sglang 0.4.9.post1__py3-none-any.whl → 0.4.9.post2__py3-none-any.whl

sglang/srt/multimodal/processors/pixtral.py

@@ -106,7 +106,7 @@ class PixtralProcessor(BaseMultimodalProcessor):
                     pixel_values=processor_output["pixel_values"],
                     image_sizes=processor_output["image_sizes"],
                     modality=Modality.IMAGE,
-                    image_offsets=image_offsets,
+                    offsets=image_offsets,
                 )
             ]
 

sglang/srt/multimodal/processors/qwen_vl.py

@@ -1,9 +1,13 @@
 import asyncio
 import math
+import os
 import re
-from typing import Dict, List, Union
+from typing import List, Union
 
+import torch
+import torchvision
 from PIL import Image
+from torchvision.transforms import InterpolationMode
 
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
 from sglang.srt.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
@@ -12,6 +16,185 @@ from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor as SGLangBaseProcessor,
 )
 from sglang.srt.multimodal.processors.base_processor import MultimodalSpecialTokens
+from sglang.utils import logger
+
+IMAGE_FACTOR = 28
+MIN_PIXELS = 4 * 28 * 28
+MAX_PIXELS = 16384 * 28 * 28
+MAX_RATIO = 200
+VIDEO_TOTAL_PIXELS = int(
+    float(os.environ.get("VIDEO_MAX_PIXELS", 128000 * 28 * 28 * 0.9))
+)
+
+VIDEO_MIN_PIXELS = 128 * 28 * 28
+VIDEO_MAX_PIXELS = 768 * 28 * 28
+FRAME_FACTOR = 2
+FPS = 2.0
+FPS_MIN_FRAMES = 4
+FPS_MAX_FRAMES = 768
+
+
+def smart_resize(
+    height: int,
+    width: int,
+    factor: int = IMAGE_FACTOR,
+    min_pixels: int = MIN_PIXELS,
+    max_pixels: int = MAX_PIXELS,
+) -> tuple[int, int]:
+    """
+    Rescales the image so that the following conditions are met:
+
+    1. Both dimensions (height and width) are divisible by 'factor'.
+
+    2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
+
+    3. The aspect ratio of the image is maintained as closely as possible.
+    """
+    if max(height, width) / min(height, width) > MAX_RATIO:
+        raise ValueError(
+            f"absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(height, width) / min(height, width)}"
+        )
+    h_bar = max(factor, round_by_factor(height, factor))
+    w_bar = max(factor, round_by_factor(width, factor))
+    if h_bar * w_bar > max_pixels:
+        beta = math.sqrt((height * width) / max_pixels)
+        h_bar = floor_by_factor(height / beta, factor)
+        w_bar = floor_by_factor(width / beta, factor)
+    elif h_bar * w_bar < min_pixels:
+        beta = math.sqrt(min_pixels / (height * width))
+        h_bar = ceil_by_factor(height * beta, factor)
+        w_bar = ceil_by_factor(width * beta, factor)
+    return h_bar, w_bar
+
+
+def resize_image(image, size_factor: int = IMAGE_FACTOR) -> Image.Image:
+    width, height = image.size
+    min_pixels = MIN_PIXELS
+    max_pixels = MAX_PIXELS
+    resized_height, resized_width = smart_resize(
+        height,
+        width,
+        factor=size_factor,
+        min_pixels=min_pixels,
+        max_pixels=max_pixels,
+    )
+    image = image.resize((resized_width, resized_height))
+    return image
+
+
+def round_by_factor(number: int, factor: int) -> int:
+    """Returns the closest integer to 'number' that is divisible by 'factor'."""
+    return round(number / factor) * factor
+
+
+def ceil_by_factor(number: int, factor: int) -> int:
+    """Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'."""
+    return math.ceil(number / factor) * factor
+
+
+def floor_by_factor(number: int, factor: int) -> int:
+    """Returns the largest integer less than or equal to 'number' that is divisible by 'factor'."""
+    return math.floor(number / factor) * factor
+
+
+async def resize_image_async(image):
+    return resize_image(image)
+
+
+def smart_nframes(
+    ele: dict,
+    total_frames: int,
+    video_fps: int | float,
+) -> int:
+    """calculate the number of frames for video used for model inputs.
+
+    Args:
+        ele (dict): a dict contains the configuration of video.
+            support either `fps` or `nframes`:
+                - nframes: the number of frames to extract for model inputs.
+                - fps: the fps to extract frames for model inputs.
+                    - min_frames: the minimum number of frames of the video, only used when fps is provided.
+                    - max_frames: the maximum number of frames of the video, only used when fps is provided.
+        total_frames (int): the original total number of frames of the video.
+        video_fps (int | float): the original fps of the video.
+
+    Raises:
+        ValueError: nframes should in interval [FRAME_FACTOR, total_frames].
+
+    Returns:
+        int: the number of frames for video used for model inputs.
+    """
+    assert not (
+        "fps" in ele and "nframes" in ele
+    ), "Only accept either `fps` or `nframes`"
+    if "nframes" in ele:
+        nframes = round_by_factor(ele["nframes"], FRAME_FACTOR)
+    else:
+        fps = ele.get("fps", FPS)
+        min_frames = ceil_by_factor(ele.get("min_frames", FPS_MIN_FRAMES), FRAME_FACTOR)
+        max_frames = floor_by_factor(
+            ele.get("max_frames", min(FPS_MAX_FRAMES, total_frames)), FRAME_FACTOR
+        )
+        nframes = total_frames / video_fps * fps
+        if nframes > total_frames:
+            logger.warning(
+                f"smart_nframes: nframes[{nframes}] > total_frames[{total_frames}]"
+            )
+        nframes = min(min(max(nframes, min_frames), max_frames), total_frames)
+        nframes = floor_by_factor(nframes, FRAME_FACTOR)
+    if not (FRAME_FACTOR <= nframes and nframes <= total_frames):
+        raise ValueError(
+            f"nframes should in interval [{FRAME_FACTOR}, {total_frames}], but got {nframes}."
+        )
+    return nframes
+
+
+# process video, qwen-specific
+async def preprocess_video(
+    vr,
+    image_factor: int = IMAGE_FACTOR,
+    # vr: VideoReader, image_factor: int = IMAGE_FACTOR
+) -> torch.Tensor:
+    ele = {}
+    total_frames, video_fps = len(vr), vr.get_avg_fps()
+    nframes = smart_nframes({}, total_frames=total_frames, video_fps=video_fps)
+    idx = torch.linspace(0, total_frames - 1, nframes).round().long().tolist()
+    video = vr.get_batch(idx).asnumpy()
+    video = torch.tensor(video).permute(0, 3, 1, 2)  # Convert to TCHW format
+    nframes, _, height, width = video.shape
+    min_pixels = ele.get("min_pixels", VIDEO_MIN_PIXELS)
+    total_pixels = ele.get("total_pixels", VIDEO_TOTAL_PIXELS)
+    max_pixels = max(
+        min(VIDEO_MAX_PIXELS, total_pixels / nframes * FRAME_FACTOR),
+        int(min_pixels * 1.05),
+    )
+    max_pixels_supposed = ele.get("max_pixels", max_pixels)
+    if max_pixels_supposed > max_pixels:
+        logger.warning(
+            f"The given max_pixels[{max_pixels_supposed}] exceeds limit[{max_pixels}]."
+        )
+    max_pixels = min(max_pixels_supposed, max_pixels)
+    if "resized_height" in ele and "resized_width" in ele:
+        resized_height, resized_width = smart_resize(
+            ele["resized_height"],
+            ele["resized_width"],
+            factor=image_factor,
+        )
+    else:
+        resized_height, resized_width = smart_resize(
+            height,
+            width,
+            factor=image_factor,
+            min_pixels=min_pixels,
+            max_pixels=max_pixels,
+        )
+    video = torchvision.transforms.functional.resize(
+        video,
+        [resized_height, resized_width],
+        interpolation=InterpolationMode.BICUBIC,
+        antialias=True,
+    ).float()
+    return video
 
 
 # Compatible with Qwen2VL and Qwen2_5VL
@@ -37,104 +220,44 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
         self.MIN_PIXELS = 4 * 28 * 28
         self.MAX_PIXELS = 16384 * 28 * 28
         self.MAX_RATIO = 200
+        # TODO(mick): move all MultimodalSpecialTokens initializations into processor init
+        self.mm_special_tokens = MultimodalSpecialTokens(
+            image_token=self.IMAGE_TOKEN,
+            image_token_regex=self.IMAGE_TOKEN_REGEX,
+            video_token=self.VIDEO_TOKEN_ID,
+        )
 
     async def process_mm_data_async(
         self,
-        image_data: List[Union[str, bytes, Dict]],
+        image_data: List[Union[str, bytes]],
         input_text,
         request_obj,
         max_req_input_len,
         *args,
         **kwargs,
     ):
+
         base_output = self.load_mm_data(
             prompt=input_text,
             image_data=image_data,
-            multimodal_tokens=MultimodalSpecialTokens(
-                image_token=self.IMAGE_TOKEN,
-                image_token_regex=self.IMAGE_TOKEN_REGEX,
-            ),
+            video_data=request_obj.video_data,
+            multimodal_tokens=self.mm_special_tokens,
             max_req_input_len=max_req_input_len,
         )
 
-        def smart_resize(
-            height: int,
-            width: int,
-            factor: int = self.IMAGE_FACTOR,
-            min_pixels: int = self.MIN_PIXELS,
-            max_pixels: int = self.MAX_PIXELS,
-        ) -> tuple[int, int]:
-            """
-            Rescales the image so that the following conditions are met:
-
-            1. Both dimensions (height and width) are divisible by 'factor'.
-
-            2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
-
-            3. The aspect ratio of the image is maintained as closely as possible.
-            """
-            if max(height, width) / min(height, width) > self.MAX_RATIO:
-                raise ValueError(
-                    f"absolute aspect ratio must be smaller than {self.MAX_RATIO}, got {max(height, width) / min(height, width)}"
-                )
-            h_bar = max(factor, round_by_factor(height, factor))
-            w_bar = max(factor, round_by_factor(width, factor))
-            if h_bar * w_bar > max_pixels:
-                beta = math.sqrt((height * width) / max_pixels)
-                h_bar = floor_by_factor(height / beta, factor)
-                w_bar = floor_by_factor(width / beta, factor)
-            elif h_bar * w_bar < min_pixels:
-                beta = math.sqrt(min_pixels / (height * width))
-                h_bar = ceil_by_factor(height * beta, factor)
-                w_bar = ceil_by_factor(width * beta, factor)
-            return h_bar, w_bar
-
-        def resize_image(image, size_factor: int = self.IMAGE_FACTOR) -> Image.Image:
-            width, height = image.size
-            min_pixels = self.MIN_PIXELS
-            max_pixels = self.MAX_PIXELS
-            resized_height, resized_width = smart_resize(
-                height,
-                width,
-                factor=size_factor,
-                min_pixels=min_pixels,
-                max_pixels=max_pixels,
-            )
-            image = image.resize((resized_width, resized_height))
-            return image
-
-        def round_by_factor(number: int, factor: int) -> int:
-            """Returns the closest integer to 'number' that is divisible by 'factor'."""
-            return round(number / factor) * factor
-
-        def ceil_by_factor(number: int, factor: int) -> int:
-            """Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'."""
-            return math.ceil(number / factor) * factor
-
-        def floor_by_factor(number: int, factor: int) -> int:
-            """Returns the largest integer less than or equal to 'number' that is divisible by 'factor'."""
-            return math.floor(number / factor) * factor
-
-        async def resize_image_async(image):
-            return resize_image(image)
-
         # Qwen-specific: resize images if they are raw Image objects
         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)
 
-        video_grid_thw = None  # TODO
-
-        mm_items, input_ids = self.process_and_combine_mm_data(base_output)
-
-        if not mm_items:
-            # Note(Xinyuan): This is the case where image loading fails.
-            return None
+        if base_output.videos:
+            base_output.videos = [
+                await preprocess_video(video) for video in base_output.videos
+            ]
 
-        combined_mm_item = mm_items[0]  # only image is supported for now
-        video_grid_thw = None  # TODO
-        second_per_grid_ts = getattr(combined_mm_item, "second_per_grid_ts", None)
+        mm_items, input_ids, ret = self.process_and_combine_mm_data(base_output)
 
+        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.IM_TOKEN_ID,
@@ -145,9 +268,9 @@ class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
                 self.hf_config.vision_config, "tokens_per_second", None
             ),
             input_ids=input_ids.unsqueeze(0),
-            image_grid_thw=combined_mm_item.image_grid_thw,
-            video_grid_thw=video_grid_thw,
-            second_per_grid_ts=second_per_grid_ts,
+            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),
         )
         mrope_positions = mrope_positions.squeeze(1)
 

sglang/srt/multimodal/processors/vila.py

@@ -57,7 +57,7 @@ class VILAMultimodalProcessor(BaseMultimodalProcessor):
             image_data=image_data,
         )
 
-        mm_items, input_ids = self.process_and_combine_mm_data(base_output)
+        mm_items, input_ids, _ = self.process_and_combine_mm_data(base_output)
 
         return {
             "input_ids": input_ids.tolist(),

sglang/srt/server_args.py

@@ -416,7 +416,7 @@ class ServerArgs:
         if self.enable_dp_lm_head:
             assert (
                 self.enable_dp_attention
-            ), "Please enable dp attention when setting enable_dp_attention. "
+            ), "Please enable dp attention when setting enable_dp_lm_head. "
 
         # DeepEP MoE
         if self.enable_deepep_moe:
@@ -1048,9 +1048,16 @@ class ServerArgs:
         parser.add_argument(
             "--tool-call-parser",
             type=str,
-            choices=["qwen25", "mistral", "llama3", "deepseekv3", "pythonic"],
+            choices=[
+                "qwen25",
+                "mistral",
+                "llama3",
+                "deepseekv3",
+                "pythonic",
+                "kimi_k2",
+            ],
             default=ServerArgs.tool_call_parser,
-            help="Specify the parser for handling tool-call interactions. Options include: 'qwen25', 'mistral', 'llama3', 'deepseekv3', and 'pythonic'.",
+            help="Specify the parser for handling tool-call interactions. Options include: 'qwen25', 'mistral', 'llama3', 'deepseekv3', 'pythonic', and 'kimi_k2'.",
         )
 
         # Data parallelism
@@ -1614,7 +1621,7 @@ class ServerArgs:
             "--disaggregation-transfer-backend",
             type=str,
             default=ServerArgs.disaggregation_transfer_backend,
-            choices=["mooncake", "nixl"],
+            choices=["mooncake", "nixl", "ascend"],
             help="The backend for disaggregation transfer. Default is mooncake.",
         )
         parser.add_argument(

sglang/srt/utils.py

@@ -197,7 +197,7 @@ def get_int_env_var(name: str, default: int = 0) -> int:
 
 
 def support_triton(backend: str) -> bool:
-    return backend not in ["torch_native", "intel_amx"]
+    return backend not in ["torch_native", "intel_amx", "ascend"]
 
 
 try:
@@ -728,33 +728,6 @@ def load_audio(audio_file: str, sr: int = 16000, mono: bool = True) -> np.ndarra
     return audio
 
 
-def encode_video(video_path, frame_count_limit=None):
-    # Lazy import because decord is not available on some arm platforms.
-    from decord import VideoReader, cpu
-
-    if not os.path.exists(video_path):
-        logger.error(f"Video {video_path} does not exist")
-        return []
-
-    if frame_count_limit == 0:
-        return []
-
-    def uniform_sample(l, n):
-        gap = len(l) / n
-        idxs = [int(i * gap + gap / 2) for i in range(n)]
-        return [l[i] for i in idxs]
-
-    vr = VideoReader(video_path, ctx=cpu(0))
-    sample_fps = round(vr.get_avg_fps() / 1)  # FPS
-    frame_indices = [i for i in range(0, len(vr), sample_fps)]
-    if frame_count_limit is not None and len(frame_indices) > frame_count_limit:
-        frame_indices = uniform_sample(frame_indices, frame_count_limit)
-
-    frames = vr.get_batch(frame_indices).asnumpy()
-    frames = [Image.fromarray(v.astype("uint8")) for v in frames]
-    return frames
-
-
 def load_image(
     image_file: Union[Image.Image, str, bytes],
 ) -> tuple[Image.Image, tuple[int, int]]:
@@ -774,9 +747,6 @@ def load_image(
     elif image_file.startswith("data:"):
         image_file = image_file.split(",")[1]
         image = Image.open(BytesIO(pybase64.b64decode(image_file, validate=True)))
-    elif image_file.startswith("video:"):
-        image_file = image_file.replace("video:", "")
-        image, image_size = decode_video_base64(image_file)
     elif isinstance(image_file, str):
         image = Image.open(BytesIO(pybase64.b64decode(image_file, validate=True)))
     else:
@@ -785,6 +755,61 @@ def load_image(
     return image, image_size
 
 
+def load_video(video_file: Union[str, bytes], use_gpu: bool = True):
+    # We import decord here to avoid a strange Segmentation fault (core dumped) issue.
+    from decord import VideoReader, cpu, gpu
+
+    try:
+        from decord.bridge import decord_bridge
+
+        ctx = gpu(0)
+        _ = decord_bridge.get_ctx_device(ctx)
+    except Exception:
+        ctx = cpu(0)
+
+    tmp_file = None
+    vr = None
+    try:
+        if isinstance(video_file, bytes):
+            tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
+            tmp_file.write(video_file)
+            tmp_file.close()
+            vr = VideoReader(tmp_file.name, ctx=ctx)
+        elif isinstance(video_file, str):
+            if video_file.startswith(("http://", "https://")):
+                timeout = int(os.getenv("REQUEST_TIMEOUT", "10"))
+                response = requests.get(video_file, stream=True, timeout=timeout)
+                response.raise_for_status()
+                tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
+                for chunk in response.iter_content(chunk_size=8192):
+                    tmp_file.write(chunk)
+                tmp_file.close()
+                vr = VideoReader(tmp_file.name, ctx=ctx)
+            elif video_file.startswith("data:"):
+                _, encoded = video_file.split(",", 1)
+                video_bytes = base64.b64decode(encoded)
+                tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
+                tmp_file.write(video_bytes)
+                tmp_file.close()
+                vr = VideoReader(tmp_file.name, ctx=ctx)
+            elif os.path.isfile(video_file):
+                vr = VideoReader(video_file, ctx=ctx)
+            else:
+                video_bytes = base64.b64decode(video_file)
+                tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp4")
+                tmp_file.write(video_bytes)
+                tmp_file.close()
+                vr = VideoReader(tmp_file.name, ctx=ctx)
+        else:
+            raise ValueError(f"Unsupported video input type: {type(video_file)}")
+
+        return vr
+
+    finally:
+        if tmp_file and os.path.exists(tmp_file.name):
+            os.unlink(tmp_file.name)
+
+
 def suppress_other_loggers():
     warnings.filterwarnings(
         "ignore", category=UserWarning, message="The given NumPy array is not writable"
@@ -2757,3 +2782,101 @@ def lru_cache_frozenset(maxsize=128):
         return wrapper
 
     return decorator
+
+
+def apply_module_patch(target_module, target_function, wrappers):
+    original_module, original_function = parse_module_path(
+        target_module, target_function, False
+    )
+
+    original_function_id = id(original_function)
+
+    candidate = original_function
+    for wrapper in wrappers:
+        candidate = wrapper(candidate)
+    if target_function is not None:
+        setattr(original_module, target_function, candidate)
+
+    for key, value in sys.modules.copy().items():
+        if (
+            target_function is not None
+            and hasattr(value, target_function)
+            and id(getattr(value, target_function)) == original_function_id
+        ):
+            setattr(value, target_function, candidate)
+
+
+def parse_module_path(module_path, function_name, create_dummy):
+    from importlib.machinery import ModuleSpec
+
+    def create_dummy_module(full_path, parent=None):
+        """Create and register a placeholder module"""
+        dummy = types.ModuleType(full_path)
+        dummy.__file__ = "vllm_ascend.dummy_module.py"
+        dummy.__spec__ = ModuleSpec(full_path, None)
+        sys.modules[full_path] = dummy
+        if parent:
+            setattr(parent, full_path.split(".")[-1], dummy)
+        return dummy
+
+    def create_placeholder_function(func_name):
+        """Create dummy function that raises when called"""
+
+        def placeholder(*args, **kwargs):
+            raise NotImplementedError(f"Function {func_name} is a placeholder")
+
+        placeholder.__name__ = func_name
+        return placeholder
+
+    modules = module_path.split(".")
+    current_module = None
+    processed_path = []
+
+    for idx, part in enumerate(modules):
+        current_path = ".".join(modules[: idx + 1])
+        parent_path = ".".join(modules[:idx]) if idx > 0 else None
+
+        try:
+            current_module = importlib.import_module(current_path)
+        except ModuleNotFoundError:
+            # Handle missing module
+            parent = importlib.import_module(parent_path) if parent_path else None
+            if parent and hasattr(parent, part):
+                # Use existing attribute from parent
+                current_module = getattr(parent, part)
+                # Check for early function resolution
+                if function_name and hasattr(current_module, function_name):
+                    return current_module, getattr(current_module, function_name)
+                if function_name and create_dummy:
+                    ph_func = create_placeholder_function(function_name)
+                    setattr(current_module, function_name, ph_func)
+                    return current_module, ph_func
+                if function_name:
+                    raise AttributeError(
+                        f"Function {function_name} missing in {current_path}"
+                    )
+            else:
+                if not create_dummy:
+                    raise
+                # Create and register dummy module
+                current_module = create_dummy_module(
+                    current_path,
+                    parent=(
+                        importlib.import_module(parent_path) if parent_path else None
+                    ),
+                )
+
+        processed_path.append(part)
+
+    # Final function handling
+    final_module = sys.modules[module_path]
+    if function_name is not None:
+        if not hasattr(final_module, function_name):
+            if create_dummy:
+                ph_func = create_placeholder_function(function_name)
+                setattr(final_module, function_name, ph_func)
+            else:
+                setattr(final_module, function_name, None)
+        return final_module, getattr(final_module, function_name)
+
+    return final_module, None

sglang/version.py

@@ -1 +1 @@
-__version__ = "0.4.9.post1"
+__version__ = "0.4.9.post2"

{sglang-0.4.9.post1.dist-info → sglang-0.4.9.post2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sglang
-Version: 0.4.9.post1
+Version: 0.4.9.post2
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -242,6 +242,7 @@ Requires-Dist: pynvml; extra == "runtime-common"
 Requires-Dist: pybase64; extra == "runtime-common"
 Requires-Dist: python-multipart; extra == "runtime-common"
 Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
+Requires-Dist: sentencepiece; extra == "runtime-common"
 Requires-Dist: soundfile==0.13.1; extra == "runtime-common"
 Requires-Dist: scipy; extra == "runtime-common"
 Requires-Dist: torchao==0.9.0; extra == "runtime-common"
@@ -249,10 +250,10 @@ Requires-Dist: transformers==4.53.0; extra == "runtime-common"
 Requires-Dist: timm==1.0.16; extra == "runtime-common"
 Requires-Dist: uvicorn; extra == "runtime-common"
 Requires-Dist: uvloop; extra == "runtime-common"
-Requires-Dist: xgrammar==0.1.20; extra == "runtime-common"
+Requires-Dist: xgrammar==0.1.21; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
-Requires-Dist: sgl-kernel==0.2.4; extra == "srt"
+Requires-Dist: sgl-kernel==0.2.5; extra == "srt"
 Requires-Dist: torch==2.7.1; extra == "srt"
 Requires-Dist: torchaudio==2.7.1; extra == "srt"
 Requires-Dist: torchvision==0.22.1; extra == "srt"