Revert "lang: support direct video inference" (#12038)

python/sglang/utils.py

@@ -209,6 +209,52 @@ def encode_frame(frame):
     return frame_bytes
 
 
+def encode_video_base64(video_path: str, num_frames: int = 16):
+    import cv2  # pip install opencv-python-headless
+
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        raise IOError(f"Could not open video file:{video_path}")
+
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    print(f"target_frames: {num_frames}")
+
+    frame_indices = np.linspace(0, total_frames - 1, num_frames, dtype=int)
+
+    frames = []
+    for _ in range(total_frames):
+        ret, frame = cap.read()
+        if ret:
+            frames.append(frame)
+        else:
+            # Handle the case where the frame could not be read
+            # print(f"Warning: Could not read frame at index {i}.")
+            pass
+
+    cap.release()
+
+    # Safely select frames based on frame_indices, avoiding IndexError
+    frames = [frames[i] for i in frame_indices if i < len(frames)]
+
+    # If there are not enough frames, duplicate the last frame until we reach the target
+    while len(frames) < num_frames:
+        frames.append(frames[-1])
+
+    # Use ThreadPoolExecutor to process and encode frames in parallel
+    with ThreadPoolExecutor() as executor:
+        encoded_frames = list(executor.map(encode_frame, frames))
+
+    # encoded_frames = list(map(encode_frame, frames))
+
+    # Concatenate all frames bytes
+    video_bytes = b"".join(encoded_frames)
+
+    # Encode the concatenated bytes to base64
+    video_base64 = "video:" + pybase64.b64encode(video_bytes).decode("utf-8")
+
+    return video_base64
+
+
 def _is_chinese_char(cp: int):
     """Checks whether CP is the codepoint of a CJK character."""
     # This defines a "chinese character" as anything in the CJK Unicode block: