more stuff

config-streaming-runpod.yaml

@@ -43,8 +43,8 @@ output:
   path: "/workspace/output_video.mp4"  # Update with your output path
   format: "greenscreen"  # "greenscreen" or "alpha"
   background_color: [0, 255, 0]  # RGB for green screen
-  video_codec: "h264_nvenc"  # GPU encoding (or "hevc_nvenc" for better compression)
-  quality_preset: "p4"  # NVENC preset (p1-p7, higher = better quality)
+  video_codec: "libx264"  # CPU encoding (use "h264_nvenc" if GPU encoding works)
+  quality_preset: "medium"  # CPU preset (ultrafast/fast/medium/slow/veryslow)
   crf: 18  # Quality (0-51, lower = better, 18 = high quality)
   maintain_sbs: true  # Keep side-by-side format with audio
 

requirements.txt

@@ -12,4 +12,4 @@ ffmpeg-python>=0.2.0
 decord>=0.6.0
 # GPU acceleration (optional but recommended for stereo validation speedup)
 # cupy-cuda11x>=12.0.0  # For CUDA 11.x
-# cupy-cuda12x>=12.0.0  # For CUDA 12.x - uncomment appropriate version
+cupy-cuda12x>=12.0.0  # For CUDA 12.x (most common on modern systems)

runpod_setup.sh

@@ -58,20 +58,11 @@ pip install -r requirements.txt
 print_status "Installing video processing libraries..."
 pip install decord ffmpeg-python
 
-# Install CuPy for GPU acceleration of stereo validation
+# Install CuPy for GPU acceleration (CUDA 12 is standard on modern RunPod)
 print_status "Installing CuPy for GPU acceleration..."
-# Auto-detect CUDA version and install appropriate CuPy
 if command -v nvidia-smi &> /dev/null; then
-    CUDA_VERSION=$(nvidia-smi | grep "CUDA Version" | awk '{print $9}' | cut -d. -f1-2)
-    echo "Detected CUDA version: $CUDA_VERSION"
-    
-    if [[ "$CUDA_VERSION" == "11."* ]]; then
-        pip install cupy-cuda11x>=12.0.0 && print_success "Installed CuPy for CUDA 11.x"
-    elif [[ "$CUDA_VERSION" == "12."* ]]; then
-        pip install cupy-cuda12x>=12.0.0 && print_success "Installed CuPy for CUDA 12.x"
-    else
-        print_error "Unknown CUDA version, skipping CuPy installation"
-    fi
+    print_status "Installing CuPy for CUDA 12.x (standard on RunPod)..."
+    pip install cupy-cuda12x>=12.0.0 && print_success "Installed CuPy for CUDA 12.x"
 else
     print_error "NVIDIA GPU not detected, skipping CuPy installation"
 fi

vr180_streaming/frame_writer.py

@@ -132,13 +132,22 @@ class StreamingFrameWriter:
                 bufsize=10**8  # Large buffer for performance
             )
             
+            # Test if ffmpeg starts successfully (quick check)
+            import time
+            time.sleep(0.1)  # Give ffmpeg time to fail if it's going to
+            
+            if self.ffmpeg_process.poll() is not None:
+                # Process already died - read error
+                stderr = self.ffmpeg_process.stderr.read().decode()
+                raise RuntimeError(f"FFmpeg failed immediately: {stderr}")
+            
             # Set process to ignore SIGINT (Ctrl+C) - we'll handle it
             if hasattr(signal, 'pthread_sigmask'):
                 signal.pthread_sigmask(signal.SIG_BLOCK, [signal.SIGINT])
                 
         except Exception as e:
             # Try CPU fallback if GPU encoding fails
-            if 'nvenc' in self.ffmpeg_cmd:
+            if 'nvenc' in ' '.join(self.ffmpeg_cmd):
                 print(f"⚠️  GPU encoding failed, trying CPU fallback...")
                 self.ffmpeg_cmd = self._build_ffmpeg_command('libx264', 'medium', 18)
                 self._start_ffmpeg()

vr180_streaming/sam2_streaming.py

@@ -34,6 +34,11 @@ class SAM2StreamingProcessor:
         self.config = config
         self.device = torch.device(config.get('hardware', {}).get('device', 'cuda'))
         
+        # Processing parameters (set before _init_predictor)
+        self.memory_offload = config.get('matting', {}).get('memory_offload', True)
+        self.fp16 = config.get('matting', {}).get('fp16', True)
+        self.correction_interval = config.get('matting', {}).get('correction_interval', 300)
+        
         # SAM2 model configuration
         model_cfg = config.get('matting', {}).get('sam2_model_cfg', 'sam2.1_hiera_l')
         checkpoint = config.get('matting', {}).get('sam2_checkpoint', 
@@ -43,11 +48,6 @@ class SAM2StreamingProcessor:
         self.predictor = None
         self._init_predictor(model_cfg, checkpoint)
         
-        # Processing parameters
-        self.memory_offload = config.get('matting', {}).get('memory_offload', True)
-        self.fp16 = config.get('matting', {}).get('fp16', True)
-        self.correction_interval = config.get('matting', {}).get('correction_interval', 300)
-        
         # State management
         self.states = {}  # eye -> inference state
         self.object_ids = []