[VLM] Adopt fast image processor by default (#5065)

benchmark/mmmu/bench_sglang.py

@@ -89,5 +89,4 @@ if __name__ == "__main__":
     EvalArgs.add_cli_args(parser)
     args = add_common_sglang_args_and_parse(parser)
     args = parser.parse_args()
-
     eval_mmmu(args)

benchmark/mmmu/eval_utils.py

@@ -7,6 +7,7 @@ import os
 import pprint
 import random
 import re
+from concurrent.futures import ThreadPoolExecutor, as_completed
 from typing import Dict, Optional
 
 import numpy as np
@@ -117,29 +118,38 @@ def prepare_samples(eval_args: EvalArgs):
     # merge all dataset
     dataset = concatenate_datasets(sub_dataset_list)
 
-    ## prepare images
+    # Prepare images in parallel
-    samples = []
-    skip_count = 0
-
-    # use image file as input to ensure the consistency between sglang and hf
     images_path = os.path.expanduser("~/.cache/mmmu/images")
     os.makedirs(images_path, exist_ok=True)
     print(f"Saving images to: {images_path}")
 
-    for i, sample in enumerate(tqdm(dataset)):
+    samples = []
+    skip_count = 0
+
+    def process_sample(i, sample):
         sample = process_single_sample(sample)
         sample = construct_prompt(sample, eval_args.config)
         image = sample["image"]
-
         width, height = image.size
         if width * height >= eval_args.image_pixels_limit:
-            skip_count += 1
+            return None, True
-            continue
         image_path = f"{images_path}/image_{i}.png"
         if not os.path.exists(image_path):
             image.save(image_path)
         sample["image_path"] = image_path
-        samples.append(sample)
+        return sample, False
+
+    with ThreadPoolExecutor() as executor:
+        futures = [
+            executor.submit(process_sample, i, sample)
+            for i, sample in enumerate(dataset)
+        ]
+        for future in tqdm(as_completed(futures), total=len(futures)):
+            sample, skipped = future.result()
+            if skipped:
+                skip_count += 1
+            elif sample:
+                samples.append(sample)
 
     print(
         f"skipping {skip_count} samples with large images, {round((float(skip_count) / len(dataset)) * 100, 2)}% of dataset"

docs/backend/server_arguments.md

@@ -45,7 +45,7 @@ python -m sglang.launch_server --model-path meta-llama/Meta-Llama-3-8B-Instruct
 Please consult the documentation below to learn more about the parameters you may provide when launching a server.
 
 
-## Model and tokenizer
+## Model, processor and  tokenizer
 
 * `model_path`: Path to the model that will be served.
 * `tokenizer_path`: Defaults to the `model_path`.
@@ -62,6 +62,7 @@ Please consult the documentation below to learn more about the parameters you ma
 * `skip_tokenizer_init`: Set to true to provide the tokens to the engine and get the output tokens directly, typically used in RLHF. Please see this [example for reference](https://github.com/sgl-project/sglang/blob/main/examples/runtime/token_in_token_out/).
 * `json_model_override_args`: Override model config with the provided JSON.
 * `delete_ckpt_after_loading`: Delete the model checkpoint after loading the model.
+* `disable_fast_image_processor`: Adopt base image processor instead of fast image processor(which is by default). For more detail, see: https://huggingface.co/docs/transformers/main/en/main_classes/image_processor#image-processor
 
 
 ## Serving: HTTP & API

python/sglang/srt/hf_transformers_utils.py

@@ -215,6 +215,7 @@ def get_processor(
     tokenizer_mode: str = "auto",
     trust_remote_code: bool = False,
     tokenizer_revision: Optional[str] = None,
+    use_fast: Optional[bool] = True,
     **kwargs,
 ):
     # pop 'revision' from kwargs if present.
@@ -232,6 +233,9 @@ def get_processor(
         if "size" not in kwargs:
             kwargs["size"] = {"shortest_edge": 3136, "longest_edge": 1003520}
 
+    if config.model_type not in {"llava", "clip"}:
+        kwargs["use_fast"] = use_fast
+
     processor = AutoProcessor.from_pretrained(
         tokenizer_name,
         *args,

python/sglang/srt/managers/multimodal_processors/base_processor.py

@@ -4,14 +4,16 @@ import dataclasses
 import multiprocessing as mp
 import os
 from abc import ABC, abstractmethod
-from typing import Optional
+from typing import List, Optional
 
 import numpy as np
 import PIL
 from decord import VideoReader, cpu
 from PIL import Image
+from transformers import BaseImageProcessorFast
 
-from sglang.srt.utils import encode_video, load_audio, load_image, logger
+from sglang.srt.managers.schedule_batch import Modality
+from sglang.srt.utils import encode_video, load_audio, load_image
 
 
 @dataclasses.dataclass
@@ -78,6 +80,10 @@ class BaseMultimodalProcessor(ABC):
             kwargs["audios"] = audios
 
         processor = self._processor
+        if hasattr(processor, "image_processor") and isinstance(
+            processor.image_processor, BaseImageProcessorFast
+        ):
+            kwargs["device"] = "cuda"
         result = processor.__call__(
             text=[input_text],
             padding=True,
@@ -111,6 +117,84 @@ class BaseMultimodalProcessor(ABC):
 
         return estimated_frames_list
 
+    @staticmethod
+    def _load_single_item(
+        data, is_video, is_audio, frame_count_limit=None, discard_alpha_channel=True
+    ):
+        """Static method that can be pickled for multiprocessing"""
+        try:
+            if is_audio:
+                return load_audio(data)
+            elif is_video:
+                path = data[len("video:") :]
+                return encode_video(path, frame_count_limit)
+            else:
+                img, _ = load_image(data)
+                return img.convert("RGB") if discard_alpha_channel else img
+        except Exception as e:
+            raise RuntimeError(f"Error while loading data {data}: {e}")
+
+    def submit_data_loading_tasks(
+        self,
+        text_parts: List[str],
+        multimodal_tokens: MultimodalSpecialTokens,
+        image_data: Optional[list] = None,
+        audio_data: Optional[list] = None,
+        discard_alpha_channel: bool = True,
+    ):
+        """
+        load multimodal data parallelly
+        """
+
+        # TODO(mick): load from server_args, env, or sampling_params
+        MAX_NUM_FRAMES = 30
+        estimated_frames_list = self.get_estimated_frames_list(image_data=image_data)
+        total_frame_count = sum(estimated_frames_list)
+        # a heuristic value, suggesting the maximum fraction of frames to embed from all visual inputs.
+        # e.g., 0.1 suggests that 1 frame out of 10 input frames should be used
+        scaling_factor = min(1.0, MAX_NUM_FRAMES / max(1, total_frame_count))
+
+        assert len(image_data) == len(estimated_frames_list)
+        # Submit all tasks
+        futures = []
+        task_info = []
+        image_index, audio_index = 0, 0
+
+        for text_part in text_parts:
+            if text_part == multimodal_tokens.image_token:
+                data = image_data[image_index]
+                is_video = isinstance(data, str) and data.startswith("video:")
+                estimated_frames = estimated_frames_list[image_index]
+                frame_count_limit = max(1, int(estimated_frames * scaling_factor))
+                futures.append(
+                    self.io_executor.submit(
+                        BaseMultimodalProcessor._load_single_item,
+                        data,
+                        is_video,
+                        False,
+                        frame_count_limit,
+                        discard_alpha_channel,
+                    )
+                )
+                task_info.append((Modality.IMAGE, data, frame_count_limit))
+                image_index += 1
+            elif text_part == multimodal_tokens.audio_token:
+                data = audio_data[audio_index]
+                futures.append(
+                    self.io_executor.submit(
+                        BaseMultimodalProcessor._load_single_item,
+                        data,
+                        False,
+                        True,
+                        None,
+                        discard_alpha_channel,
+                    )
+                )
+                task_info.append((Modality.AUDIO, data, None))
+                audio_index += 1
+
+        return futures, task_info
+
     def load_mm_data(
         self,
         prompt: str,
@@ -155,84 +239,37 @@ class BaseMultimodalProcessor(ABC):
             # split text into list of normal text and special tokens
             text_parts = re.split(pattern, prompt)
 
-        # TODO(mick): load from server_args, env, or sampling_params
+        futures, task_info = self.submit_data_loading_tasks(
-        MAX_NUM_FRAMES = 30
+            text_parts=text_parts,
-        estimated_frames_list = self.get_estimated_frames_list(image_data=image_data)
+            multimodal_tokens=multimodal_tokens,
-        total_frame_count = sum(estimated_frames_list)
+            image_data=image_data,
-        # a heuristic value, suggesting the maximum fraction of frames to embed from all visual inputs.
+            audio_data=audio_data,
-        # e.g., 0.1 suggests that 1 frame out of 10 input frames should be used
+            discard_alpha_channel=discard_alpha_channel,
-        scaling_factor = min(1.0, MAX_NUM_FRAMES / max(1, total_frame_count))
+        )
-
+        # Process results
-        assert len(image_data) == len(estimated_frames_list)
+        image_sizes, images, audios = [], [], []
-
-        image_index, audio_index = 0, 0
-        hashes, image_sizes, images, audios = [], [], [], []
         new_text = ""
-        for index, text_part in enumerate(text_parts):
+        task_ptr = 0
-            try:
-                if text_part == multimodal_tokens.image_token:
-                    # load as image
-                    if len(images) >= MAX_NUM_FRAMES:
-                        frames_to_process = 0
-                    else:
-                        estimated_frames = estimated_frames_list[image_index]
-                        frames_to_process = max(
-                            1, int(estimated_frames * scaling_factor)
-                        )
 
-                    if frames_to_process == 0:
+        for text_part in text_parts:
-                        frames = []
+            if text_part in multimodal_tokens.collect():
-                    else:
+                task_type, data, frame_limit = task_info[task_ptr]
-                        image_file = image_data[image_index]
+                result = futures[task_ptr].result()
-                        if isinstance(image_file, str) and image_file.startswith(
+                task_ptr += 1
-                            "video:"
-                        ):
-                            # video
-                            path = image_file[len("video:") :]
-                            frames = encode_video(
-                                path, frame_count_limit=frames_to_process
-                            )
-                        else:
-                            # image
-                            raw_image, _size = load_image(image_file)
-                            if discard_alpha_channel:
-                                raw_image = raw_image.convert("RGB")
-                            frames = [raw_image]
-                        if len(frames) == 0:
-                            continue
 
-                    image_sizes += frames[0].size * len(frames)
+                if task_type == Modality.IMAGE:
-
+                    frames = [result] if not isinstance(result, list) else result
-                    # Generate a hashable value for the image file
+                    if frames:
-                    if isinstance(image_file, Image.Image):
+                        image_sizes += frames[0].size * len(frames)
-                        # For PIL.Image objects, use the ID as a hashable value
+                        images += frames
-                        hash_value = hash(id(image_file))
-                    else:
-                        # For other types (strings, etc.), use the regular hash
-                        hash_value = hash(image_file)
-
-                    hashes += [hash_value] * len(frames)
-                    images += frames
-                    image_index += 1
-                    if frames_to_process != 0:
                         new_text += multimodal_tokens.image_token * len(frames)
-                    assert frames_to_process == len(frames)
+                elif task_type == Modality.AUDIO:
-                elif text_part == multimodal_tokens.audio_token:
+                    # audio
-                    # load as audio
+                    audios.append(result)
-                    audio_file = audio_data[audio_index]
-                    audio = load_audio(audio_file)
-                    hashes += [hash(audio_file)]
-                    audios += [audio]
-                    audio_index += 1
                     new_text += multimodal_tokens.audio_token
-                else:
+                # TODO: handle video
-                    # TODO(mick): handle video
+            else:
-                    # normal text
+                new_text += text_part
-                    new_text += text_part
-
-            except Exception as e:
-                logger.error(f"An exception occurred while loading images: {e}")
-                raise RuntimeError(f"An exception occurred while loading images: {e}")
 
         out = BaseMultiModalProcessorOutput(
             images=images,

python/sglang/srt/managers/multimodal_processors/janus_pro.py

@@ -33,7 +33,9 @@ class JanusProImageProcessor(BaseMultimodalProcessor):
         base_out = self.load_mm_data(
             prompt=input_ids,
             image_data=image_data,
-            multimodal_tokens=MultimodalSpecialTokens(image_token=processor.image_tag),
+            multimodal_tokens=MultimodalSpecialTokens(
+                image_token=processor.image_token
+            ),
             max_req_input_len=max_req_input_len,
         )
 

python/sglang/srt/managers/schedule_batch.py

@@ -222,10 +222,10 @@ class MultimodalDataItem:
                 # memoryview() doesn't support PyTorch's BFloat16 dtype
                 tensor = tensor.float()
 
+            assert isinstance(tensor, torch.Tensor)
             if tensor.is_cuda:
-                tensor_cpu = torch.frombuffer(
+                # TODO: improve this
-                    tensor.storage().untyped(), dtype=tensor.dtype, count=tensor.numel()
+                tensor_cpu = tensor.cpu()
-                ).clone()
             else:
                 tensor_cpu = tensor
 
@@ -321,7 +321,6 @@ class MultimodalInputs:
             item.set_pad_value()
 
         optional_args = [
-            "modalities",
             "im_token_id",
             "im_start_id",
             "im_end_id",

python/sglang/srt/managers/scheduler.py

@@ -452,6 +452,7 @@ class Scheduler(
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
                     revision=server_args.revision,
+                    use_fast=not server_args.disable_fast_image_processor,
                 )
                 self.tokenizer = self.processor.tokenizer
             else:

python/sglang/srt/managers/tokenizer_manager.py

@@ -180,6 +180,7 @@ class TokenizerManager:
                 tokenizer_mode=server_args.tokenizer_mode,
                 trust_remote_code=server_args.trust_remote_code,
                 revision=server_args.revision,
+                use_fast=not server_args.disable_fast_image_processor,
             )
 
             # We want to parallelize the image pre-processing so we create an executor for it

python/sglang/srt/models/qwen2_5_vl.py

@@ -462,6 +462,7 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                 quant_config=quant_config,
                 prefix=add_prefix("lm_head", prefix),
             )
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
 
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
@@ -515,15 +516,14 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                 otherwise it will be `(seq_len,).
                 (Use input_metadata.mrope_positions to replace it)
         """
-        is_mrope_enabled = "mrope_section" in self.config.rope_scaling
+        if self.is_mrope_enabled:
-        if is_mrope_enabled:
             positions = forward_batch.mrope_positions
 
         if not (
             forward_batch.forward_mode.is_decode()
             or not forward_batch.contains_image_inputs()
         ):
-            if is_mrope_enabled:
+            if self.is_mrope_enabled:
                 assert positions.ndim == 2 and positions.size(0) == 3, (
                     "multimodal section rotary embedding requires "
                     f"(3, seq_len) positions, but got {positions.size()}"

python/sglang/srt/models/qwen2_vl.py

@@ -467,6 +467,7 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 prefix=add_prefix("lm_head", prefix),
             )
 
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
@@ -521,15 +522,14 @@ class Qwen2VLForConditionalGeneration(nn.Module):
                 otherwise it will be `(seq_len,).
                 (Use input_metadata.mrope_positions to replace it)
         """
-        is_mrope_enabled = "mrope_section" in self.config.rope_scaling
+        if self.is_mrope_enabled:
-        if is_mrope_enabled:
             positions = forward_batch.mrope_positions
 
         if not (
             forward_batch.forward_mode.is_decode()
             or not forward_batch.contains_image_inputs()
         ):
-            if is_mrope_enabled:
+            if self.is_mrope_enabled:
                 assert positions.ndim == 2 and positions.size(0) == 3, (
                     "multimodal section rotary embedding requires "
                     f"(3, seq_len) positions, but got {positions.size()}"

python/sglang/srt/server_args.py

@@ -196,6 +196,9 @@ class ServerArgs:
     disaggregation_mode: str = "null"
     disaggregation_bootstrap_port: int = 8998
 
+    # multimodal
+    disable_fast_image_processor: bool = False
+
     def __post_init__(self):
         # Expert parallelism
         if self.enable_ep_moe:
@@ -979,6 +982,7 @@ class ServerArgs:
         )
         parser.add_argument(
             "--enable-llama4-multimodal",
+            default=ServerArgs.enable_llama4_multimodal,
             action="store_true",
             help="Enable the multimodal functionality for Llama-4.",
         )
@@ -1170,6 +1174,13 @@ class ServerArgs:
             help="Bootstrap server port on the prefill server. Default is 8998.",
         )
 
+        # Multimodal
+        parser.add_argument(
+            "--disable-fast-image-processor",
+            action="store_true",
+            help="Adopt base image processor instead of fast image processor.",
+        )
+
     @classmethod
     def from_cli_args(cls, args: argparse.Namespace):
         args.tp_size = args.tensor_parallel_size