Return 400 instead of 500 for unfetchable or unparseable multimodal inputs (#31417)

Co-authored-by: hnyls2002 <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>

python/sglang/srt/disaggregation/encode_server.py

@@ -69,6 +69,7 @@ from sglang.srt.server_args import (
     set_global_server_args_for_scheduler,
 )
 from sglang.srt.utils import (
+    CLIENT_MEDIA_EXCEPTIONS,
     add_prometheus_middleware,
     configure_logger,
     load_audio,
@@ -624,6 +625,9 @@ class MMEncoder:
             elif modality == Modality.AUDIO:
                 return load_audio(data, self.model_audio_sr)
 
+        except CLIENT_MEDIA_EXCEPTIONS as e:
+            # Not ValueError: the DP envelope classifies by `.code`, which only MMError carries.
+            raise BadRequestError(f"Error while loading data {data}: {e}") from e
         except Exception as e:
             raise RuntimeError(f"Error while loading data {data}: {e}")
 

python/sglang/srt/multimodal/processors/base_processor.py

@@ -22,6 +22,7 @@ from sglang.srt.managers.schedule_batch import (
 from sglang.srt.multimodal.processors.executor import MultimodalProcessorExecutor
 from sglang.srt.runtime_context import get_server_args
 from sglang.srt.utils import (
+    CLIENT_MEDIA_EXCEPTIONS,
     envs,
     is_cpu,
     is_npu,
@@ -652,8 +653,7 @@ class BaseMultimodalProcessor(ABC):
             elif modality == Modality.AUDIO:
                 return load_audio(data, audio_sample_rate)
 
-        except ValueError as e:
-            # Bad input (e.g. invalid base64) -> 400, not 500.
+        except CLIENT_MEDIA_EXCEPTIONS as e:
             data_str = str(data)
             if len(data_str) > 100:
                 data_str = data_str[:100] + "..."

python/sglang/srt/utils/common.py

@@ -86,7 +86,7 @@ import torch
 import torch.distributed as dist
 import triton
 from packaging import version as pkg_version
-from PIL import Image
+from PIL import Image, UnidentifiedImageError
 from starlette.routing import Mount
 from torch import nn
 from torch.library import Library
@@ -1514,6 +1514,15 @@ def get_mm_http_session() -> requests.Session:
     return session
 
 
+# Raised by the loaders below when client-supplied media cannot be fetched or
+# decoded. ValueError is in the set because invalid base64 raises binascii.Error.
+CLIENT_MEDIA_EXCEPTIONS = (
+    ValueError,
+    UnidentifiedImageError,
+    requests.exceptions.RequestException,
+)
+
+
 def load_audio(
     audio_file: str, sr: Optional[int] = None, mono: bool = True
 ) -> np.ndarray:
@@ -1582,10 +1591,13 @@ def load_audio(
     import torch
     import torchaudio
 
-    if isinstance(source, bytes):
-        audio, original_sr = sf.read(BytesIO(source))
-    else:
-        audio, original_sr = sf.read(source)
+    try:
+        if isinstance(source, bytes):
+            audio, original_sr = sf.read(BytesIO(source))
+        else:
+            audio, original_sr = sf.read(source)
+    except sf.LibsndfileError as e:
+        raise ValueError(f"Could not decode audio: {e}") from e
 
     if mono and len(audio.shape) > 1:
         audio = np.mean(audio, axis=1)
@@ -1785,7 +1797,13 @@ def load_video(video_file: Union[str, bytes, VideoData], use_gpu: bool = True):
         raise ValueError(f"Unsupported video input type: {type(video_file)}")
 
     device = "cuda" if use_gpu else "cpu"
-    return VideoDecoderWrapper(source, device=device)
+    try:
+        return VideoDecoderWrapper(source, device=device)
+    except (ImportError, MemoryError):
+        raise  # missing backend / OOM is not a bad payload
+    except Exception as e:
+        # Broad on purpose: torchcodec raises RuntimeError, decord its own type.
+        raise ValueError(f"Could not decode video: {e}") from e
 
 
 def sample_video_frames(video, *, desired_fps: int, max_frames: int) -> list[int]: