sglang 0.5.3rc2__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/models/utils.py

@@ -27,7 +27,11 @@ if _is_cuda:
 
 def enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
     """Enable fused set_kv_buffer only on CUDA with bfloat16 KV cache."""
-    return _is_cuda and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
+    return (
+        _is_cuda
+        and hasattr(forward_batch.token_to_kv_pool, "dtype")
+        and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
+    )
 
 
 def create_fused_set_kv_buffer_arg(
@@ -49,3 +53,9 @@ def create_fused_set_kv_buffer_arg(
         v_scale=layer.v_scale,
         cache_loc=forward_batch.out_cache_loc,
     )
+
+
+def permute_inv(perm: torch.Tensor) -> torch.Tensor:
+    inv_perm = torch.empty_like(perm)
+    inv_perm[perm] = torch.arange(perm.numel(), device=perm.device, dtype=perm.dtype)
+    return inv_perm

sglang/srt/multimodal/processors/base_processor.py

@@ -155,7 +155,6 @@ class BaseMultimodalProcessor(ABC):
     ):
         self.hf_config = hf_config
         self._processor = _processor
-        self.arch = hf_config.architectures[0]
         self.server_args = server_args
         self.transport_mode = transport_mode
 
@@ -179,6 +178,7 @@ class BaseMultimodalProcessor(ABC):
             "image_attention_mask": Modality.IMAGE,
             "image_emb_mask": Modality.IMAGE,
             "images_spatial_crop": Modality.IMAGE,
+            "images_crop": Modality.IMAGE,
             "tgt_size": Modality.IMAGE,
             "image_grid_hws": Modality.IMAGE,
             "aspect_ratio_ids": Modality.IMAGE,
@@ -191,6 +191,7 @@ class BaseMultimodalProcessor(ABC):
             "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,
@@ -222,6 +223,7 @@ class BaseMultimodalProcessor(ABC):
             if self._processor.__class__.__name__ in {
                 "Gemma3nProcessor",
                 "Qwen2AudioProcessor",
+                "Qwen3OmniMoeProcessor",
             }:
                 # Note(Xinyuan): for gemma3n, ref: https://github.com/huggingface/transformers/blob/ccf2ca162e33f381e454cdb74bf4b41a51ab976d/src/transformers/models/gemma3n/processing_gemma3n.py#L107
                 kwargs["audio"] = audios
@@ -312,7 +314,9 @@ class BaseMultimodalProcessor(ABC):
         try:
             if modality == Modality.IMAGE:
                 img, _ = load_image(data)
-                return img.convert("RGB") if discard_alpha_channel else img
+                if discard_alpha_channel and img.mode != "RGB":
+                    img = img.convert("RGB")
+                return img
             elif modality == Modality.VIDEO:
                 return load_video(data, frame_count_limit)
             elif modality == Modality.AUDIO:

sglang/srt/multimodal/processors/deepseek_ocr.py

@@ -0,0 +1,37 @@
+from typing import List, Union
+
+from sglang.srt.models.deepseek_ocr import DeepseekOCRForCausalLM
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor,
+    MultimodalSpecialTokens,
+)
+
+
+class DeepseekOCRProcessor(BaseMultimodalProcessor):
+    models = [DeepseekOCRForCausalLM]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        _processor.image_size = 640
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+        self.mm_tokens = MultimodalSpecialTokens(
+            image_token="<image>", image_token_id=self._processor.image_token_id
+        ).build(_processor)
+
+    async def process_mm_data_async(
+        self, image_data: List[Union[str, bytes]], input_text, *args, **kwargs
+    ):
+        base_output = self.load_mm_data(
+            prompt=input_text,
+            multimodal_tokens=self.mm_tokens,
+            image_data=image_data,
+        )
+
+        mm_items, input_ids, _ = self.process_and_combine_mm_data(
+            base_output, self.mm_tokens
+        )
+
+        return {
+            "input_ids": input_ids.tolist(),
+            "mm_items": mm_items,
+            "im_token_id": self.mm_tokens.image_token_id,
+        }

sglang/srt/multimodal/processors/deepseek_vl_v2.py

@@ -18,9 +18,6 @@
 # CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 from typing import List, Union
 
-import torch
-
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.deepseek_vl2 import DeepseekVL2ForCausalLM
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,

sglang/srt/multimodal/processors/dots_vlm.py

@@ -1,5 +1,4 @@
 import asyncio
-import math
 import re
 from typing import Dict, List, Union
 

sglang/srt/multimodal/processors/glm4v.py

@@ -1,4 +1,3 @@
-import re
 from typing import List, Union
 
 from decord import VideoReader
@@ -9,10 +8,7 @@ from sglang.srt.models.glm4v_moe import Glm4vMoeForConditionalGeneration
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor as SGLangBaseProcessor,
 )
-from sglang.srt.multimodal.processors.base_processor import (
-    BaseMultiModalProcessorOutput,
-    MultimodalSpecialTokens,
-)
+from sglang.srt.multimodal.processors.base_processor import MultimodalSpecialTokens
 
 
 class Glm4vImageProcessor(SGLangBaseProcessor):

sglang/srt/multimodal/processors/internvl.py

@@ -4,10 +4,8 @@ from functools import lru_cache
 
 import numpy as np
 import torch
-import torchvision.transforms as T
 from decord import VideoReader, cpu, gpu
 from PIL import Image
-from torchvision.transforms import InterpolationMode
 
 from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.interns1 import InternS1ForConditionalGeneration

sglang/srt/multimodal/processors/janus_pro.py

@@ -1,6 +1,5 @@
 from typing import List, Union
 
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.deepseek_janus_pro import MultiModalityCausalLM
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,

sglang/srt/multimodal/processors/mllama4.py

@@ -1,13 +1,5 @@
 from typing import List, Union
 
-import torch
-from transformers.image_utils import SizeDict
-from transformers.models.llama4.image_processing_llama4_fast import (
-    find_supported_resolutions,
-    get_best_fit,
-)
-
-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,

sglang/srt/multimodal/processors/phi4mm.py

@@ -3,7 +3,6 @@ from typing import List, Union
 
 from transformers.processing_utils import ProcessorMixin
 
-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,

sglang/srt/multimodal/processors/points_v15_chat.py

@@ -0,0 +1,52 @@
+# Copy from qwen_vl.py, adapted for points-v15-chat
+
+import asyncio
+from typing import List, Union
+
+from PIL import Image
+
+from sglang.srt.models.points_v15_chat import POINTSV15ChatModel
+from sglang.srt.multimodal.processors.qwen_vl import (
+    Qwen2_5VLImageProcessor,
+    resize_image_async,
+)
+
+
+class POINTSV15ChatProcessor(Qwen2_5VLImageProcessor):
+    models = [POINTSV15ChatModel]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        # Compatible with POINTSV15Chat
+        hf_config.vision_start_token_id = None
+        hf_config.vision_end_token_id = None
+        hf_config.video_token_id = None
+
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+
+    async def process_mm_data_async(
+        self,
+        image_data: List[Union[str, bytes]],
+        input_text,
+        request_obj,
+        *args,
+        **kwargs,
+    ):
+        base_output = self.load_mm_data(
+            prompt=input_text,
+            image_data=image_data,
+            multimodal_tokens=self.mm_tokens,
+        )
+
+        if base_output.images and isinstance(base_output.images[0], Image.Image):
+            resize_tasks = [resize_image_async(image) for image in base_output.images]
+            base_output.images = await asyncio.gather(*resize_tasks)
+
+        mm_items, input_ids, _ = self.process_and_combine_mm_data(
+            base_output, self.mm_tokens
+        )
+
+        return {
+            "input_ids": input_ids.tolist(),
+            "mm_items": mm_items,
+            "im_token_id": self.mm_tokens.image_token_id,
+        }

sglang/srt/multimodal/processors/qwen_vl.py

@@ -9,9 +9,11 @@ import torchvision
 from PIL import Image
 from torchvision.transforms import InterpolationMode
 
+from sglang.srt.environ import envs
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
 from sglang.srt.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
 from sglang.srt.models.qwen2_vl import Qwen2VLForConditionalGeneration
+from sglang.srt.models.qwen3_omni_moe import Qwen3OmniMoeForConditionalGeneration
 from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
 from sglang.srt.models.qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration
 from sglang.srt.multimodal.processors.base_processor import (
@@ -22,8 +24,14 @@ from sglang.utils import logger
 
 IMAGE_FACTOR = 28
 MIN_PIXELS = 4 * 28 * 28
-MAX_PIXELS = 16384 * 28 * 28
+MAX_PIXELS = envs.SGLANG_IMAGE_MAX_PIXELS.get()
 MAX_RATIO = 200
+RESIZE_RESAMPLE = getattr(Image, envs.SGLANG_RESIZE_RESAMPLE.get(), None)
+if envs.SGLANG_RESIZE_RESAMPLE.is_set() and RESIZE_RESAMPLE is None:
+    logger.warning(
+        f"Invalid RESIZE_RESAMPLE value: '{envs.SGLANG_RESIZE_RESAMPLE.get()}'. "
+        f"Ignoring and using default."
+    )
 VIDEO_TOTAL_PIXELS = int(
     float(os.environ.get("VIDEO_MAX_PIXELS", 128000 * 28 * 28 * 0.9))
 )
@@ -85,7 +93,7 @@ def resize_image(
         min_pixels=min_pixels,
         max_pixels=max_pixels,
     )
-    image = image.resize((resized_width, resized_height))
+    image = image.resize((resized_width, resized_height), resample=RESIZE_RESAMPLE)
     return image
 
 
@@ -206,25 +214,41 @@ async def preprocess_video(
         interpolation=InterpolationMode.BICUBIC,
         antialias=True,
     ).float()
-    return video
-
-
-# Compatible with Qwen2VL and Qwen2_5VL
-class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
+    video_metadata = {
+        "fps": video_fps,
+        "duration": total_frames / video_fps,
+        "total_num_frames": total_frames,
+        "frames_indices": idx,
+        "video_backend": "torchvision",
+    }
+    return video, video_metadata
+
+
+# Compatible with Qwen-VL & Qwen-Omni Series
+class QwenVLImageProcessor(SGLangBaseProcessor):
     models = [
         Qwen2VLForConditionalGeneration,
         Qwen2_5_VLForConditionalGeneration,
         Qwen3VLForConditionalGeneration,
         Qwen3VLMoeForConditionalGeneration,
+        Qwen3OmniMoeForConditionalGeneration,
     ]
 
     def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        self.model_type = hf_config.model_type
+        if hf_config.model_type == "qwen3_omni_moe":
+            hf_config = hf_config.thinker_config
+
         super().__init__(hf_config, server_args, _processor, *args, **kwargs)
-        # The regex that matches expanded image tokens.
+
         self.IM_START_TOKEN_ID = hf_config.vision_start_token_id
         self.IM_END_TOKEN_ID = hf_config.vision_end_token_id
         self.vision_start_token_id = hf_config.vision_start_token_id
         self.vision_end_token_id = hf_config.vision_end_token_id
+
+        self.audio_start_token_id = getattr(hf_config, "audio_start_token_id", None)
+        self.audio_token_id = getattr(hf_config, "audio_token_id", None)
+
         self.NUM_TOKEN_PER_FRAME = 770
         self.IMAGE_FACTOR = 28
         self.MIN_PIXELS = 4 * 28 * 28
@@ -233,10 +257,12 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
         self.mm_tokens = MultimodalSpecialTokens(
             image_token="<|vision_start|><|image_pad|><|vision_end|>",
             image_token_id=hf_config.image_token_id,
+            # The regex that matches expanded image tokens.
             image_token_regex=re.compile(
                 r"<\|vision_start\|>(?:<\|image_pad\|>)+<\|vision_end\|>"
             ),
             video_token_id=hf_config.video_token_id,
+            audio_token_id=self.audio_token_id,
         ).build(_processor)
 
     async def process_mm_data_async(
@@ -247,11 +273,11 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
         *args,
         **kwargs,
     ):
-
         base_output = self.load_mm_data(
             prompt=input_text,
             image_data=image_data,
             video_data=request_obj.video_data,
+            audio_data=request_obj.audio_data,
             multimodal_tokens=self.mm_tokens,
         )
 
@@ -260,29 +286,61 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
             resize_tasks = [resize_image_async(image) for image in base_output.images]
             base_output.images = await asyncio.gather(*resize_tasks)
 
+        video_metadata = None
         if base_output.videos:
-            base_output.videos = [
-                await preprocess_video(video) for video in base_output.videos
-            ]
+            video_results = await asyncio.gather(
+                *[preprocess_video(video) for video in base_output.videos]
+            )
+            base_output.videos, video_metadata = map(list, zip(*video_results))
+
+        # NOTE: for qwen3-vl, video_meta need to be passed in, since do_sample_frames is already done in preprocess_video
+        if self.hf_config.model_type in ("qwen3_vl", "qwen3_vl_moe"):
+            mm_items, input_ids, ret = self.process_and_combine_mm_data(
+                base_output,
+                self.mm_tokens,
+                video_metadata=video_metadata,
+                do_sample_frames=False,
+            )
+        else:
+            mm_items, input_ids, ret = self.process_and_combine_mm_data(
+                base_output, self.mm_tokens
+            )
+
+        audio_feature_lengths = None
 
-        mm_items, input_ids, ret = self.process_and_combine_mm_data(
-            base_output, self.mm_tokens
+        if self.model_type == "qwen3_omni_moe":
+            audio_item = next((mm for mm in mm_items if mm.is_audio()), None)
+            if audio_item:
+                audio_feature_lengths = torch.sum(
+                    audio_item.feature_attention_mask, dim=1
+                )
+
+        second_per_grid_ts = getattr(ret, "second_per_grid_ts", None) or getattr(
+            ret, "video_second_per_grid", None
         )
 
         input_ids = input_ids.flatten()
+
         mrope_positions, mrope_position_delta = MRotaryEmbedding.get_rope_index(
             spatial_merge_size=self.hf_config.vision_config.spatial_merge_size,
             image_token_id=self.mm_tokens.image_token_id,
             video_token_id=self.mm_tokens.video_token_id,
             vision_start_token_id=self.vision_start_token_id,
-            model_type=self.hf_config.model_type,
+            model_type=self.model_type,
             tokens_per_second=getattr(
                 self.hf_config.vision_config, "tokens_per_second", None
             ),
             input_ids=input_ids.unsqueeze(0),
             image_grid_thw=getattr(ret, "image_grid_thw", None),
             video_grid_thw=getattr(ret, "video_grid_thw", None),
-            second_per_grid_ts=getattr(ret, "second_per_grid_ts", None),
+            second_per_grid_ts=second_per_grid_ts,
+            use_audio_in_video=False,
+            audio_seqlens=audio_feature_lengths,
+            audio_token_id=getattr(self.hf_config, "audio_token_id", None),
+            audio_start_token_id=self.audio_start_token_id,
+            position_id_per_seconds=getattr(
+                self.hf_config, "position_id_per_seconds", None
+            ),
         )
         mrope_positions = mrope_positions.squeeze(1)
 
@@ -293,6 +351,7 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
             "im_end_id": self.IM_END_TOKEN_ID,
             "im_token_id": self.mm_tokens.image_token_id,
             "video_token_id": self.mm_tokens.video_token_id,
+            "audio_token_id": self.mm_tokens.audio_token_id,
             "mrope_positions": mrope_positions,
             "mrope_position_delta": mrope_position_delta,
         }

sglang/srt/multimodal/processors/step3_vl.py

@@ -1,7 +1,7 @@
 import math
 import re
 from itertools import product
-from typing import List, Literal, Optional, TypedDict, Union
+from typing import List, Optional, Union
 
 import numpy as np
 import torch

sglang/srt/parser/conversation.py

@@ -838,6 +838,19 @@ register_conv_template(
     )
 )
 
+register_conv_template(
+    Conversation(
+        name="deepseek-ocr",
+        system_message="",
+        system_template="",
+        roles=("", ""),
+        sep="",
+        sep_style=SeparatorStyle.NO_COLON_SINGLE,
+        stop_str=["<｜end▁of▁sentence｜>"],
+        image_token="<image>",
+    )
+)
+
 register_conv_template(
     Conversation(
         name="deepseek-vl2",
@@ -960,6 +973,19 @@ register_conv_template(
     )
 )
 
+register_conv_template(
+    Conversation(
+        name="points-v15-chat",
+        system_message="",
+        system_template="",
+        roles=("<|im_start|>user", "<|im_start|>assistant"),
+        sep="<|im_end|>\n",
+        sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
+        stop_str=["<|im_end|>"],
+        image_token="<|vision_start|><|image_pad|><|vision_end|>",
+        video_token="<|vision_start|><|video_pad|><|vision_end|>",
+    )
+)
 
 MODEL_TYPE_TO_TEMPLATE = {
     "internvl_chat": "internvl-2-5",
@@ -968,9 +994,16 @@ MODEL_TYPE_TO_TEMPLATE = {
     "phi4mm": "phi-4-mm",
     "minicpmv": "minicpmv",
     "minicpmo": "minicpmo",
+    "deepseek-ocr": "deepseek-ocr",
 }
 
 
+@register_conv_template_matching_function
+def match_points_v15_chat(model_path: str):
+    if re.search(r"points", model_path, re.IGNORECASE):
+        return "points-v15-chat"
+
+
 def get_model_type(model_path: str) -> Optional[str]:
     config_path = os.path.join(model_path, "config.json")
     if not os.path.exists(config_path):
@@ -1038,3 +1071,11 @@ def match_phi_4_mm(model_path: str):
         return "phi-4-mm"
     model_type = get_model_type(model_path)
     return MODEL_TYPE_TO_TEMPLATE.get(model_type)
+
+
+@register_conv_template_matching_function
+def match_deepseek_ocr(model_path: str):
+    if "deepseek-ocr" in model_path.lower():
+        return "deepseek-ocr"
+    model_type = get_model_type(model_path)
+    return MODEL_TYPE_TO_TEMPLATE.get(model_type)

sglang/srt/parser/reasoning_parser.py

@@ -1,4 +1,3 @@
-import re
 from typing import Dict, Optional, Tuple, Type
 
 from sglang.srt.parser.harmony_parser import HarmonyParser

sglang/srt/sampling/custom_logit_processor.py

@@ -1,18 +1,22 @@
 import json
 from abc import ABC, abstractmethod
 from functools import lru_cache
-from typing import Any, Dict, List, Optional
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 import dill
+import orjson
 import torch
 
+if TYPE_CHECKING:
+    from sglang.srt.managers.schedule_batch import Req
+
 
 @lru_cache(maxsize=None)
 def _cache_from_str(json_str: str):
     """Deserialize a json string to a Callable object.
     This function is cached to avoid redundant deserialization.
     """
-    data = json.loads(json_str)
+    data = orjson.loads(json_str)
     return dill.loads(bytes.fromhex(data["callable"]))
 
 
@@ -51,3 +55,74 @@ class DisallowedTokensLogitsProcessor(CustomLogitProcessor):
         ), f"{custom_param_list=}"
         logits[..., disallowed_token_ids] = -float("inf")
         return logits
+
+
+class ThinkingBudgetLogitProcessor(CustomLogitProcessor):
+    """A logit processor that controls the length of thinking."""
+
+    THINKING_START_TOKEN_ID: int
+    THINKING_END_TOKEN_ID: int
+    NEW_LINE_TOKEN_ID: int
+
+    def __call__(self, logits, custom_param_list: list[dict[str, Any]]):
+        if custom_param_list is None or not custom_param_list:
+            return logits
+        for i, param_dict in enumerate(custom_param_list):
+            if param_dict is None:
+                continue
+
+            thinking_budget: int | None = param_dict.get("thinking_budget")
+
+            # Skip if thinking_budget is unset, or not an integer, or negative
+            if (
+                thinking_budget is None
+                or not isinstance(thinking_budget, int)
+                or thinking_budget < 0
+            ):
+                continue
+            req: Req = param_dict.get("__req__")
+            cur_ids: list[int] = [*req.origin_input_ids, *req.output_ids]
+
+            # Check if out of thinking stage
+            if (
+                self.THINKING_START_TOKEN_ID not in cur_ids
+                or self.THINKING_END_TOKEN_ID in cur_ids
+            ):
+                continue
+
+            # Find the index of the thinking start token
+            start_index = cur_ids.index(self.THINKING_START_TOKEN_ID)
+
+            # Count the number of tokens after the thinking start token
+            num_tokens_after_start = len(cur_ids) - start_index - 1
+
+            if num_tokens_after_start < thinking_budget:
+                continue
+
+            # Ensure new line token before thinking end token
+            if not req.output_ids or req.output_ids[-1] != self.NEW_LINE_TOKEN_ID:
+                logits[i, :] = -float("inf")
+                logits[i, self.NEW_LINE_TOKEN_ID] = 0.0
+                continue
+
+            # Assign highest probability to the thinking end token
+            logits[i, :] = -float("inf")
+            logits[i, self.THINKING_END_TOKEN_ID] = 0.0
+
+        return logits
+
+
+class Qwen3ThinkingBudgetLogitProcessor(ThinkingBudgetLogitProcessor):
+    """A logit processor that controls the length of thinking for Qwen3 models."""
+
+    THINKING_START_TOKEN_ID: int = 151667
+    THINKING_END_TOKEN_ID: int = 151668
+    NEW_LINE_TOKEN_ID: int = 198
+
+
+class DeepSeekR1ThinkingBudgetLogitProcessor(ThinkingBudgetLogitProcessor):
+    """A logit processor that controls the length of thinking for DeepSeek-R1 models."""
+
+    THINKING_START_TOKEN_ID: int = 128798
+    THINKING_END_TOKEN_ID: int = 128799
+    NEW_LINE_TOKEN_ID: int = 201