Set DeepSpeed environment variables to prevent CUDA compilation issues on Windows

app.py

@@ -1,8 +1,15 @@
+import os
+# CRITICAL: Set DeepSpeed environment variables BEFORE any imports
+# This prevents DeepSpeed from trying to compile CUDA ops on Windows
+os.environ['DS_BUILD_OPS'] = '0'
+os.environ['DS_BUILD_AIO'] = '0'
+os.environ['DS_BUILD_FUSED_ADAM'] = '0'
+os.environ['DS_BUILD_SPARSE_ATTN'] = '0'
+
 import gradio as gr
 import torch
 import llava
 from peft import PeftModel
-import os
 from huggingface_hub import snapshot_download