Debug decoder output, fix button color, inline examples

app.py

@@ -330,15 +330,48 @@ def generate(prompt: str, temperature: float, max_tokens: int, num_steps: int, g
                 output_type="latent",
             )
             latent = result.images.to(dtype)
+            print(f"Latent shape: {latent.shape}, dtype: {latent.dtype}")
+            print(f"Latent stats: min={latent.min():.3f}, max={latent.max():.3f}, mean={latent.mean():.3f}")
         
-        # Latent -> Gcode via trained decoder
+        # Latent -> Gcode via trained decoder (with debug)
         with torch.no_grad():
-            gcode = gcode_decoder.generate(
-                latent,
-                gcode_tokenizer,
-                max_length=max_tokens,
-                temperature=temperature,
-            )
+            batch_size = latent.shape[0]
+            input_ids = torch.full((batch_size, 1), gcode_tokenizer.pad_token_id, dtype=torch.long, device=device)
+            
+            generated_tokens = []
+            for step in range(min(max_tokens, 1024) - 1):
+                logits = gcode_decoder(latent, input_ids)
+                next_logits = logits[:, -1, :] / temperature
+                
+                # Top-p sampling
+                sorted_logits, sorted_indices = torch.sort(next_logits, descending=True)
+                cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
+                sorted_indices_to_remove = cumulative_probs > 0.9
+                sorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()
+                sorted_indices_to_remove[:, 0] = False
+                
+                for b in range(batch_size):
+                    next_logits[b, sorted_indices[b, sorted_indices_to_remove[b]]] = float('-inf')
+                
+                probs = torch.softmax(next_logits, dim=-1)
+                next_token = torch.multinomial(probs, num_samples=1)
+                input_ids = torch.cat([input_ids, next_token], dim=1)
+                
+                token_id = next_token.item()
+                generated_tokens.append(token_id)
+                
+                # Debug first few tokens
+                if step < 5:
+                    token_str = gcode_tokenizer.decode([token_id])
+                    print(f"Step {step}: token_id={token_id}, token='{token_str}'")
+                
+                if token_id == gcode_tokenizer.eos_token_id:
+                    print(f"Hit EOS at step {step}")
+                    break
+            
+            print(f"Generated {len(generated_tokens)} tokens")
+            gcode = gcode_tokenizer.decode(input_ids[0], skip_special_tokens=True)
+            print(f"Decoded gcode length: {len(gcode)} chars")
         
         gcode = validate_gcode(gcode)
         line_count = len([l for l in gcode.split("\n") if l.strip()])
@@ -368,14 +401,23 @@ css = """
 }
 
 .gr-button-primary {
-    background: #000 !important;
-    border: none !important;
-    color: #fff !important;
+    background: #e8e8e8 !important;
+    border: 1px solid #ccc !important;
+    color: #333 !important;
     font-weight: 500 !important;
 }
 
 .gr-button-primary:hover {
-    background: #333 !important;
+    background: #d8d8d8 !important;
+}
+
+.gr-examples {
+    margin-top: 8px !important;
+}
+
+.gr-examples .gr-sample-textbox {
+    display: inline-block !important;
+    margin-right: 8px !important;
 }
 
 footer {
@@ -425,6 +467,8 @@ with gr.Blocks(css=css, theme=gr.themes.Base()) as demo:
                     ["geometric shapes"],
                 ],
                 inputs=prompt,
+                label=None,
+                examples_per_page=3,
             )
         
         with gr.Column(scale=2):