fix: jangan paksa thinking & tampilkan raw output saat thinking off

app.py

@@ -284,9 +284,9 @@ def run_inference_raw(user_prompt, max_tokens, temperature, top_k, top_p, rep_pe
     model = model_manager.model
     thinking = model_manager.thinking
 
-    if thinking and thinking[0] not in user_prompt:
-        user_prompt += thinking[0] + "\n"
-
+    # NOTE: Jangan paksa thinking dengan append start-token. Kalau
+    # enable_thinking=False sudah dipakai di format_prompt (Chat mode),
+    # append ini justru membatalkannya dan memaksa model berpikir.
     inputs = tokenizer([user_prompt], return_tensors="pt")
 
     if use_cuda:
@@ -343,19 +343,24 @@ def run_inference_raw(user_prompt, max_tokens, temperature, top_k, top_p, rep_pe
 
             if thinking:
                 start_tok, end_tok, think_label, answer_label = thinking
-                clean = generated_text.replace("<s>", "").replace("</s>", "")
-                clean = clean.replace(start_tok, "").replace(end_tok, "")
-                clean = clean.replace("<|begin_of_solution|>", "").replace("<|end_of_solution|>", "")
-                if end_tok in generated_text:
-                    parts = generated_text.split(end_tok, 1)
-                    think_raw = parts[0].replace(start_tok, "").strip()
-                    answer_raw = parts[1].replace("<|begin_of_solution|>", "").replace("<|end_of_solution|>", "").strip()
-                    think_block = "\n".join("> " + line for line in think_raw.splitlines()) if think_raw else "> _thinking..._"
-                    display_text = f"{think_block}\n\n**{answer_raw}**"
-                else:
-                    think_raw = clean.strip()
-                    think_block = "\n".join("> " + line for line in think_raw.splitlines()) if think_raw else "> _thinking..._"
-                    display_text = f"{think_block}\n\n*...*"
+                # Hanya bungkus sebagai thinking kalau output BENAR-BENAR
+                # mengandung token thinking. Kalau thinking dimatikan
+                # (enable_thinking=False), output adalah teks biasa (mis. JSON)
+                # dan harus ditampilkan apa adanya tanpa prefix "> " / "*...*".
+                if start_tok in generated_text or end_tok in generated_text:
+                    clean = generated_text.replace("<s>", "").replace("</s>", "")
+                    clean = clean.replace(start_tok, "").replace(end_tok, "")
+                    clean = clean.replace("<|begin_of_solution|>", "").replace("<|end_of_solution|>", "")
+                    if end_tok in generated_text:
+                        parts = generated_text.split(end_tok, 1)
+                        think_raw = parts[0].replace(start_tok, "").strip()
+                        answer_raw = parts[1].replace("<|begin_of_solution|>", "").replace("<|end_of_solution|>", "").strip()
+                        think_block = "\n".join("> " + line for line in think_raw.splitlines()) if think_raw else "> _thinking..._"
+                        display_text = f"{think_block}\n\n**{answer_raw}**"
+                    else:
+                        think_raw = clean.strip()
+                        think_block = "\n".join("> " + line for line in think_raw.splitlines()) if think_raw else "> _thinking..._"
+                        display_text = f"{think_block}\n\n*...*"
 
             device_label = "CUDA" if use_cuda else "CPU"
             yield base_display + display_text, f"Speed: {tps:.2f} tokens/sec ({device_label})"