[Cherry-Pick][Others] support arm env

fastdeploy/input/ernie4_5_vl_processor/__init__.py

@@ -16,13 +16,13 @@
 
 from .ernie4_5_vl_processor import Ernie4_5_VLProcessor
 from .process import DataProcessor, fancy_print
-from .process_video import read_video_decord
+from .process_video import read_video_paddlecodec
 from .utils.video_utils import VideoReaderWrapper
 
 __all__ = [
     "DataProcessor",
     "fancy_print",
     "VideoReaderWrapper",
-    "read_video_decord",
+    "read_video_paddlecodec",
     "Ernie4_5_VLProcessor",
 ]

fastdeploy/input/ernie4_5_vl_processor/process.py

@@ -38,7 +38,7 @@
 from fastdeploy.utils import data_processor_logger
 
 from .image_preprocessor.image_preprocessor_adaptive import AdaptiveImageProcessor
-from .process_video import read_frames_decord, read_video_decord
+from .process_video import read_frames_paddlecodec, read_video_paddlecodec
 from .utils.render_timestamp import render_frame_timestamp
 
 
@@ -630,7 +630,7 @@ def _extract_labels(self, outputs: Dict, tgts: List[str]) -> None:
         outputs["labels"] = labels
 
     def _load_and_process_video(self, url: str, item: Dict) -> List[Image.Image]:
-        reader, meta, path = read_video_decord(url, save_to_disk=False)
+        reader, meta, path = read_video_paddlecodec(url, save_to_disk=False)
 
         video_frame_args = dict()
         video_frame_args["fps"] = item.get("fps", self.fps)
@@ -641,7 +641,7 @@ def _load_and_process_video(self, url: str, item: Dict) -> List[Image.Image]:
 
         video_frame_args = self._set_video_frame_args(video_frame_args, meta)
 
-        frames_data, _, timestamps = read_frames_decord(
+        frames_data, _, timestamps = read_frames_paddlecodec(
             path,
             reader,
             meta,

fastdeploy/input/ernie4_5_vl_processor/process_video.py

@@ -27,8 +27,8 @@
 from .utils.video_utils import VideoReaderWrapper
 
 
-def read_video_decord(video_path, save_to_disk):
-    """get reader and meta by decord"""
+def read_video_paddlecodec(video_path, save_to_disk):
+    """get reader and meta by paddlecodec"""
     # video_path = get_downloadable(video_path, save_to_disk=save_to_disk)
     if isinstance(video_path, VideoReaderWrapper):
         video_reader = video_path
@@ -113,7 +113,7 @@ def get_frame_indices(
     return frame_indices
 
 
-def read_frames_decord(
+def read_frames_paddlecodec(
     video_path,
     video_reader,
     video_meta,
@@ -126,7 +126,7 @@ def read_frames_decord(
     frame_indices=None,
     tol=10,
 ):
-    """get frames by decord"""
+    """get frames by paddlecodec"""
 
     if frame_indices is None:
         frame_indices = get_frame_indices(

fastdeploy/input/ernie4_5_vl_processor/utils/video_utils.py

@@ -18,7 +18,10 @@
 import os
 from tempfile import NamedTemporaryFile as ntf
 
-import decord
+import numpy as np
+import paddle
+
+from fastdeploy.utils import get_logger
 
 try:
     # moviepy 1.0
@@ -27,6 +30,8 @@
     # moviepy 2.0
     import moviepy as mp
 
+logger = get_logger("video_utils")
+
 
 def is_gif(data: bytes) -> bool:
     """
@@ -35,19 +40,24 @@ def is_gif(data: bytes) -> bool:
     return data[:6] in (b"GIF87a", b"GIF89a")
 
 
-class VideoReaderWrapper(decord.VideoReader):
-    """
-    Solving memory leak bug
+class _NumpyFrame:
+    """Wrapper so that frame[idx].asnumpy() keeps working with paddlecodec."""
+
+    def __init__(self, array):
+        self._array = array
+
+    def asnumpy(self):
+        return self._array
 
-    https://github.com/dmlc/decord/issues/208
-    """
+
+class VideoReaderWrapper:
+    """paddlecodec VideoDecoder wrapper with GIF support."""
 
     def __init__(self, video_path, *args, **kwargs):
         with ntf(delete=True, suffix=".gif") as gif_file:
             gif_input = None
             self.original_file = None
             if isinstance(video_path, str):
-                self.original_file = video_path
                 if video_path.lower().endswith(".gif"):
                     gif_input = video_path
             elif isinstance(video_path, bytes):
@@ -70,14 +80,58 @@ def __init__(self, video_path, *args, **kwargs):
                 video_path = mp4_file.name
                 self.original_file = video_path
 
-            super().__init__(video_path, *args, **kwargs)
-            self.seek(0)
+            with paddle.use_compat_guard(enable=True, scope={"torchcodec"}):
+                try:
+                    import sys
+
+                    from torchcodec.decoders import VideoDecoder
+
+                    sys.modules["torchcodec"] = None
+                except (ImportError, RuntimeError) as e:
+                    logger.error(
+                        f"Failed to load 'torchcodec' backend via Paddle proxy.\n"
+                        f"  - Common Causes:\n"
+                        f"    1. Conflict with official 'torch' or 'torchcodec' packages.\n"
+                        f"    2. Missing FFmpeg libraries or System library mismatch (CXXABI).\n"
+                        f"  - Recommended Fix Steps:\n"
+                        f"    1. Install dependencies: `conda install ffmpeg -c conda-forge` or `apt-get update && apt-get install ffmpeg` \n"
+                        f"    2. Uninstall conflicts: `pip uninstall torchcodec paddlecodec -y`\n"
+                        f"    3. Reinstall packages: `pip install paddlecodec --force-reinstall`\n"
+                        f"  - If you encounter 'CXXABI' or 'libstdc++' errors, your system libraries might be outdated.\n"
+                        f"    Try prioritizing Conda libraries by running: `LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH python your_script.py`\n"
+                        f"  - Original Error: {e}"
+                    )
+                    raise
+                PADDLECODEC_NUM_THREADS = int(os.environ.get("PADDLECODEC_NUM_THREADS", 0))
+                self._decoder = VideoDecoder(
+                    video_path,
+                    seek_mode="exact",
+                    num_ffmpeg_threads=PADDLECODEC_NUM_THREADS,
+                    device=kwargs.get("device", "cpu"),
+                    dimension_order="NHWC",
+                )
+
+    def __len__(self):
+        return self._decoder.metadata.num_frames
 
     def __getitem__(self, key):
-        frames = super().__getitem__(key)
-        self.seek(0)
-        return frames
+        if isinstance(key, (int, np.integer)):
+            frame = self._decoder.get_frames_at(indices=[int(key)]).data[0]
+            return _NumpyFrame(frame.numpy())
+        if isinstance(key, slice):
+            indices = list(range(*key.indices(len(self))))
+        else:
+            indices = list(key) if not isinstance(key, list) else key
+        frames = self._decoder.get_frames_at(indices=indices).data
+        return _NumpyFrame(frames.numpy())
+
+    def get_avg_fps(self):
+        return self._decoder.metadata.average_fps
 
     def __del__(self):
-        if self.original_file and os.path.exists(self.original_file):
-            os.remove(self.original_file)
+        original_file = getattr(self, "original_file", None)
+        if original_file and os.path.exists(original_file):
+            try:
+                os.remove(original_file)
+            except OSError:
+                pass

fastdeploy/input/paddleocr_vl_processor/process.py

@@ -28,7 +28,7 @@
 from fastdeploy.entrypoints.chat_utils import parse_chat_messages
 from fastdeploy.input.mm_data_processor import MMBaseDataProcessor
 from fastdeploy.input.utils import IDS_TYPE_FLAG
-from fastdeploy.input.video_utils import read_video_decord
+from fastdeploy.input.video_utils import read_video_paddlecodec
 from fastdeploy.input.video_utils import sample_frames_paddleocr as sample_frames
 from fastdeploy.multimodal.hasher import MultimodalHasher
 from fastdeploy.utils import data_processor_logger
@@ -530,7 +530,7 @@ def _load_and_process_video(self, url: str, item: Dict) -> Tuple[np.ndarray, Dic
                 - frames: Processed video frames as numpy array
                 - metadata: Updated video metadata dictionary
         """
-        reader, meta, _ = read_video_decord(url, save_to_disk=False)
+        reader, meta, _ = read_video_paddlecodec(url, save_to_disk=False)
 
         # Apply frame sampling if fps or target_frames specified
         fps = item.get("fps", self.fps)

fastdeploy/input/qwen3_vl_processor/process.py

@@ -28,7 +28,7 @@
 from fastdeploy.entrypoints.chat_utils import parse_chat_messages
 from fastdeploy.input.mm_data_processor import MMBaseDataProcessor
 from fastdeploy.input.utils import IDS_TYPE_FLAG
-from fastdeploy.input.video_utils import read_video_decord
+from fastdeploy.input.video_utils import read_video_paddlecodec
 from fastdeploy.input.video_utils import sample_frames_qwen as sample_frames
 from fastdeploy.multimodal.hasher import MultimodalHasher
 from fastdeploy.utils import data_processor_logger
@@ -681,7 +681,7 @@ def _load_and_process_video(self, url: str, item: Dict) -> Tuple[np.ndarray, Dic
                 - frames: Processed video frames as numpy array
                 - metadata: Updated video metadata dictionary
         """
-        reader, meta, _ = read_video_decord(url, save_to_disk=False)
+        reader, meta, _ = read_video_paddlecodec(url, save_to_disk=False)
 
         # Apply frame sampling if fps or target_frames specified
         fps = item.get("fps", self.fps)

fastdeploy/input/qwen_vl_processor/process.py

@@ -28,7 +28,7 @@
 from fastdeploy.entrypoints.chat_utils import parse_chat_messages
 from fastdeploy.input.mm_data_processor import MMBaseDataProcessor
 from fastdeploy.input.utils import IDS_TYPE_FLAG
-from fastdeploy.input.video_utils import read_video_decord
+from fastdeploy.input.video_utils import read_video_paddlecodec
 from fastdeploy.input.video_utils import sample_frames_qwen as sample_frames
 from fastdeploy.multimodal.hasher import MultimodalHasher
 from fastdeploy.utils import data_processor_logger
@@ -531,7 +531,7 @@ def _load_and_process_video(self, url: str, item: Dict) -> Tuple[np.ndarray, Dic
                 - frames: Processed video frames as numpy array
                 - metadata: Updated video metadata dictionary
         """
-        reader, meta, _ = read_video_decord(url, save_to_disk=False)
+        reader, meta, _ = read_video_paddlecodec(url, save_to_disk=False)
 
         # Apply frame sampling if fps or target_frames specified
         fps = item.get("fps", self.fps)

fastdeploy/input/video_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Shared video utilities: VideoReaderWrapper, read_video_decord, and sample_frames."""
+"""Shared video utilities: VideoReaderWrapper, read_video_paddlecodec, and sample_frames."""
 
 import io
 import math
@@ -21,13 +21,16 @@
 from typing import Optional, Union
 
 import numpy as np
+import paddle
 
 from fastdeploy.input.image_processors.common import ceil_by_factor, floor_by_factor
-from fastdeploy.utils import data_processor_logger
+from fastdeploy.utils import data_processor_logger, get_logger
+
+logger = get_logger("video_utils")
 
 __all__ = [
     "VideoReaderWrapper",
-    "read_video_decord",
+    "read_video_paddlecodec",
     "sample_frames",
     "sample_frames_qwen",
     "sample_frames_paddleocr",
@@ -44,15 +47,20 @@ def _is_gif(data: bytes) -> bool:
     return data[:6] in (b"GIF87a", b"GIF89a")
 
 
-class VideoReaderWrapper:
-    """decord.VideoReader wrapper that fixes a memory leak and adds GIF support.
+class _NumpyFrame:
+    """Wrapper so that frame[idx].asnumpy() keeps working with paddlecodec."""
 
-    Reference: https://github.com/dmlc/decord/issues/208
-    """
+    def __init__(self, array):
+        self._array = array
+
+    def asnumpy(self):
+        return self._array
 
-    def __init__(self, video_path, *args, **kwargs):
-        import decord
 
+class VideoReaderWrapper:
+    """paddlecodec VideoDecoder wrapper with GIF support."""
+
+    def __init__(self, video_path, *args, **kwargs):
         try:
             # moviepy 1.0
             import moviepy.editor as mp
@@ -91,22 +99,53 @@ def __init__(self, video_path, *args, **kwargs):
                 video_path = mp4_path
                 self.original_file = video_path  # temp mp4, cleaned up in __del__
 
-            self._reader = decord.VideoReader(video_path, *args, **kwargs)
-            self._reader.seek(0)
+            with paddle.use_compat_guard(enable=True, scope={"torchcodec"}):
+                try:
+                    import sys
+
+                    from torchcodec.decoders import VideoDecoder
+
+                    sys.modules["torchcodec"] = None
+                except (ImportError, RuntimeError) as e:
+                    logger.error(
+                        f"Failed to load 'torchcodec' backend via Paddle proxy.\n"
+                        f"  - Common Causes:\n"
+                        f"    1. Conflict with official 'torch' or 'torchcodec' packages.\n"
+                        f"    2. Missing FFmpeg libraries or System library mismatch (CXXABI).\n"
+                        f"  - Recommended Fix Steps:\n"
+                        f"    1. Install dependencies: `conda install ffmpeg -c conda-forge` or `apt-get update && apt-get install ffmpeg` \n"
+                        f"    2. Uninstall conflicts: `pip uninstall torchcodec paddlecodec -y`\n"
+                        f"    3. Reinstall packages: `pip install paddlecodec --force-reinstall`\n"
+                        f"  - If you encounter 'CXXABI' or 'libstdc++' errors, your system libraries might be outdated.\n"
+                        f"    Try prioritizing Conda libraries by running: `LD_LIBRARY_PATH=$CONDA_PREFIX/lib:$LD_LIBRARY_PATH python your_script.py`\n"
+                        f"  - Original Error: {e}"
+                    )
+                    raise
+                PADDLECODEC_NUM_THREADS = int(os.environ.get("PADDLECODEC_NUM_THREADS", 0))
+                self._decoder = VideoDecoder(
+                    video_path,
+                    seek_mode="exact",
+                    num_ffmpeg_threads=PADDLECODEC_NUM_THREADS,
+                    device=kwargs.get("device", "cpu"),
+                    dimension_order="NHWC",
+                )
 
     def __len__(self):
-        return len(self._reader)
+        return self._decoder.metadata.num_frames
 
     def __getitem__(self, key):
-        frames = self._reader[key]
-        self._reader.seek(0)
-        return frames
+        if isinstance(key, (int, np.integer)):
+            frame = self._decoder.get_frames_at(indices=[int(key)]).data[0]
+            return _NumpyFrame(frame.numpy())
+        if isinstance(key, slice):
+            indices = list(range(*key.indices(len(self))))
+        else:
+            indices = list(key) if not isinstance(key, list) else key
+        frames = self._decoder.get_frames_at(indices=indices).data
+        return _NumpyFrame(frames.numpy())
 
     def get_avg_fps(self):
-        return self._reader.get_avg_fps()
-
-    def seek(self, pos):
-        return self._reader.seek(pos)
+        return self._decoder.metadata.average_fps
 
     def __del__(self):
         original_file = getattr(self, "original_file", None)
@@ -118,11 +157,11 @@ def __del__(self):
 
 
 # ---------------------------------------------------------------------------
-# read_video_decord
+# read_video_paddlecodec
 # ---------------------------------------------------------------------------
 
 
-def read_video_decord(video_path, save_to_disk: bool = False):
+def read_video_paddlecodec(video_path, save_to_disk: bool = False):
     """Load a video file and return (video_reader, video_meta, video_path).
 
     video_meta contains keys: "fps", "duration", "num_of_frame".

requirements.txt

@@ -14,7 +14,7 @@ paddleformers>=1.1.1
 redis
 etcd3
 httpx
-tool_helpers
+fast_dataindex
 cupy-cuda12x
 pybind11[global]
 tabulate
@@ -23,7 +23,7 @@ xlwt
 visualdl
 setuptools-scm>=8
 prometheus-client
-decord
+paddlecodec
 moviepy
 triton
 crcmod