fix: move expandable_segments to app.py before torch imports

CRITICAL FIX: Previous attempt to set PYTORCH_CUDA_ALLOC_CONF failed
because it was set in chronos_inference.py AFTER torch was already
imported by Gradio or other dependencies.

Solution: Move environment variable to top of app.py BEFORE all imports
(except os/sys). This ensures PyTorch sees the config on first import.

Validation: OOM errors showed fragmentation (12.61 GB reserved but
unallocated) and PyTorch still suggested setting expandable_segments,
proving the previous fix wasn't being applied.

This fixes memory fragmentation that prevented allocating 10.75 GB
contiguous blocks even when 12.61 GB was reserved.

Co-Authored-By: Claude <[email protected]>

app.py

@@ -2,15 +2,21 @@
 """
 FBMC Chronos-2 Forecasting API
 HuggingFace Space Gradio Interface
-Version: 1.0.1 (fixed target column bug)
+Version: 1.0.2 (fixed memory fragmentation - expandable_segments)
 """
 
+# CRITICAL: Set PyTorch memory allocator config BEFORE any imports
+# This prevents memory fragmentation issues that cause OOM even with sufficient free memory
+# Must be set before torch is imported the first time (including via gradio or other dependencies)
+import os
+os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'
+
 import sys
 print(f"[STARTUP] Python version: {sys.version}", flush=True)
 print(f"[STARTUP] Python path: {sys.path[:3]}", flush=True)
+print(f"[STARTUP] PyTorch memory config: {os.environ.get('PYTORCH_CUDA_ALLOC_CONF')}", flush=True)
 
 import gradio as gr
-import os
 from datetime import datetime
 
 print("[STARTUP] Basic imports successful", flush=True)