Deploy from GitHub main

.gitattributes

@@ -2,3 +2,4 @@
 *.bin filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
 *.ckpt filter=lfs diff=lfs merge=lfs -text
+exports/compi_export_20250823_171107.zip filter=lfs diff=lfs merge=lfs -text

exports/compi_export_20250823_171107.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bd409356326d49795e641662c102b854fb2177798c827f84b05ec190b8bd197
+size 330651

src/ui/compi_phase3_final_dashboard.py

@@ -18,6 +18,10 @@ Features:
 
 import gc
 import os
+
+# Set PyTorch memory management for better VRAM handling
+os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
 import io
 import csv
 import json
@@ -292,17 +296,87 @@ def load_sd15(txt2img=True):
         )
     return pipe.to(DEVICE)
 
+def force_clear_vram():
+    """Nuclear VRAM cleanup - clears everything possible"""
+    if DEVICE == "cuda":
+        try:
+            # Clear PyTorch cache multiple times
+            for _ in range(3):
+                torch.cuda.empty_cache()
+                torch.cuda.synchronize()
+
+            # Force Python garbage collection
+            import gc
+            gc.collect()
+
+            # Try to reset memory stats (if available)
+            try:
+                torch.cuda.reset_peak_memory_stats()
+                torch.cuda.reset_accumulated_memory_stats()
+            except:
+                pass
+
+            # Final cache clear
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+
+            # Show memory status
+            allocated = torch.cuda.memory_allocated() / (1024**3)
+            reserved = torch.cuda.memory_reserved() / (1024**3)
+            st.info(f"🧹 Memory cleared - Allocated: {allocated:.2f}GB, Reserved: {reserved:.2f}GB")
+
+        except Exception as e:
+            st.warning(f"Memory clearing failed: {e}")
+
 @st.cache_resource(show_spinner=True)
 def load_sdxl():
-    """Load SDXL pipeline"""
+    """Load SDXL pipeline with nuclear VRAM management"""
     if not HAS_SDXL:
         return None
-    pipe = StableDiffusionXLPipeline.from_pretrained(
-        "stabilityai/stable-diffusion-xl-base-1.0",
-        safety_checker=None,
-        torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
-    )
-    return pipe.to(DEVICE)
+
+    # Nuclear cleanup before loading
+    force_clear_vram()
+
+    # Try loading SDXL with retry logic
+    for attempt in range(3):  # Try up to 3 times
+        try:
+            if attempt > 0:
+                st.info(f"🔄 SDXL loading attempt {attempt + 1}/3 - nuclear VRAM cleanup...")
+                force_clear_vram()
+                # Wait a moment for cleanup to take effect
+                import time
+                time.sleep(1)
+
+            pipe = StableDiffusionXLPipeline.from_pretrained(
+                "stabilityai/stable-diffusion-xl-base-1.0",
+                safety_checker=None,
+                torch_dtype=torch.float16 if DEVICE == "cuda" else torch.float32,
+                low_cpu_mem_usage=True,  # Load model parts progressively
+                use_safetensors=True,    # More memory efficient loading
+            )
+            result = pipe.to(DEVICE)
+            if attempt > 0:
+                st.success(f"✅ SDXL loaded successfully on attempt {attempt + 1}")
+            return result
+
+        except torch.OutOfMemoryError as e:
+            if attempt < 2:  # Not the last attempt
+                st.warning(f"⚠️ CUDA OOM on attempt {attempt + 1} - nuclear cleanup and retry...")
+                force_clear_vram()
+                # Longer wait for memory to actually be freed
+                import time
+                time.sleep(2)
+                continue
+            else:
+                st.error(f"🚫 SDXL failed after 3 attempts. Try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True")
+                st.error(f"Error details: {e}")
+                force_clear_vram()
+                return None
+        except Exception as e:
+            st.error(f"Failed to load SDXL: {e}")
+            return None
+
+    return None
 
 @st.cache_resource(show_spinner=True)
 def load_upscaler():
@@ -1046,12 +1120,14 @@ with tab_inputs:
         # Use first chosen style reference as init image
         init_image = ref_images[style_idxs[0]-1].resize((int(width), int(height)))
 
-    # Generation + Clear buttons side-by-side
-    col_gen, col_clear = st.columns([3, 1])
+    # Generation + Clear + Memory buttons
+    col_gen, col_clear, col_mem = st.columns([3, 1, 1])
     with col_gen:
         go = st.button("🚀 Generate Multimodal Art", type="primary", use_container_width=True)
     with col_clear:
         clear = st.button("🧹 Clear", use_container_width=True)
+    with col_mem:
+        clear_mem = st.button("💾 Free VRAM", use_container_width=True, help="Clear model cache and free VRAM")
 
     # Clear logic: reset prompt fields and any generated output state
     if 'generated_images' not in st.session_state:
@@ -1065,6 +1141,17 @@ with tab_inputs:
         st.success("Cleared current prompt and output. Ready for a new prompt.")
         st.rerun()
 
+    # Define clear function before using it
+    def clear_model_cache():
+        """Clear all cached models to free VRAM"""
+        st.cache_resource.clear()
+        force_clear_vram()
+        st.success("🧹 All model caches cleared!")
+
+    if clear_mem:
+        clear_model_cache()
+        st.rerun()
+
     # Cached pipeline getters
     @st.cache_resource(show_spinner=True)
     def get_txt2img():
@@ -1078,6 +1165,8 @@ with tab_inputs:
     def get_sdxl():
         return load_sdxl()
 
+
+
     @st.cache_resource(show_spinner=True)
     def get_upscaler():
         return load_upscaler()
@@ -1142,7 +1231,7 @@ with tab_inputs:
         # Choose pipeline based on model selection
         if model_choice.startswith("SDXL") and HAS_SDXL and gen_mode == "txt2img":
             pipe = get_sdxl()
-            model_id = "SDXL-Base-1.0"
+            model_id = "SDXL-Base-1.0" if pipe else "SD-1.5-fallback"
         else:
             if gen_mode == "txt2img":
                 pipe = get_txt2img()
@@ -1151,6 +1240,10 @@ with tab_inputs:
                 pipe = get_img2img()
                 model_id = "SD-1.5 (img2img)"
 
+        if not pipe:
+            st.error("❌ Failed to load pipeline after all retry attempts. Try restarting the app or use a different model.")
+            st.stop()
+
         # Apply performance optimizations
         xformed = attempt_enable_xformers(pipe) if use_xformers else False
         apply_perf(pipe, attn_slice, vae_slice, vae_tile)