Update Dockerfile

Dockerfile

@@ -101,7 +101,7 @@ HTML_TEMPLATE = r"""
         const { useState, useEffect, useRef } = React;
         function App() {
             const [desktopImage, setDesktopImage] = useState(null);
-            const [logs, setLogs] = useState(["Neural Bios v9.5", "Booting Kernel..."]);
+            const [logs, setLogs] = useState(["Neural Bios v9.6", "Booting Kernel..."]);
             const socketRef = useRef(null);
             const canvasRef = useRef(null);
 
@@ -180,14 +180,12 @@ PROGRAMS = {
 DRIVERS = {}
 
 def initialize_drivers():
-    # 1. Desktop Background Latent (Blue-ish)
-    bg = torch.zeros((1, 4, 128, 128), dtype=torch.float32) # Using float32 for safety
+    bg = torch.zeros((1, 4, 128, 128), dtype=torch.float32)
     bg[:, 0, :, :] = 0.5   
     bg[:, 1, :, :] = 0.8   
     bg[:, 2, :, :] = 0.2   
     DRIVERS["DESKTOP_BG"] = bg
 
-    # 2. Icon Placeholders
     icon = torch.zeros((1, 4, 8, 8), dtype=torch.float32)
     icon[:, 0, 2:6, 2:6] = 2.0 
     DRIVERS["ICON_GENERIC"] = icon
@@ -210,7 +208,6 @@ class OSKernel:
         pid = self.next_pid
         self.next_pid += 1
         
-        # Initialize empty window latent
         w, h = app.default_size
         latent = torch.zeros((1, 4, h, w), dtype=torch.float32)
         
@@ -261,13 +258,9 @@ class OSKernel:
 class NeuralSystem:
     def __init__(self):
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        # CPU must use float32 to avoid "Half not implemented" errors
         self.dt = torch.float16 if self.device == "cuda" else torch.float32
         print(f"[*] System Device: {self.device} | Type: {self.dt}")
         
-        # A. LOAD DIFFUSION
-        # We do NOT call .to(device) immediately for the whole pipeline if on CPU
-        # to prevent meta-tensor conversion issues with accelerate.
         print("[*] Loading Neural GPU...")
         self.pipe = StableDiffusionPipeline.from_pretrained(
             "runwayml/stable-diffusion-v1-5", 
@@ -278,25 +271,23 @@ class NeuralSystem:
         if self.device == "cuda":
             self.pipe = self.pipe.to("cuda")
         
-        # Load LCM and VAE
         self.pipe.load_lora_weights("latent-consistency/lcm-lora-sdv1-5")
         self.pipe.scheduler = LCMScheduler.from_config(self.pipe.scheduler.config)
         self.pipe.vae = AutoencoderTiny.from_pretrained("madebyollin/taesd", torch_dtype=self.dt).to(self.device)
         
-        # B. LOAD QWEN
         print("[*] Loading Qwen 2.5...")
         self.model_id = "Qwen/Qwen2.5-Coder-0.5B-Instruct"
         self.tokenizer = AutoTokenizer.from_pretrained(self.model_id)
         if self.tokenizer.pad_token_id is None:
             self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
         
-        # FIX: Use device_map instead of .to() to prevent "Cannot copy out of meta tensor"
+        # [FIXED] Removed device_map and low_cpu_mem_usage to prevent meta-tensor crash
+        # This forces a standard, synchronous load into RAM.
         self.llm = AutoModelForCausalLM.from_pretrained(
             self.model_id, 
-            torch_dtype=self.dt,
-            low_cpu_mem_usage=True,
-            device_map=self.device  # Let accelerate handle the placement
-        )
+            torch_dtype=self.dt
+        ).to(self.device)
+        
         print("[*] Systems Online.")
 
     def think(self, prompt_text):
@@ -319,26 +310,21 @@ class NeuralSystem:
         return response
 
     def render_frame(self, kernel: OSKernel):
-        # 1. Base Layer
         canvas = DRIVERS["DESKTOP_BG"].clone().to(self.device)
-
-        # 2. Draw Icons
         icon_dna = DRIVERS["ICON_GENERIC"].to(self.device)
+        
         for icon in kernel.desktop_icons:
             x, y = icon['x'], icon['y']
             canvas[:, :, y:y+8, x:x+8] = icon_dna
 
-        # 3. Draw Windows
         sorted_procs = sorted(kernel.processes.values(), key=lambda p: p.z_order)
         for proc in sorted_procs:
             x, y = proc.position
             w, h = proc.size
             if x+w <= 128 and y+h <= 128:
-                # Ensure latent is on correct device/dtype before blitting
                 proc_latent = proc.latent_state.to(self.device, dtype=self.dt)
                 canvas[:, :, y:y+h, x:x+w] = proc_latent
 
-        # 4. Decode
         with torch.no_grad():
             img = self.pipe.vae.decode(canvas / 0.18215).sample
             img = (img / 2 + 0.5).clamp(0, 1).cpu().permute(0, 2, 3, 1).numpy()
@@ -363,10 +349,8 @@ class NeuralSystem:
                 output_type="latent"
             ).images 
             
-            # Simple Title Bar Injection
             img_latents[:, 1, 0:4, :] = 1.5 
             img_latents[:, 0, 0:4, :] = -0.5 
-            
             proc.latent_state = img_latents
 
 # ============================================================================