need explicit handling, skip generic item.set continue # Get modality for this attribute current_modality = explicit_modality or self.ATTR_NAME_TO_MODALITY.get( attr_name ) if attr_name == "precomputed_embeddings": current_modality = current_modality or Modality.IMAGE if current_modality: # Create item if needed if current_modality not in items: items[current_modality] = MultimodalDataItem( modality=current_modality ) if attr_name in self.FEATURE_NAMES: attr_name = "feature" items[current_modality].set(attr_name, value) # deal with metadata fields when data_dict is preprocessed input: convert from tensor to expected python types # the attribution of the metadata fields is only clear when number of MultimodalDataItem is 1 if len(items) == 1: item = next(iter(items.values())) # adjust offset offsets = get_data_value("offsets") if offsets is not None: if isinstance(offsets, torch.Tensor): offsets = offsets.detach().cpu().tolist() item.offsets = [(int(start), int(end)) for start, end in offsets] # adjust hash_value hash_value = get_data_value("hash") if hash_value is not None: if isinstance(hash_value, torch.Tensor): hash_value = hash_value.item() item.hash = int(hash_value) pad_value = get_data_value("pad_value") if pad_value is not None: if isinstance(pad_value, torch.Tensor): pad_value = pad_value.item() item.pad_value = int(pad_value) return list(items.values()) def _process_and_collect_mm_items( self, input_text: str, images=None, audios=None, videos=None, **kwargs ) -> Tuple[List[MultimodalDataItem], torch.Tensor, dict]: """ Helper method to process multimodal data and create mm_items in one step. Returns: Tuple of (created mm_items, input_ids) """ ret = self.process_mm_data( input_text=input_text, images=images, audios=audios, videos=videos, **kwargs ) input_ids = ret["input_ids"].flatten() collected_items = self.collect_mm_items_from_processor_output(ret) return collected_items, input_ids, ret @staticmethod def _ensure_input_ids_is_tensor(input_ids) -> Optional[torch.Tensor]: """make sure the input_ids is a flattened tensor""" if input_ids is None: return None if isinstance(input_ids, torch.Tensor): return input_ids.flatten().to(dtype=torch.long) return torch.tensor(input_ids, dtype=torch.long).flatten() def _wrap_tensor_for_cuda_ipc(self, tensor: torch.Tensor): """helper function to turn a tensor into a cuda-ipc tensor""" if not tensor.is_cuda: return tensor sync_flag, available_slice, byte_offset = ( self.cudaipc_mmfeature_pool.return_a_slice_tensor_with_flag(tensor) ) if isinstance(available_slice, torch.Tensor): available_slice.copy_(tensor.view(torch.int8).view(-1), non_blocking=True) return CudaIpcTensorTransportProxy( data=available_slice, info_data=tensor, sync_buffer_meta=sync_flag, pool_ipc_handle=( self.cudaipc_mmfeature_pool._pool_ipc_handle if _IPC_POOL_HANDLE_CACHE else None ), pool_byte_offset=byte_offset, pool_device_index=self.cudaipc_mmfeature_pool._pool_device_index, ) if self.keep_mm_feature_on_device: return tensor return tensor.cpu() def resolve_image_token_counts(self, images: List) -> List[int]: """Per-image expanded token counts, computed without re-tokenizing. Default implementation uses the transformers in-tree convention ``_get_num_multimodal_tokens(image_sizes=...)`` (present on the in-tree VLM processors, e.g. Qwen-VL, Gemma3, GLM4V). Models whose processor does not implement it (e.g. Kimi) override this method. """ assert images is not None image_sizes = [(image.height, image.width) for image in images] num_image_tokens = self._processor._get_num_multimodal_tokens( image_sizes=image_sizes ).num_image_tokens return [int(count) for count in num_image_tokens] @staticmethod def _expand_input_ids( original_ids: List[int], counts: List[int], placeholder_token_id: Optional[int], ) -> List[int]: """Rebuild final input_ids for a pre-tokenized (list[int]) prompt. Keep the user's ORIGINAL tokens verbatim and expand the i-th image placeholder into ``counts[i]`` copies of ``placeholder_token_id``. The HF processor's re-tokenization is discarded, so non-media tokens cannot drift. """ if placeholder_token_id is None: raise ValueError("placeholder_token_id is not set for this processor") num_placeholders = sum( 1 for token_id in original_ids if token_id == placeholder_token_id ) if num_placeholders != len(counts): raise ValueError( f"prompt has {num_placeholders} image placeholder token(s) but " f"{len(counts)} image(s) were provided" ) rebuilt: List[int] = [] next_image_idx = 0 for token_id in original_ids: if token_id == placeholder_token_id: rebuilt.extend([placeholder_token_id] * counts[next_image_idx]) next_image_idx += 1 else: rebuilt.append(token_id) return rebuilt def process_and_combine_mm_data( self, base_output: BaseMultiModalProcessorOutput, mm_tokens: MultimodalSpecialTokens, **kwargs, ) -> Tuple[List[MultimodalDataItem], torch.Tensor, dict]: """ Process multimodal data and return the combined multimodal items and input_ids. Supports mixed modalities (images and audio in the same request). Returns: Tuple of (list of mm_items, input_ids) """ # Collect all items and categorize them all_loaded_data = base_output.organize_results() # Handle text-only case if not all_loaded_data: input_ids = self._tokenizer( base_output.input_text, return_tensors="pt", add_special_tokens=True, ).input_ids.flatten() return [], input_ids, {} dict_items, raw_images, raw_audios, raw_videos = [], [], [], [] for modality, item in all_loaded_data: if isinstance(item, dict): dict_items.append((modality, item)) elif modality == Modality.IMAGE: raw_images.append(item) elif modality == Modality.AUDIO: raw_audios.append(item) elif modality == Modality.VIDEO: raw_videos.append(item) else: raise ValueError(f"Unknown multimodal item type: {type(item)}") # Process items and get input_ids all_collected_items: list[MultimodalDataItem] = [] input_ids = None # Handle raw items (need processing) if raw_images or raw_audios or raw_videos: collected_items, input_ids, ret = self._process_and_collect_mm_items( input_text=base_output.input_text, images=raw_images, audios=raw_audios, videos=raw_videos, **kwargs, ) all_collected_items = collected_items # When SGLANG_MM_AVOID_RETOKENIZE is on, keep the user's exact tokens to avoid retokenize drift. # Drift happens when Retokenization is not identity: Decode(X) => String => Re-tokenize => Y, X != Y. if ( envs.SGLANG_MM_AVOID_RETOKENIZE.get() and base_output.input_ids is not None and input_ids is not None and raw_images and not raw_audios and not raw_videos ): assert isinstance( base_output.input_ids, list ), f"expected list[int] input_ids, got {type(base_output.input_ids)}" try: counts = self.resolve_image_token_counts(raw_images) image_placeholder_token_id = mm_tokens.image_token_id if image_placeholder_token_id is None: raise ValueError( "image placeholder token id is not set for this processor" ) processor_placeholder_count = int( (input_ids == image_placeholder_token_id).sum().item() ) if processor_placeholder_count != sum(counts): raise ValueError( "processor image placeholder count mismatch: " f"processor={processor_placeholder_count}, " f"resolved={sum(counts)}" ) input_ids = torch.tensor( self._expand_input_ids( base_output.input_ids, counts, image_placeholder_token_id, ), dtype=input_ids.dtype, ) except Exception as e: logger.warning( f"Due to {e}, falling back to decode+retokenize, which may change prompt length (token drift)." ) else: ret = None # Handle dict items (processed or precomputed) dict_ret = None for modality, dict_item in dict_items: input_format = self._get_preprocessed_input_format(dict_item) if input_format is not None and dict_ret is None: dict_ret = dict_item if input_format == MultimodalInputFormat.PROCESSOR_OUTPUT: items = self.collect_mm_items_from_processor_output(dict_item) for item in items: item.format = MultimodalInputFormat.PROCESSOR_OUTPUT all_collected_items.extend(items) elif input_format == MultimodalInputFormat.PRECOMPUTED_EMBEDDING: dict_item = dict(dict_item) feature = dict_item.pop("feature") all_collected_items.append( MultimodalDataItem( modality=modality, feature=feature, format=MultimodalInputFormat.PRECOMPUTED_EMBEDDING, model_specific_data=dict_item, ) ) # Fallback tokenization if no raw items were processed if ret is None and dict_ret is not None: ret = dict_ret if input_ids is None: input_ids = self._ensure_input_ids_is_tensor(base_output.input_ids) if input_ids is None: for _, dict_item in dict_items: input_ids = self._ensure_input_ids_is_tensor(dict_item.get("input_ids")) if input_ids is not None: break if input_ids is None: input_ids = self._tokenizer( base_output.input_text, return_tensors="pt", add_special_tokens=True, ).input_ids.flatten() # Add offsets to all items for mm_item in all_collected_items: if mm_item.offsets is not None: continue mm_token_id = mm_tokens.get_token_id_by_modality(mm_item.modality) if mm_token_id is None: raise ValueError(f"No token id found for modality: {mm_item.modality}") mm_item.offsets = self.get_mm_items_offset( input_ids=input_ids, mm_token_id=mm_token_id, ) # Split bundled items into per-image/video items for better cache granularity from sglang.srt.managers.mm_utils import get_new_expanded_mm_items all_collected_items = get_new_expanded_mm_items(all_collected_items) for item in all_collected_items: if item.format in ( MultimodalInputFormat.PROCESSOR_OUTPUT, MultimodalInputFormat.PRECOMPUTED_EMBEDDING, ): item.set_pad_value() """ solution for cuda-ipc memory-leak: 1. memory-pool: each time get a slice from memory-pool and use it as transport-data (with async lock guard) 2. if can not get a slice , transport normal tensor 3. copy tensor in scheduler and release it (use position mark) 4. copy """ if SGL_USE_CUDA_IPC: # post-process, prepare for cuda-ipc transfer for item in all_collected_items: if isinstance(item.feature, torch.Tensor): item.feature = self._wrap_tensor_for_cuda_ipc(item.feature) if isinstance(item.precomputed_embeddings, torch.Tensor): item.precomputed_embeddings = self._wrap_tensor_for_cuda_ipc( item.precomputed_embeddings ) return all_collected_items, input_ids, ret