sglang 0.4.4.post3__py3-none-any.whl → 0.4.5__py3-none-any.whl

sglang/srt/managers/multimodal_processors/minicpm.py

@@ -1,13 +1,13 @@
-import asyncio
 from typing import List, Union
 
 import torch
+from transformers import BaseImageProcessorFast
 
 from sglang.srt.managers.multimodal_processors.base_processor import (
     BaseMultimodalProcessor,
     MultimodalSpecialTokens,
-    get_global_processor,
 )
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.minicpmo import MiniCPMO
 from sglang.srt.models.minicpmv import MiniCPMV
 
@@ -21,19 +21,23 @@ class MiniCPMMultimodalProcessor(BaseMultimodalProcessor):
         self.image_token = "(<image>./</image>)"
         self.audio_token = "(<audio>./</audio>)"
 
-    @staticmethod
-    def _process_data_task(input_text, images=None, audios=None):
+    def process_data_task(self, input_text, images=None, audios=None):
 
         if isinstance(images, list) and len(images) == 0:
             images = None
         if isinstance(audios, list) and len(audios) == 0:
             audios = None
-        result = get_global_processor().__call__(
+        processor = self._processor
+        args = {}
+        if isinstance(processor, BaseImageProcessorFast):
+            args["device"] = "cuda"
+        result = self._processor.__call__(
             text=input_text,
             images=images,
             audios=audios,
             return_tensors="pt",
             chunk_input=True,
+            **args,
         )
         return {
             "input_ids": result.input_ids,
@@ -44,23 +48,6 @@ class MiniCPMMultimodalProcessor(BaseMultimodalProcessor):
             "audio_bounds": getattr(result, "audio_bounds", None),
         }
 
-    async def _process_data(self, images, input_text, audios=None):
-        if self.executor is not None:
-            loop = asyncio.get_event_loop()
-            multimodal_data_inputs = await loop.run_in_executor(
-                self.executor,
-                MiniCPMMultimodalProcessor._process_data_task,
-                input_text,
-                images,
-                audios,
-            )
-        else:
-            multimodal_data_inputs = self._processor(
-                images=images, text=input_text, audios=audios, return_tensors="pt"
-            )
-
-        return multimodal_data_inputs
-
     async def process_mm_data_async(
         self,
         image_data: List[Union[str, bytes]],
@@ -77,7 +64,7 @@ class MiniCPMMultimodalProcessor(BaseMultimodalProcessor):
             audio_data = [audio_data]
 
         base_output = self.load_mm_data(
-            input_ids=input_ids,
+            prompt=input_ids,
             max_req_input_len=max_req_input_len,
             audio_data=audio_data,
             image_data=image_data,
@@ -88,9 +75,9 @@ class MiniCPMMultimodalProcessor(BaseMultimodalProcessor):
         if base_output is None:
             return None
 
-        res = await self._process_data(
-            images=base_output.images,
+        res = self.process_mm_data(
             input_text=base_output.input_text,
+            images=base_output.images,
             audios=base_output.audios,
         )
 
@@ -142,23 +129,33 @@ class MiniCPMMultimodalProcessor(BaseMultimodalProcessor):
                 tgt_sizes_flat += [tgt_n]
 
         pixel_values = pixel_values_flat
-        if len(tgt_sizes_flat) == 0:
-            tgt_sizes = None
-        else:
-            tgt_sizes = torch.stack(tgt_sizes_flat)
-        if not isinstance(res["audio_features"], list):
-            res["audio_features"] = [res["audio_features"]]
+
+        items = []
+        if len(pixel_values) != 0:
+            item = MultimodalDataItem(
+                pixel_values=pixel_values,
+                tgt_size=tgt_sizes_flat,
+                modality=Modality.IMAGE,
+            )
+            items += [item]
+
+        if (
+            "audio_features" in res
+            and res["audio_features"] is not None
+            and len(res["audio_features"]) != 0
+        ):
+            item = MultimodalDataItem(
+                audio_features=[res["audio_features"]],
+                audio_feature_lens=res["audio_feature_lens"],
+                modality=Modality.AUDIO,
+            )
+            items += [item]
+
         return {
+            "mm_items": items,
             "input_ids": res["input_ids"].flatten().tolist(),
-            "pixel_values": pixel_values,
-            "tgt_sizes": tgt_sizes,
-            "data_hashes": base_output.mm_data_hashes,
-            "modalities": request_obj.modalities or ["image"],
             "audio_start_id": audio_start_id,
             "audio_end_id": audio_end_id,
-            "audio_features": res["audio_features"],
-            "audio_bounds": res["audio_bounds"],
-            "audio_feature_lens": res["audio_feature_lens"],
             "im_token_id": im_token_id,
             "im_start_id": tokenizer.im_start_id,
             "im_end_id": tokenizer.im_end_id,

sglang/srt/managers/multimodal_processors/mlama.py

@@ -1,10 +1,9 @@
-import asyncio
 from typing import List, Union
 
 from sglang.srt.managers.multimodal_processors.base_processor import (
     BaseMultimodalProcessor,
-    get_global_processor,
 )
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.mllama import MllamaForConditionalGeneration
 from sglang.srt.utils import load_image
 
@@ -15,25 +14,6 @@ class MllamaImageProcessor(BaseMultimodalProcessor):
     def __init__(self, hf_config, server_args, _processor):
         super().__init__(hf_config, server_args, _processor)
 
-    @staticmethod
-    def _process_single_image_task(images, input_text):
-        # input_ids', 'attention_mask', 'pixel_values', 'aspect_ratio_ids', 'aspect_ratio_mask', 'cross_attention_mask'
-        return get_global_processor()(images, input_text, return_tensors="pt")
-
-    async def _process_single_image(self, images, input_text):
-        if self.executor is not None:
-            loop = asyncio.get_event_loop()
-            image_inputs = await loop.run_in_executor(
-                self.executor,
-                MllamaImageProcessor._process_single_image_task,
-                images,
-                input_text,
-            )
-        else:
-            image_inputs = self._processor(images, input_text, return_tensors="pt")
-
-        return image_inputs
-
     async def process_mm_data_async(
         self, image_data: List[Union[str, bytes]], input_text, *args, **kwargs
     ):
@@ -52,8 +32,15 @@ class MllamaImageProcessor(BaseMultimodalProcessor):
         else:
             images = load_image(image_data[0])[0]
 
-        image_inputs = await self._process_single_image(images, input_text)
-        image_inputs["data_hashes"] = [hash(str(image_data))]
+        image_inputs = self.process_mm_data(input_text=input_text, images=images)
         image_inputs["input_ids"] = image_inputs["input_ids"].tolist()[0]
+        image_inputs["mm_items"] = [
+            MultimodalDataItem(
+                pixel_values=image_inputs["pixel_values"],
+                aspect_ratio_id=image_inputs["aspect_ratio_ids"],
+                aspect_ratio_mask=image_inputs["aspect_ratio_mask"],
+                modality=Modality.IMAGE,
+            )
+        ]
 
         return image_inputs

sglang/srt/managers/multimodal_processors/mllama4.py

@@ -0,0 +1,161 @@
+from typing import List, Mapping, Optional, Tuple, Union
+
+import torch
+from PIL import Image
+from transformers import Llama4Processor
+from transformers.image_utils import SizeDict
+from transformers.models.llama4.image_processing_llama4 import (
+    find_supported_resolutions,
+    get_best_fit,
+)
+
+from sglang.srt.managers.multimodal_processors.base_processor import (
+    BaseMultimodalProcessor,
+    MultimodalSpecialTokens,
+)
+from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.models.mllama4 import Llama4ForConditionalGeneration
+from sglang.srt.utils import load_image
+
+
+class Mllama4ImageProcessor(BaseMultimodalProcessor):
+    models = [Llama4ForConditionalGeneration]
+
+    def __init__(self, hf_config, server_args, _processor):
+        super().__init__(hf_config, server_args, _processor)
+        self.vision_config = hf_config.vision_config
+        self.text_config = hf_config.text_config
+        self.multimodal_tokens = MultimodalSpecialTokens(
+            image_token=_processor.image_token
+        )
+
+    async def process_mm_data_async(
+        self,
+        image_data: List[Union[str, bytes]],
+        input_text,
+        max_req_input_len=None,
+        *args,
+        **kwargs,
+    ):
+        if not image_data:
+            return None
+
+        if isinstance(input_text, list):
+            assert len(input_text) and isinstance(input_text[0], int)
+            input_text = self._processor.tokenizer.decode(input_text)
+
+        # Process images and text using the base processor's load_mm_data method
+        processed_data = self.load_mm_data(
+            prompt=input_text,
+            multimodal_tokens=self.multimodal_tokens,
+            max_req_input_len=max_req_input_len or 4096,
+            image_data=image_data,
+            return_text=True,
+        )
+
+        # Process the images using the processor
+        processor = Llama4Processor.from_pretrained(
+            self.server_args.model_path, **kwargs
+        )
+
+        # Process the prompt and images
+        image_inputs = processor(
+            text=processed_data.input_text,
+            images=processed_data.images,
+            return_tensors="pt",
+        )
+
+        # Handle image resolutions and aspect ratios
+        if "pixel_values" in image_inputs:
+            image_processor = processor.image_processor
+            tokenizer = self._processor.tokenizer
+
+            # Calculate tile size and find supported resolutions
+            tile_size = self.vision_config.image_size
+            max_num_tiles = getattr(self.vision_config, "max_patches", 1)
+
+            possible_resolutions = find_supported_resolutions(
+                max_num_chunks=max_num_tiles,
+                patch_size=SizeDict(height=tile_size, width=tile_size),
+            )
+
+            # Find best fit for each image
+            best_fit_sizes = [
+                get_best_fit(
+                    (image.size[1], image.size[0]),  # (height, width)
+                    torch.tensor(possible_resolutions),
+                    resize_to_max_canvas=image_processor.resize_to_max_canvas,
+                )
+                for image in processed_data.images
+            ]
+
+            # Calculate aspect ratios and patches per image
+            aspect_ratios = [
+                (image_size[0] // tile_size, image_size[1] // tile_size)
+                for image_size in best_fit_sizes
+            ]
+
+            patches_per_image = [
+                1 if r_h * r_w == 1 else 1 + r_h * r_w for (r_h, r_w) in aspect_ratios
+            ]
+
+            # Add to image_inputs
+            image_inputs["aspect_ratios"] = aspect_ratios
+            image_inputs["patches_per_image"] = torch.tensor(patches_per_image)
+
+            # Process embed_is_patch
+            vocab = tokenizer.get_vocab()
+            patch_id = vocab.get(processor.img_patch_token, -1)
+            image_end_id = vocab.get(processor.end_of_img_token, -1)
+
+            if patch_id != -1 and image_end_id != -1:
+                input_ids = image_inputs["input_ids"].view(-1)
+
+                # Remove BOS token if present
+                if input_ids.size(0) > 0 and input_ids[0] == tokenizer.bos_token_id:
+                    input_ids = input_ids[1:]
+
+                # Find image end indices and split input_ids
+                image_end_indices = (input_ids == image_end_id).nonzero().view(-1)
+
+                if image_end_indices.size(0) > 0:
+                    # Split at image boundaries
+                    split_indices = (image_end_indices + 1)[:-1]
+                    split_input_ids = torch.tensor_split(input_ids, split_indices)
+                    split_input_ids = [x for x in split_input_ids if x.numel() > 0]
+
+                    # Create embed_is_patch for each image
+                    embed_is_patch = []
+                    for per_image_input_ids in split_input_ids:
+                        embed_is_patch.append(per_image_input_ids == patch_id)
+
+                    image_inputs["embed_is_patch"] = embed_is_patch
+
+        # Convert to the format expected by SGLang
+        image_inputs["input_ids"] = image_inputs["input_ids"].tolist()[0]
+
+        # Add metadata for image processing
+        image_inputs["mm_items"] = [
+            MultimodalDataItem(
+                pixel_values=image_inputs["pixel_values"],
+                modality=Modality.IMAGE,
+                # Add additional metadata needed for Llama4 vision processing
+                embed_is_patch=image_inputs.get("embed_is_patch", None),
+                aspect_ratios=image_inputs.get("aspect_ratios", None),
+                patches_per_image=image_inputs.get("patches_per_image", None),
+            )
+        ]
+
+        return image_inputs
+
+    def get_patch_per_chunk(self):
+        """Calculate patches per chunk based on vision config"""
+        image_size = self.vision_config.image_size
+        patch_size = self.vision_config.patch_size
+
+        assert (
+            image_size % patch_size == 0
+        ), f"chunk size {image_size} should be multiple of patch_size {patch_size}"
+
+        ds_ratio = int(round(1.0 / (self.vision_config.pixel_shuffle_ratio**2)))
+        return (image_size // patch_size) ** 2 // ds_ratio

sglang/srt/managers/multimodal_processors/qwen_vl.py

@@ -1,18 +1,17 @@
 import asyncio
 import math
-import time
 from typing import List, Union
 
 import torch
 from PIL import Image
 
-from sglang.srt.managers.multimodal_processor import (
+from sglang.srt.managers.multimodal_processors.base_processor import (
     BaseMultimodalProcessor as SGLangBaseProcessor,
 )
 from sglang.srt.managers.multimodal_processors.base_processor import (
     MultimodalSpecialTokens,
-    get_global_processor,
 )
+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
 
@@ -34,45 +33,15 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
         self.MAX_PIXELS = 16384 * 28 * 28
         self.MAX_RATIO = 200
 
-    @staticmethod
-    def _process_images_task(images, input_text, _hf_config):
-        if isinstance(images, list) and len(images) == 0:
-            images = None
-        result = get_global_processor().__call__(
-            text=[input_text], images=images, padding=True, return_tensors="pt"
-        )
-
-        return {
-            "input_ids": result.input_ids,
-            "pixel_values": getattr(result, "pixel_values", None),
-            "image_grid_thw": getattr(result, "image_grid_thw", None),
-            "second_per_grid_ts": getattr(result, "second_per_grid_ts", None),
-            "video_grid_thws": getattr(result, "video_grid_thws", None),
-        }
-
-    async def _process_single_image(self, images, input_text) -> dict:
-        if self.executor is not None:
-            loop = asyncio.get_event_loop()
-            return await loop.run_in_executor(
-                self.executor,
-                Qwen2_5VLImageProcessor._process_images_task,
-                images,
-                input_text,
-                self.hf_config,
-            )
-        else:
-            return self._process_images_task(images, input_text, self.hf_config)
-
     async def process_mm_data_async(
         self,
         image_data: List[Union[str, bytes]],
-        input_ids,
+        prompt,
         request_obj,
         max_req_input_len,
         *args,
         **kwargs,
     ):
-        start = time.time()
         if not image_data:
             return None
         if isinstance(image_data, str):
@@ -80,7 +49,7 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
 
         image_token = self.IMAGE_TOKEN
         base_output = self.load_mm_data(
-            input_ids=input_ids,
+            prompt=prompt,
             image_data=image_data,
             multimodal_tokens=MultimodalSpecialTokens(image_token=image_token),
             max_req_input_len=max_req_input_len,
@@ -144,24 +113,32 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
             """Returns the largest integer less than or equal to 'number' that is divisible by 'factor'."""
             return math.floor(number / factor) * factor
 
-        images = [resize_image(image) for image in base_output.images]
+        async def resize_image_async(image):
+            return resize_image(image)
 
-        ret = await self._process_single_image(
-            images=images, input_text=base_output.input_text
+        resize_tasks = [resize_image_async(image) for image in base_output.images]
+        resized_images = await asyncio.gather(*resize_tasks)
+
+        ret = self.process_mm_data(
+            input_text=base_output.input_text,
+            images=resized_images,
         )
 
         image_grid_thws = torch.concat([ret["image_grid_thw"]])
-        video_grid_thws = None
         return {
             "input_ids": ret["input_ids"].flatten().tolist(),
-            "pixel_values": ret["pixel_values"],
-            "data_hashes": base_output.mm_data_hashes,
-            "modalities": request_obj.modalities or ["image"],
-            "image_grid_thws": image_grid_thws,
-            "video_grid_thws": video_grid_thws,
+            "mm_items": [
+                MultimodalDataItem(
+                    pixel_values=ret["pixel_values"],
+                    image_grid_thws=image_grid_thws,
+                    # TODO
+                    video_grid_thws=None,
+                    second_per_grid_ts=ret.get("second_per_grid_ts", None),
+                    modality=Modality.IMAGE,
+                )
+            ],
             "im_start_id": self.IM_START_TOKEN_ID,
             "im_end_id": self.IM_END_TOKEN_ID,
             "im_token_id": self.image_token_id,
             "video_token_id": self.video_token_id,
-            "second_per_grid_ts": ret["second_per_grid_ts"],
         }