Fix TensorRT/ESRGAN fallback: split requirements for RunPod cu128

- requirements.txt: Remove torch-tensorrt (wrong CUDA build breaks RunPod)
- requirements-runpod.txt: New file with --extra-index-url cu128 for
  torch-tensorrt and torchaudio pinned to correct CUDA toolkit
- goofer_video_concat.py: Loud warnings when TensorRT unavailable so
  silent fallback to Real-ESRGAN is obvious in logs

Prevents the libcudart.so.13 error from recurring on CUDA 12.8 pods.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

goofer_video_concat.py

@@ -387,10 +387,21 @@ def _upscale(images, target_resolution=1080, quality="ULTRA"):
                     n_frames, result.shape[2], result.shape[1])
         return result
 
-    except ImportError:
-        logger.info("torch-tensorrt not installed. Trying standard Real-ESRGAN...")
+    except ImportError as ie:
+        logger.warning(
+            "torch-tensorrt NOT INSTALLED — falling back to standard Real-ESRGAN. "
+            "To fix: pip install -r requirements-runpod.txt  (import error: %s)", ie
+        )
+        print(
+            "[VideoConcat] WARNING: torch-tensorrt not found! Using slow Real-ESRGAN fallback.\n"
+            "[VideoConcat] Fix: pip install -r requirements-runpod.txt"
+        )
     except Exception as e:
-        logger.warning("TensorRT path failed: %s. Trying standard Real-ESRGAN...", e)
+        logger.warning(
+            "TensorRT path FAILED: %s — falling back to standard Real-ESRGAN. "
+            "Check CUDA version matches torch-tensorrt build (nvcc --version).", e
+        )
+        print(f"[VideoConcat] WARNING: TensorRT failed: {e}. Using Real-ESRGAN fallback.")
 
     # --- Strategy 3: Real-ESRGAN via PyTorch CUDA (no TensorRT) ---
     try:

requirements-runpod.txt

@@ -0,0 +1,20 @@
+# ComfyUI-Goofer — RunPod / Linux NVIDIA TensorRT requirements
+# Install AFTER base requirements.txt:
+#   pip install -r requirements.txt
+#   pip install -r requirements-runpod.txt
+#
+# CRITICAL: The --extra-index-url MUST match your CUDA toolkit.
+#   RunPod default (March 2026): CUDA 12.8 → cu128
+#   Check yours:  nvcc --version
+#
+# If you see "libcudart.so.13: cannot open shared object file"
+# you installed the wrong CUDA build. Uninstall and reinstall
+# with the correct index URL below.
+
+--extra-index-url https://download.pytorch.org/whl/cu128
+
+# NVIDIA TensorRT — compiled for tensor cores (Blackwell, Ada, Hopper)
+torch-tensorrt>=2.2.0
+
+# Ensure torchaudio matches PyTorch CUDA build
+torchaudio>=2.0.0

requirements.txt

@@ -11,7 +11,10 @@ opencv-python>=4.8.0      # Optional: denoise/frame prep (graceful fallback)
 # Upscaling — NVIDIA hardware chain
 realesrgan>=0.3.0         # Real-ESRGAN upscaler (auto-downloads ~64 MB model)
 basicsr>=1.4.2            # Required by Real-ESRGAN
-torch-tensorrt>=2.2.0     # NVIDIA TensorRT acceleration (Linux/RunPod — uses tensor cores)
+# NOTE: torch-tensorrt moved to requirements-runpod.txt
+# It MUST match your CUDA toolkit version (cu128/cu130).
+# Installing the wrong build causes libcudart.so errors.
+# On RunPod/Linux:  pip install -r requirements-runpod.txt
 
 # Prompt generation (Qwen2.5 mode) — models auto-download on first use
 transformers>=4.37.0      # Qwen2.5-Instruct + MusicGen