fix: add sampling to model generation + surface errors in logs

app.py

@@ -47,12 +47,18 @@ async def health() -> JSONResponse:
 @app.api(name="summon_npc")
 def summon_npc(image_path: Any) -> dict:
     """Analyze an uploaded image and summon a complete NPC profile."""
+    import traceback
     try:
         path = _extract_file_path(image_path)
+        print(f"[NPCverse] summon_npc: extracted path = {path!r}")
         description = analyze_image(path)
+        print(f"[NPCverse] summon_npc: image description = {description!r}")
         npc = generate_npc(description)
+        print(f"[NPCverse] summon_npc: generated NPC name = {npc.get('name')!r}")
         return _json_safe_dict(npc)
-    except Exception:
+    except Exception as e:
+        print(f"[NPCverse] summon_npc FAILED: {e}")
+        traceback.print_exc()
         return _json_safe_dict(DEFAULT_NPC)
 
 

model_engine.py

@@ -142,6 +142,9 @@ def _generate_from_messages(
         downsample_mode=downsample_mode,
         max_new_tokens=max_new_tokens,
         pad_token_id=processor.tokenizer.eos_token_id,
+        do_sample=True,
+        temperature=0.8,
+        top_p=0.9,
     )
     generated_ids_trimmed = [
         output_ids[len(input_ids):]
@@ -294,16 +297,21 @@ No markdown. No extra text.
 @spaces.GPU
 def generate_npc(description: str) -> dict:
     """Generate a complete RPG NPC JSON object from a visual description."""
+    import traceback
     try:
         msgs = [{'role': 'user', 'content': [{"type": "text", "text": _npc_generation_prompt(description)}]}]
-        result = _generate_from_messages(msgs, max_new_tokens=900)
+        result = _generate_from_messages(msgs, max_new_tokens=1200)
         return _validate_npc_payload(parse_json_safe(str(result)))
-    except Exception:
+    except Exception as e:
+        print(f"[NPCverse] generate_npc first attempt failed: {e}")
+        traceback.print_exc()
         try:
             retry_msgs = [{'role': 'user', 'content': [{"type": "text", "text": _npc_retry_prompt(description)}]}]
-            retry_result = _generate_from_messages(retry_msgs, max_new_tokens=900)
+            retry_result = _generate_from_messages(retry_msgs, max_new_tokens=1200)
             return _validate_npc_payload(parse_json_safe(str(retry_result)))
-        except Exception:
+        except Exception as e2:
+            print(f"[NPCverse] generate_npc retry also failed: {e2}")
+            traceback.print_exc()
             return DEFAULT_NPC