Move multimodal processors into a separate folder (#7581)

python/sglang/srt/managers/multimodal_processors/minicpm.py → python/sglang/srt/multimodal/processors/minicpm.py

@@ -2,13 +2,13 @@ from typing import List, Union
 
 import torch
 
-from sglang.srt.managers.multimodal_processors.base_processor import (
-    BaseMultimodalProcessor,
-    MultimodalSpecialTokens,
-)
 from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.minicpmo import MiniCPMO
 from sglang.srt.models.minicpmv import MiniCPMV
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor,
+    MultimodalSpecialTokens,
+)
 
 
 # Compatible with both 'O' and 'V'

python/sglang/srt/managers/multimodal_processors/mlama.py → python/sglang/srt/multimodal/processors/mlama.py

 from typing import List, Union
 
-from sglang.srt.managers.multimodal_processors.base_processor import (
-    BaseMultimodalProcessor,
-)
 from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.mllama import MllamaForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import BaseMultimodalProcessor
 from sglang.srt.utils import load_image
 
 

python/sglang/srt/managers/multimodal_processors/mllama4.py → python/sglang/srt/multimodal/processors/mllama4.py

@@ -7,12 +7,12 @@ from transformers.models.llama4.image_processing_llama4_fast import (
     get_best_fit,
 )
 
-from sglang.srt.managers.multimodal_processors.base_processor import (
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.models.mllama4 import Llama4ForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,
     MultimodalSpecialTokens,
 )
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
-from sglang.srt.models.mllama4 import Llama4ForConditionalGeneration
 
 
 class Mllama4ImageProcessor(BaseMultimodalProcessor):

python/sglang/srt/managers/multimodal_processors/phi4mm.py → python/sglang/srt/multimodal/processors/phi4mm.py

 import logging
 from typing import List, Union
 
-from sglang.srt.managers.multimodal_processors.base_processor import (
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.models.phi4mm import Phi4MMForCausalLM
+from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,
     MultimodalSpecialTokens,
 )
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
-from sglang.srt.models.phi4mm import Phi4MMForCausalLM
 
 logger = logging.getLogger(__name__)
 

python/sglang/srt/managers/multimodal_processors/pixtral.py → python/sglang/srt/multimodal/processors/pixtral.py

@@ -6,12 +6,12 @@ from transformers.models.pixtral.image_processing_pixtral import (
     _num_image_tokens as _get_pixtral_hf_num_image_tokens,
 )
 
-from sglang.srt.managers.multimodal_processors.base_processor import (
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.models.pixtral import PixtralVisionModel
+from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,
     MultimodalSpecialTokens,
 )
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
-from sglang.srt.models.pixtral import PixtralVisionModel
 
 
 class PixtralProcessor(BaseMultimodalProcessor):

python/sglang/srt/managers/multimodal_processors/qwen_vl.py → python/sglang/srt/multimodal/processors/qwen_vl.py

@@ -7,15 +7,13 @@ import torch
 from PIL import Image
 
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
-from sglang.srt.managers.multimodal_processors.base_processor import (
-    BaseMultimodalProcessor as SGLangBaseProcessor,
-)
-from sglang.srt.managers.multimodal_processors.base_processor import (
-    MultimodalSpecialTokens,
-)
 from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
 from sglang.srt.models.qwen2_vl import Qwen2VLForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor as SGLangBaseProcessor,
+)
+from sglang.srt.multimodal.processors.base_processor import MultimodalSpecialTokens
 
 
 # Compatible with Qwen2VL and Qwen2_5VL

python/sglang/srt/managers/multimodal_processors/vila.py → python/sglang/srt/multimodal/processors/vila.py

@@ -10,12 +10,12 @@ from sglang.srt.managers.io_struct import (
     GenerateReqInput,
     ImageDataItem,
 )
-from sglang.srt.managers.multimodal_processors.base_processor import (
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.models.vila import VILAForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,
     MultimodalSpecialTokens,
 )
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
-from sglang.srt.models.vila import VILAForConditionalGeneration
 from sglang.srt.server_args import ServerArgs
 
 

python/sglang/srt/utils.py

@@ -2577,3 +2577,13 @@ def configure_gc_logger():
             )
 
     gc.callbacks.append(gc_callback)
+
+
+# COPIED FROM DeepGEMM
+def align(x: int, y: int) -> int:
+    return ceil_div(x, y) * y
+
+
+# COPIED FROM DeepGEMM
+def ceil_div(x: int, y: int) -> int:
+    return (x + y - 1) // y

test/srt/test_vlm_accuracy.py

@@ -23,15 +23,13 @@ from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.conversation import generate_chat_conv
 from sglang.srt.entrypoints.openai.protocol import ChatCompletionRequest
 from sglang.srt.managers.mm_utils import embed_mm_inputs, init_embedding_cache
-from sglang.srt.managers.multimodal_processors.base_processor import (
-    BaseMultimodalProcessor,
-)
 from sglang.srt.managers.schedule_batch import (
     Modality,
     MultimodalDataItem,
     MultimodalInputs,
 )
 from sglang.srt.model_executor.model_runner import ModelRunner
+from sglang.srt.multimodal.processors.base_processor import BaseMultimodalProcessor
 from sglang.srt.server_args import ServerArgs