Add --allow-bigger-models option to allow models larger than VRAM using system RAM

videogen

@@ -164,6 +164,16 @@ def clear_memory(clear_cuda=True, aggressive=False):
         gc.collect()
 
 
+def get_available_ram_gb():
+    """Get available system RAM in GB"""
+    try:
+        import psutil
+        mem = psutil.virtual_memory()
+        return mem.available / (1024 ** 3)  # Convert bytes to GB
+    except Exception:
+        return 0.0
+
+
 def get_memory_usage():
     """Get current memory usage statistics
     
@@ -3661,7 +3671,7 @@ def detect_generation_type(prompt, prompt_image=None, prompt_animation=None, arg
     return result
 
 
-def select_best_model(gen_type, models, vram_gb=24, prefer_quality=True, return_all=False, offload_strategy=None):
+def select_best_model(gen_type, models, vram_gb=24, prefer_quality=True, return_all=False, offload_strategy=None, allow_bigger_models=False):
     """Select the best model based on generation type and constraints
     
     Args:
@@ -3671,6 +3681,7 @@ def select_best_model(gen_type, models, vram_gb=24, prefer_quality=True, return_
         prefer_quality: Prefer quality over speed
         return_all: If True, return all candidates sorted by score
         offload_strategy: If an offload strategy is specified, allow larger models
+        allow_bigger_models: If True, allow models larger than VRAM by using system RAM for offloading
     
     Returns: (model_name, model_info, reason) or [(model_name, model_info, reason), ...] if return_all=True
     
@@ -3754,7 +3765,13 @@ def select_best_model(gen_type, models, vram_gb=24, prefer_quality=True, return_
             # Check VRAM compatibility using base model requirements
             # LoRAs add a small overhead (~1-2GB)
             vram_est = parse_vram_estimate(base_model_info.get("vram", "~10 GB")) + 2
-            if offload_strategy:
+            if allow_bigger_models:
+                # If allowing bigger models, check if VRAM + 75% of available RAM is sufficient
+                available_ram = get_available_ram_gb()
+                total_available = vram_gb + (available_ram * 0.75)
+                if vram_est > total_available:
+                    continue
+            elif offload_strategy:
                 # If using offload, allow models up to full VRAM
                 if vram_est > vram_gb:
                     continue
@@ -3912,7 +3929,13 @@ def select_best_model(gen_type, models, vram_gb=24, prefer_quality=True, return_
             # Non-LoRA model handling (original logic)
             # Check VRAM compatibility
             vram_est = parse_vram_estimate(info.get("vram", "~10 GB"))
-            if offload_strategy:
+            if allow_bigger_models:
+                # If allowing bigger models, check if VRAM + 75% of available RAM is sufficient
+                available_ram = get_available_ram_gb()
+                total_available = vram_gb + (available_ram * 0.75)
+                if vram_est > total_available:
+                    continue
+            elif offload_strategy:
                 # If using offload, allow models up to full VRAM
                 if vram_est > vram_gb:
                     continue
@@ -4255,6 +4278,11 @@ def run_auto_mode(args, models):
     print("🤖 AUTO MODE - Analyzing prompts and selecting models")
     print("=" * 60)
     
+    # If --allow-bigger-models is specified, enable sequential offload strategy
+    if args.allow_bigger_models and args.offload_strategy == "model":
+        args.offload_strategy = "sequential"
+        print(f"  📦 --allow-bigger-models enabled, using sequential offload strategy")
+    
     # Track which settings were explicitly provided by user
     # These are settings that have non-default values
     user_provided = {
@@ -4278,6 +4306,7 @@ def run_auto_mode(args, models):
         'prompt_image': getattr(args, 'prompt_image', None) is not None,
         'prompt_animation': getattr(args, 'prompt_animation', None) is not None,
         'image': getattr(args, 'image', None) is not None,
+        'allow_bigger_models': args.allow_bigger_models,
     }
     
     # Store alternative models for retry in auto mode
@@ -4315,7 +4344,7 @@ def run_auto_mode(args, models):
     
     if not user_provided['model']:
         # Get all candidate models for retry support
-        all_candidates = select_best_model(gen_type, models, vram_gb, prefer_quality, return_all=True, offload_strategy=args.offload_strategy)
+        all_candidates = select_best_model(gen_type, models, vram_gb, prefer_quality, return_all=True, offload_strategy=args.offload_strategy, allow_bigger_models=args.allow_bigger_models)
         
         if not all_candidates:
             print("❌ Could not find a suitable model!")
@@ -4368,7 +4397,7 @@ def run_auto_mode(args, models):
             
             # Get all image model candidates
             all_img_candidates = select_best_model(
-                img_gen_type, models, vram_gb, prefer_quality=True, return_all=True, offload_strategy=args.offload_strategy
+                img_gen_type, models, vram_gb, prefer_quality=True, return_all=True, offload_strategy=args.offload_strategy, allow_bigger_models=args.allow_bigger_models
             )
             
             if all_img_candidates:
@@ -9585,6 +9614,8 @@ List TTS voices:
     parser.add_argument("--remove-cached-model", type=str, default=None,
                         metavar="MODEL_ID",
                         help="Remove a specific model from the local HuggingFace cache (e.g., stabilityai/stable-video-diffusion-img2vid-xt-1-1)")
+    parser.add_argument("--allow-bigger-models", action="store_true",
+                        help="Allow models larger than available VRAM by using system RAM for offloading (implies --offload_strategy sequential)")
     parser.add_argument("--clear-cache", action="store_true",
                         help="Clear the entire local HuggingFace cache")
     parser.add_argument("--update-models", action="store_true",