unload models before loading llm

modules/ui.py

@@ -393,6 +393,22 @@ def expand_prompt_standalone(prompt: str, model_path: str, system_prompt: str =
             torch.cuda.empty_cache()
             log_step("Previous model unloaded", unload_start)
 
+        # Unload diffusion models from VRAM to make room for LLM
+        log_step("Unloading diffusion models from VRAM to make room for LLM...")
+        unload_diffusion_start = time.time()
+        try:
+            from backend import memory_management
+            memory_management.unload_all_models()
+            memory_management.soft_empty_cache(force=True)
+            gc.collect()
+            torch.cuda.empty_cache()
+            gpu_mem = get_gpu_memory()
+            if gpu_mem:
+                log_step(f"  After diffusion unload: {gpu_mem}")
+            log_step("Diffusion models unloaded", unload_diffusion_start)
+        except Exception as e:
+            log_step(f"  Warning: Could not unload diffusion models: {e}")
+
         log_step(f"Loading LLM model: {model_path}")
         load_start = time.time()
 

prompt.txt

@@ -1,26 +1 @@
-Create a detailed, photorealistic image of a realistic domestic cat actively installing RAM sticks into an open consumer-grade PC case with an extreme RAM configuration.
-Cat Subject Requirements:
-
-Fully realistic tabby cat - standard quadrupedal feline anatomy
-Detailed fur textures with visible individual hairs and stripes
-Cat positioned on desk, both front paws carefully gripping a single RAM stick, placing it into a slot
-Focused, intent expression with ears slightly forward
-Natural cat proportions, sitting upright to reach the motherboard
-
-Computer & Hardware Requirements - CRITICAL:
-
-EXTREME RAM CONFIGURATION: Motherboard with 32 RAM slots arranged in a massive extended row spanning the entire top edge of the motherboard
-30 RGB RAM sticks already installed - creating a spectacular wall of illuminated memory modules in rainbow colors
-Only 2 empty slots remaining, cat installing stick #30
-The RAM sticks should be the visual focal point - an absurdly long row of colorful memory modules
-Large NVIDIA RTX 6090 graphics card with triple-fan design and illuminated "6090" branding
-Open mid-tower case with glass panel removed
-AIO liquid cooler, RGB fans, clean cable management
-
-Environment & Composition:
-
-Clean desk with anti-static mat
-RAM packaging boxes scattered nearby
-Soft lighting with RGB glow illuminating cat's fur
-
-Overall Style: Hyperrealistic photography, humorous tech enthusiast scene, emphasis on the ridiculous quantity of RAM sticks as the key visual element.
+Examine the prompt expansion pipeline when using the z image tab of the gui launched with webui-user.bat. Currently we are getting the error in console.txt. The script needs to unload the models currently in vram before loading the llm for prompt expansion.