probably works-ish

app.py

@@ -73,8 +73,8 @@ def _hf_login() -> None:
     else:
         print("[HF Auth] No token found in environment variables")
 
-# Login before loading any models
-_hf_login()
+# Login is handled by Space OAuth/session; avoid explicit CLI login here to prevent OAuth var errors
+# _hf_login()
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
@@ -364,13 +364,18 @@ def zerogpu_generate(full_prompt,
         out_ids = model.generate(
             **inputs,
             do_sample=bool(gen_kwargs.get("do_sample", True)),
-            temperature=float(gen_kwargs.get("temperature", 0.7)),
-            top_p=float(gen_kwargs.get("top_p", 0.9)),
-            top_k=(int(gen_kwargs.get("top_k")) if gen_kwargs.get("top_k") and int(gen_kwargs.get("top_k")) > 0 else None),
+            temperature=float(gen_kwargs.get("temperature", 0.6)),
+            top_p=(float(gen_kwargs.get("top_p")) if gen_kwargs.get("top_p") is not None else None),
+            top_k=(int(gen_kwargs.get("top_k")) if gen_kwargs.get("top_k") else None),
             max_new_tokens=int(gen_kwargs.get("max_new_tokens", MAX_DEF)),
             pad_token_id=model.config.pad_token_id,
             eos_token_id=tokenizer.eos_token_id,
-            bad_words_ids=bad_words_ids,
+            repetition_penalty=float(gen_kwargs.get("repetition_penalty", 1.1)),
+            no_repeat_ngram_size=int(gen_kwargs.get("no_repeat_ngram_size", 6)),
+            logits_processor=logits_processor,
+        )
+            eos_token_id=tokenizer.eos_token_id,
+
             logits_processor=logits_processor,
             repetition_penalty=float(gen_kwargs.get("repetition_penalty", 1.2)),
             no_repeat_ngram_size=int(gen_kwargs.get("no_repeat_ngram_size", 8)),
@@ -420,6 +425,59 @@ def zerogpu_generate(full_prompt,
         if torch.cuda.is_available():
             torch.cuda.empty_cache()
 
+# -----------------------
+# Simple (non-Harmony) GPU path — matches your minimal example
+# -----------------------
+@spaces.GPU(duration=120)
+def zerogpu_generate_simple(prompt_str: str, gen_kwargs: Dict[str, Any], rose_map: Optional[Dict[str, float]], rose_alpha: float, rose_score: Optional[float], seed: Optional[int]) -> Dict[str, str]:
+    """Straight chat_template path. No Harmony tokens. Slices completion from prompt_len.
+    Mirrors the minimal HF example and avoids header loops entirely."""
+    model = None
+    try:
+        if seed is not None:
+            torch.manual_seed(int(seed))
+        model = _load_model_on("auto")
+        device = next(model.parameters()).device
+
+        # Encode prompt string
+        enc = tokenizer(prompt_str, return_tensors="pt")
+        inputs = {k: v.to(device) for k, v in enc.items()}
+        prompt_len = int(inputs["input_ids"].shape[1])
+        if "attention_mask" not in inputs:
+            inputs["attention_mask"] = torch.ones_like(inputs["input_ids"], dtype=torch.long, device=device)
+
+        # Optional Rose bias
+        logits_processor = None
+        if rose_map:
+            bias = build_bias_from_tokens(tokenizer, rose_map).to(device)
+            eff_alpha = float(rose_alpha) * (float(rose_score) if rose_score is not None else 1.0)
+            logits_processor = [RoseGuidedLogits(bias, eff_alpha)]
+
+        out_ids = model.generate(
+            **inputs,
+            do_sample=bool(gen_kwargs.get("do_sample", True)),
+            temperature=float(gen_kwargs.get("temperature", 0.6)),
+            top_p=(float(gen_kwargs.get("top_p")) if gen_kwargs.get("top_p") is not None else None),
+            top_k=(int(gen_kwargs.get("top_k")) if gen_kwargs.get("top_k") else None),
+            max_new_tokens=int(gen_kwargs.get("max_new_tokens", MAX_DEF)),
+            pad_token_id=model.config.pad_token_id,
+            logits_processor=logits_processor,
+        )
+        # Slice generated continuation only
+        new_ids = out_ids[0, prompt_len:]
+        text = tokenizer.decode(new_ids, skip_special_tokens=True)
+        return {"final": text}
+    except Exception as e:
+        return {"final": f"[Error] {type(e).__name__}: {e}"}
+    finally:
+        try:
+            del model
+        except Exception:
+            pass
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
 # -----------------------
 # GPU Debug: Harmony Inspector
 # -----------------------
@@ -460,7 +518,6 @@ def zerogpu_generate_debug(full_prompt, gen_kwargs: Dict[str, Any]) -> Dict[str,
             max_new_tokens=int(gen_kwargs.get("max_new_tokens", MAX_DEF)),
             pad_token_id=model.config.pad_token_id,
             eos_token_id=tokenizer.eos_token_id,
-            bad_words_ids=bad_words_ids,
             stopping_criteria=sc,
             repetition_penalty=float(gen_kwargs.get("repetition_penalty", 1.15)),
             no_repeat_ngram_size=int(gen_kwargs.get("no_repeat_ngram_size", 6)),
@@ -517,29 +574,45 @@ def generate_response(message: str, history: List[List[str]], system_prompt: str
                     rose_enable: bool, rose_alpha: float, rose_score: Optional[float], 
                     rose_tokens: str, rose_json: str,
                     show_thinking: bool = False,
+                    simple_mode: bool = True,  # NEW: default to simple chat_template path
                     reasoning_effort: str = "high") -> str:
     """
     Generate response with proper CoT handling using Harmony format.
     """
     try:
-        # Build message list
+        # Build messages robustly for Gradio type='messages' or legacy tuple format
         messages = [{"role": "system", "content": system_prompt or SYSTEM_DEF}]
-        
-        # Add history
+
+        # Add prior turns
         if history:
-            for turn in history:
-                if isinstance(turn, (list, tuple)) and len(turn) >= 2:
-                    user_msg, assistant_msg = turn[0], turn[1]
-                    if user_msg:
-                        messages.append({"role": "user", "content": str(user_msg)})
-                    if assistant_msg:
-                        messages.append({"role": "assistant", "content": str(assistant_msg)})
-        
-        # Add current message
-        messages.append({"role": "user", "content": str(message)})
-        
-        # Create Harmony-formatted prompt
-        if HARMONY_AVAILABLE:
+            if isinstance(history, list) and history and isinstance(history[0], dict):
+                # history is already a flat list of {'role','content'} dicts
+                for m in history:
+                    role = m.get("role")
+                    content = m.get("content", "")
+                    if role in ("user", "assistant"):
+                        messages.append({"role": role, "content": str(content)})
+            else:
+                for turn in history:
+                    if isinstance(turn, (list, tuple)) and len(turn) >= 2:
+                        u, a = turn[0], turn[1]
+                        if u:
+                            messages.append({"role": "user", "content": str(u)})
+                        if a:
+                            messages.append({"role": "assistant", "content": str(a)})
+
+        # Current user message
+        if isinstance(message, dict):
+            user_text = message.get("content", "")
+        else:
+            user_text = str(message)
+        messages.append({"role": "user", "content": user_text})
+
+        # FAST PATH: simple chat_template prompt (recommended)
+        if simple_mode:
+            prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+        # Harmony path (optional)
+        elif HARMONY_AVAILABLE:
             prompt = create_harmony_prompt(messages, reasoning_effort)  # returns token IDs
         else:
             # Fallback to tokenizer template (string)
@@ -573,7 +646,23 @@ def generate_response(message: str, history: List[List[str]], system_prompt: str
                 rose_map = None
 
         # Generate with model
-        channels = zerogpu_generate(
+        if simple_mode:
+            channels = zerogpu_generate_simple(
+                prompt,
+                {
+                    "do_sample": bool(do_sample),
+                    "temperature": float(temperature),
+                    "top_p": float(top_p) if top_p is not None else None,
+                    "top_k": int(top_k) if top_k > 0 else None,
+                    "max_new_tokens": int(max_new_tokens),
+                },
+                rose_map,
+                float(rose_alpha),
+                float(rose_score) if rose_score is not None else None,
+                int(seed) if seed is not None else None,
+            )
+        else:
+            channels = zerogpu_generate(
             prompt,
             {
                 "do_sample": bool(do_sample),
@@ -641,7 +730,13 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
             lines=2
         )
     
-    with gr.Accordion("Generation Settings", open=False):
+    with gr.Accordion("Generation Settings ", open=False):
+        # NEW: toggle to bypass Harmony and use plain chat_template like your minimal script
+        simple_mode = gr.Checkbox(
+            value=True,
+            label="Use simple chat_template (no Harmony)",
+            info="Matches the minimal HF example; safest path for now"
+        )
         with gr.Row():
             temperature = gr.Slider(0.0, 2.0, value=0.7, step=0.05, label="Temperature")
             top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.01, label="Top-p")
@@ -692,9 +787,9 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
         fn=generate_response,
         type="messages",
         additional_inputs=[
-            system_prompt, temperature, top_p, top_k, max_new, 
-            do_sample, seed, rose_enable, rose_alpha, rose_score, 
-            rose_tokens, rose_json, show_thinking, reasoning_effort
+            system_prompt, temperature, top_p, top_k, max_new,
+            do_sample, seed, rose_enable, rose_alpha, rose_score,
+            rose_tokens, rose_json, show_thinking, simple_mode, reasoning_effort
         ],
         title="Chat with Mirel",
         description="A chain-of-thought model using Harmony format",