feat: Implement robust batch processing using isolated subprocesses to prevent VRAM fragmentation and OOM errors.

Author: artryazanov

.github/workflows/linting.yml

@@ -34,7 +34,9 @@ jobs:
         run: pip install -e .
 
       - name: Run Ruff
-        run: ruff check .
+        run: |
+          ruff check .
+          ruff check --select D --config 'lint.pydocstyle.convention="google"' src/
 
       - name: Run Mypy
         run: mypy src/shorts_maker

README.md

@@ -29,6 +29,7 @@ For the original CPU-only version, please visit [Shorts Maker](https://github.co
   - **Video Analysis**: Zero-copy GPU memory streaming for stable motion estimation (replaces heavy frame indices).
   - **Image Processing**: Native PyTorch operators used for heavy operations like blurring backgrounds (separable convolutions).
   - **Rendering**: Custom PyTorch+NVENC engine for high-performance rendering (MoviePy removed from render path).
+  - **Robust Batch Processing**: Video processing runs in fully isolated subprocesses, completely clearing CUDA contexts between files to prevent VRAM fragmentation and OOM crashes (especially in Docker/WSL).
 - Audio + video action scoring:
   - Combined ranking with tunable weights (defaults: audio 0.6, video 0.4).
 - Scenes ranked by combined action score rather than duration.

docs/index.md

@@ -26,6 +26,7 @@ For the original CPU-only version, please visit [Shorts Maker](https://github.co
   - **Video Analysis**: Zero-copy GPU memory streaming for stable motion estimation (replaces heavy frame indices).
   - **Image Processing**: Native PyTorch operators used for heavy operations like blurring backgrounds (separable convolutions).
   - **Rendering**: Custom PyTorch+NVENC engine for high-performance rendering (MoviePy removed from render path).
+  - **Robust Batch Processing**: Video processing runs in fully isolated subprocesses, completely clearing CUDA contexts between files to prevent VRAM fragmentation and OOM crashes (especially in Docker/WSL).
 - Audio + video action scoring:
   - Combined ranking with tunable weights (defaults: audio 0.6, video 0.4).
 - Scenes ranked by combined action score rather than duration.

src/shorts_maker/__init__.py

@@ -1,6 +1,4 @@
-"""
-Shorts Maker GPU: A high-performance video processing library.
-"""
+"""Shorts Maker GPU: A high-performance video processing library."""
 import logging
 
 from .config import ProcessingConfig

src/shorts_maker/analysis/__init__.py

@@ -0,0 +1 @@
+"""Audio and video analysis modules for generating action profiles."""

src/shorts_maker/analysis/audio.py

@@ -1,3 +1,5 @@
+"""Audio analysis module for computing an action profile based on loudness and spectral flux."""
+
 import logging
 from pathlib import Path
 from typing import Tuple

src/shorts_maker/analysis/video.py

@@ -1,3 +1,5 @@
+"""Video analysis module for computing an action profile from frame-by-frame pixel differences."""
+
 import logging
 from pathlib import Path
 from typing import Tuple
@@ -34,7 +36,6 @@ def compute_video_action_profile(
             - times (np.ndarray): Array of timestamps (in seconds) for each evaluated frame.
             - score (np.ndarray): Array of normalized, smoothed action scores.
     """
-
     # 1) Get metadata and calculate dimensions
     try:
         dmx = nvc.PyFFmpegDemuxer(str(video_path))

src/shorts_maker/cli.py

@@ -1,4 +1,7 @@
+"""Command Line Interface for Shorts Maker, orchestrating the entire processing pipeline."""
+
 import logging
+import multiprocessing
 from pathlib import Path
 
 import typer
@@ -9,6 +12,14 @@
 
 app = typer.Typer(help="GPU-accelerated shorts generator.")
 
+def _process_video_worker(config: ProcessingConfig, video_file: Path, output_dir: Path) -> None:
+    """Isolated worker for processing a single video.
+
+    Ensures that PyTorch and VPF memory is entirely cleared upon exit.
+    """
+    processor = VideoProcessor(config)
+    processor.process_video(video_file, output_dir)
+
 
 @app.command()
 def process(
@@ -53,15 +64,27 @@ def process(
         logger.warning(f"No '{input_dir}' directory found. Exiting.")
         raise typer.Exit(code=1)
 
-    processor = VideoProcessor(config)
-
     for video_file in input_dir.iterdir():
         if video_file.is_file() and video_file.suffix.lower() in [
             ".mp4",
             ".mkv",
             ".mov",
         ]:
-            processor.process_video(video_file, output_dir)
+            logger.info(f"\n--- Spawning isolated process for: {video_file.name} ---")
+            
+            # Using 'spawn' guarantees a clean process without inherited CUDA contexts
+            ctx = multiprocessing.get_context("spawn")
+            p = ctx.Process(
+                target=_process_video_worker, 
+                args=(config, video_file, output_dir)
+            )
+            p.start()
+            p.join()
+
+            if p.exitcode != 0:
+                logger.error(f"Processing failed for {video_file.name} with exit code {p.exitcode}")
+                if p.exitcode in (-9, 137):
+                    logger.error("Process was likely OOM killed by Docker/WSL.")
 
 
 if __name__ == "__main__":

src/shorts_maker/config.py

@@ -1,3 +1,5 @@
+"""Configuration definitions and loaded settings for the Shorts Maker pipeline."""
+
 from pydantic import Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 

src/shorts_maker/core/__init__.py

@@ -0,0 +1 @@
+"""Core processing logic orchestrating scene detection, scoring, and rendering."""