still initial commit

app.py

@@ -189,7 +189,9 @@ with gr.Blocks(title="⛏️ Minecraft Spatial Voxel Filter") as demo:
     input_stream.stream(
         fn=process_video_stream,
         inputs=[input_stream, mode_dropdown, is_running, prompt_input, strength_slider],
-        outputs=[output_stream]
+        outputs=[output_stream],
+        trigger_mode="always_last", # Drops intermediate backlog frames when backend is busy
+        concurrency_limit=1         # Ensures only one frame flies over the network at a time
     )
 
 if __name__ == "__main__":

backend/backend.py

@@ -27,10 +27,11 @@ image = (
 app = modal.App("flux-klein-voxel-backend", image=image)
 
 # ==============================================================================
-# 🏎️ 1. THE DEMO PIPELINE
+# 🏎️ 1. THE DEMO PIPELINE (FALLBACK ROUTE)
 # ==============================================================================
 @app.function()
-def demo_stream_frame(img_bytes: bytes) -> bytes:
+def demo_stream_frame(img_bytes: bytes, prompt: str, strength: float) -> bytes:
+    """Fallback route structurally aligned to match the frontend signature."""
     from PIL import Image, ImageDraw
     
     input_image = Image.open(io.BytesIO(img_bytes)).convert("RGB")
@@ -65,29 +66,29 @@ class VoxelModel:
         self.pipe = AutoPipelineForImage2Image.from_pretrained(
             model_id,
             torch_dtype=torch.bfloat16,
-            use_auth_token=os.environ["HF_TOKEN"] 
+            token=os.environ["HF_TOKEN"]  # Modernized from use_auth_token
         )
         self.pipe.to("cuda")
         self.pipe.enable_attention_slicing()
 
         # ---------------------------------------------------------
-        # ✨ COMPILE AND OPTIMIZE THE TRANSFORMER
+        # ✨ COMPILE AND OPTIMIZE THE INFRASTRUCTURE
         # ---------------------------------------------------------
-        print("Initializing torch.compile on the transformer block...")
+        print("Initializing torch.compile optimization loops...")
         
-        # Channels-last memory format often yields slightly faster compiled kernels
-        self.pipe.transformer.to(memory_format=torch.channels_last)
+        # Channels-last optimization applied exclusively to the VAE (CNN-based)
+        self.pipe.vae.to(memory_format=torch.channels_last)
         
-        # Compile the heaviest part of the FLUX architecture
+        # Compile the heaviest part of the FLUX architecture safely without memory layout issues
         self.pipe.transformer = torch.compile(
             self.pipe.transformer,
-            mode="reduce-overhead", # Trades a bit of compile time for faster inference
+            mode="reduce-overhead",  # Trades a bit of compile time for faster inference
             fullgraph=False
         )
 
-        # WARMUP RUN: We force a dummy inference right now. 
-        # If the cache is empty (first run ever), this traces the graph and saves to the Modal Volume.
-        # If the cache exists (subsequent cold-starts), it loads from the Volume in seconds.
+        # WARMUP RUN: Force a dummy inference immediately execution starts.
+        # If the cache is empty (first run), this traces the graph and saves to the Modal Volume.
+        # If the cache exists, it instantly hotloads from the Volume.
         print("Running warmup to build/load inductor cache...")
         dummy_image = Image.new("RGB", (512, 512), (0, 0, 0))
         with torch.inference_mode():