Update app.py

app.py

@@ -13,69 +13,40 @@ import torch.distributions as dists # Added import
 # --- START: Copied Helper functions from generation_utils.py ---
 # [Keep the copied functions: top_p_logits, top_k_logits, sample_tokens]
 def top_p_logits(logits, top_p=None):
-    """ Applies top-p filtering to logits. """
-    if top_p is None or top_p >= 1.0:
-        return logits
+    if top_p is None or top_p >= 1.0: return logits
     sorted_logits, sorted_indices = torch.sort(logits, descending=True)
     cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
     sorted_indices_to_remove = cumulative_probs > top_p
-    sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
-    sorted_indices_to_remove[..., 0] = 0
-    mask = torch.zeros_like(logits, dtype=torch.bool, device=logits.device)
-    mask = mask.scatter_(-1, sorted_indices, sorted_indices_to_remove)
-    logits = logits.masked_fill(mask, torch.finfo(logits.dtype).min)
-    return logits
+    sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone(); sorted_indices_to_remove[..., 0] = 0
+    mask = torch.zeros_like(logits, dtype=torch.bool, device=logits.device).scatter_(-1, sorted_indices, sorted_indices_to_remove)
+    return logits.masked_fill(mask, torch.finfo(logits.dtype).min)
 
 def top_k_logits(logits, top_k=None):
-    """ Applies top-k filtering to logits. """
-    if top_k is None or top_k <= 0:
-        return logits
+    if top_k is None or top_k <= 0: return logits
     top_k = min(top_k, logits.size(-1))
     indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
-    logits = logits.masked_fill(indices_to_remove, torch.finfo(logits.dtype).min)
-    return logits
+    return logits.masked_fill(indices_to_remove, torch.finfo(logits.dtype).min)
 
 def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None, margin_confidence=False, neg_entropy=False):
-    """ Samples tokens based on logits and calculates confidence. """
-    if temperature > 0:
-        safe_temp = max(temperature, 1e-6)
-        logits = logits / safe_temp
-    if top_p is not None and 0.0 < top_p < 1.0:
-        logits = top_p_logits(logits, top_p)
-    if top_k is not None and top_k > 0:
-        logits = top_k_logits(logits, top_k)
+    if temperature > 0: safe_temp = max(temperature, 1e-6); logits = logits / safe_temp
+    if top_p is not None and 0.0 < top_p < 1.0: logits = top_p_logits(logits, top_p)
+    if top_k is not None and top_k > 0: logits = top_k_logits(logits, top_k)
     is_all_neg_inf = torch.all(logits == torch.finfo(logits.dtype).min, dim=-1, keepdim=True)
-    if torch.any(is_all_neg_inf):
-        uniform_logits = torch.zeros_like(logits)
-        logits = torch.where(is_all_neg_inf, uniform_logits, logits)
+    if torch.any(is_all_neg_inf): uniform_logits = torch.zeros_like(logits); logits = torch.where(is_all_neg_inf, uniform_logits, logits)
     probs = torch.softmax(logits, dim=-1)
-    probs = torch.clamp(probs, min=0.0)
-    probs = probs / probs.sum(dim=-1, keepdim=True)
-    probs = torch.nan_to_num(probs, nan=0.0)
+    probs = torch.clamp(probs, min=0.0); probs = probs / probs.sum(dim=-1, keepdim=True); probs = torch.nan_to_num(probs, nan=0.0)
     if temperature > 0:
-        try:
-            x0 = dists.Categorical(probs=probs).sample()
-            confidence = torch.gather(probs, -1, x0.unsqueeze(-1)).squeeze(-1)
-        except Exception as e:
-            print(f"Warning: Error during Categorical sampling: {e}. Falling back to argmax.")
-            confidence, x0 = probs.max(dim=-1)
-    else:
-        confidence, x0 = probs.max(dim=-1)
-    if margin_confidence:
-        sorted_probs, _ = torch.sort(probs, dim=-1, descending=True)
-        top1_probs = sorted_probs[..., 0]
-        top2_probs = sorted_probs[..., 1] if sorted_probs.shape[-1] > 1 else top1_probs
-        confidence = top1_probs - top2_probs
-    if neg_entropy:
-        epsilon = 1e-10
-        log_probs = torch.log(probs + epsilon)
-        confidence = torch.sum(probs * log_probs, dim=-1)
+        try: x0 = dists.Categorical(probs=probs).sample(); confidence = torch.gather(probs, -1, x0.unsqueeze(-1)).squeeze(-1)
+        except Exception as e: print(f"Warning: Sampling failed: {e}. Argmax fallback."); confidence, x0 = probs.max(dim=-1)
+    else: confidence, x0 = probs.max(dim=-1)
+    if margin_confidence: sorted_probs, _ = torch.sort(probs, dim=-1, descending=True); top1_probs = sorted_probs[..., 0]; top2_probs = sorted_probs[..., 1] if sorted_probs.shape[-1] > 1 else top1_probs; confidence = top1_probs - top2_probs
+    if neg_entropy: epsilon = 1e-10; log_probs = torch.log(probs + epsilon); confidence = torch.sum(probs * log_probs, dim=-1)
     confidence = torch.nan_to_num(confidence, nan=0.0)
     return confidence, x0
 # --- END: Copied Helper functions ---
 
-
-# [Keep model loading, constants]
+# [Keep model loading, constants as before]
+# Load model configuration to get special token IDs
 config = AutoConfig.from_pretrained("Dream-org/Dream-v0-Instruct-7B", trust_remote_code=True)
 model_path = "Dream-org/Dream-v0-Instruct-7B"
 device = 'cuda' if torch.cuda.is_available() else 'cpu'
@@ -104,10 +75,8 @@ try:
     SPECIAL_TOKEN_IDS.add(IM_END_ID)
 except KeyError: IM_START_ID, IM_END_ID = None, None
 
-
 # --- Helper Functions ---
 def parse_constraints(constraints_text: str) -> Dict[int, List[int]]:
-    """ Parses word constraints. """
     constraints = {}
     if not constraints_text: return constraints
     parts = constraints_text.split(',')
@@ -119,55 +88,26 @@ def parse_constraints(constraints_text: str) -> Dict[int, List[int]]:
             pos = int(pos_str.strip())
             word = word.strip()
             token_ids = []
-            if word:
-                text_to_encode = (" " + word) if (pos > 0 and not word.startswith(" ")) else word
-                token_ids = tokenizer.encode(text_to_encode, add_special_tokens=False)
+            if word: text_to_encode = (" " + word) if (pos > 0 and not word.startswith(" ")) else word; token_ids = tokenizer.encode(text_to_encode, add_special_tokens=False)
             if token_ids and pos >= 0: constraints[pos] = token_ids
             elif not token_ids and word: print(f"Warning: Could not tokenize constraint word '{word}'")
         except ValueError: print(f"Warning: Invalid position '{pos_str}' in constraint part '{part}'")
         except Exception as e: print(f"Warning: Error processing constraint '{part}': {e}")
     return constraints
 
-def format_chat_history(history: List[List[Optional[str]]]) -> List[Dict[str, str]]:
-    """
-    Formats chat history [[user, bot], [user, bot]] into [{'role': 'user', 'content': ...}, ...]
-    for the tokenizer's chat template.
-    """
-    messages = []
-    # Ensure history is not empty and is properly structured
-    if not history:
-        return messages
-    for turn in history:
-        if not isinstance(turn, (list, tuple)) or len(turn) != 2:
-             print(f"Warning: Skipping malformed history turn: {turn}")
-             continue
-        user_msg, assistant_msg = turn
-        if user_msg is not None: # Check if user message exists
-             # Ensure content is a string
-             user_content = str(user_msg) if user_msg is not None else ""
-             messages.append({"role": "user", "content": user_content})
-        # Add assistant message only if it exists and is not None
-        if assistant_msg is not None:
-            assistant_content = str(assistant_msg) if assistant_msg is not None else ""
-            messages.append({"role": "assistant", "content": assistant_content})
-    # print(f"Formatted messages for template: {messages}") # Debug
-    return messages
+# Removed format_chat_history as history will be in the correct format
 
 def apply_constraints_to_state(
     x: torch.Tensor, prompt_length: int, total_length: int,
     parsed_constraints: Dict[int, List[int]], current_step: Optional[int] = None
 ) -> torch.Tensor:
-    """ Applies constraints to the state tensor `x`. """
     modified_x = x.clone()
     for rel_pos, word_token_ids in parsed_constraints.items():
-        abs_start_pos = prompt_length + rel_pos
-        abs_end_pos = abs_start_pos + len(word_token_ids)
+        abs_start_pos = prompt_length + rel_pos; abs_end_pos = abs_start_pos + len(word_token_ids)
         if abs_start_pos < total_length and abs_end_pos <= total_length:
-            try:
-                constraint_tensor = torch.tensor(word_token_ids, dtype=torch.long, device=modified_x.device)
-                modified_x[0, abs_start_pos:abs_end_pos] = constraint_tensor
-            except IndexError: print(f"Warning (Step {current_step}): Constraint OOB: {rel_pos}")
-            except Exception as e: print(f"Warning (Step {current_step}): Constraint failed {rel_pos}: {e}")
+            try: constraint_tensor = torch.tensor(word_token_ids, dtype=torch.long, device=modified_x.device); modified_x[0, abs_start_pos:abs_end_pos] = constraint_tensor
+            except IndexError: print(f"Warning (Step {current_step}): Constraint idx error at {rel_pos}")
+            except Exception as e: print(f"Warning (Step {current_step}): Constraint apply error at {rel_pos}: {e}")
     return modified_x
 
 
@@ -176,7 +116,7 @@ def apply_constraints_to_state(
 @spaces.GPU
 @torch.no_grad()
 def generate_dream_response(
-    history: List[List[Optional[str]]], # IMPORTANT: This is the *full* history from the state
+    history: List[Dict[str, str]], # MODIFIED: Expect List[Dict]
     gen_length: int,
     steps: int,
     constraints_text: str,
@@ -186,37 +126,32 @@ def generate_dream_response(
     alg: str,
     alg_temp: Optional[float],
     visualization_delay: float
-    ): # No return type annotation for generators in older Python? Or use -> Iterator[Tuple[...]]
+): # Removed -> type hint for cleaner yield handling
     """ Generates text step-by-step and yields visualization states live. """
 
-    # Ensure history is valid before proceeding
-    if not history or not history[-1] or history[-1][0] is None:
-        # Yield the current (potentially empty) history back
-        yield history, [("No valid input message found.", "red")], ""
+    if not history or history[-1]["role"] != "user": # Check last message is from user
+        yield history, [("No user message found to respond to.", "red")]
         return
 
     # --- 1. Preparation ---
-    # Use the *entire* history received from the state for context
-    messages_for_template = format_chat_history(history)
+    # History is already formatted for the template
     parsed_constraints = parse_constraints(constraints_text)
 
     try:
+        # apply_chat_template expects List[Dict[str, str]]
         inputs = tokenizer.apply_chat_template(
-            messages_for_template,
+            history, # Use history directly
             return_tensors="pt",
             return_dict=True,
-            add_generation_prompt=True # This adds the assistant prompt turn
+            add_generation_prompt=True # Crucial: Adds the "<|im_start|>assistant\n" prompt
         )
         input_ids = inputs.input_ids.to(device)
         prompt_attention_mask = inputs.attention_mask.to(device) if 'attention_mask' in inputs else torch.ones_like(input_ids)
-        prompt_length = input_ids.shape[1]
-        # print(f"Prompt length for model: {prompt_length}") # Debug
-        # print(f"Input IDs to model (first 50): {input_ids[0, :50].tolist()}") # Debug
-
+        prompt_length = input_ids.shape[1] # Length *after* adding the generation prompt
     except Exception as e:
         print(f"Error applying chat template: {e}")
-        # Yield the current history back with an error message
-        yield history, [("Error preparing input.", "red")], ""
+        # Yield current history and error message for visualization
+        yield history, [("Error preparing input.", "red")]
         return
 
     eps = 1e-3
@@ -227,132 +162,111 @@ def generate_dream_response(
     # --- 2. Initialize Generation State ---
     total_length = prompt_length + gen_length
     initial_generation_part = torch.full((1, gen_length), MASK_ID, dtype=torch.long, device=device)
+    # input_ids already includes the assistant prompt, so just append masks
     x = torch.cat((input_ids, initial_generation_part), dim=1)
 
-    # --- Prepare Attention Mask ---
+    # --- Prepare Attention Mask for SDPA ---
     generation_attention_mask = torch.ones((1, gen_length), dtype=torch.long, device=device)
+    # prompt_attention_mask corresponds to input_ids (which includes assistant prompt)
     full_attention_mask_long = torch.cat((prompt_attention_mask, generation_attention_mask), dim=1)
+
     attention_mask_for_model = full_attention_mask_long.to(model.dtype)
     large_neg_val = torch.finfo(model.dtype).min
     attention_mask_for_model = (1.0 - attention_mask_for_model) * large_neg_val
-    attention_mask_for_model = attention_mask_for_model.unsqueeze(1).unsqueeze(2) # Shape [B, 1, 1, N]
+    attention_mask_for_model = attention_mask_for_model.unsqueeze(1).unsqueeze(2) # [B, 1, 1, N]
 
+    # --- Timesteps ---
     timesteps = torch.linspace(1, eps, steps + 1, device=device)
+
+    # Apply initial constraints (relative to start of generation = prompt_length)
     x = apply_constraints_to_state(x, prompt_length, total_length, parsed_constraints, current_step=-1)
 
-    # --- 3. Visualization & State Setup ---
+    # --- 3. Visualization & History Setup ---
     previous_tokens_vis = None
-    # Use the passed-in history directly. We will modify the *last* item's assistant response.
-    # No need for history_copy if we are careful. Let's try modifying `history` directly.
-    # IMPORTANT: Gradio state needs the component to receive the *entire object* back if it's mutated.
-    # So yielding the modified `history` list itself should work.
-    history_for_yield = history # Reference the original list
+    # MODIFIED: Append placeholder assistant message to the history state *before* looping
+    history.append({"role": "assistant", "content": ""})
 
     # --- 4. Initial Yield (Masked State) ---
     initial_generated_tokens = x[0, prompt_length:].cpu()
     vis_data_initial = []
     for tok_id in initial_generated_tokens.tolist():
-        vis_data_initial.append((MASK_TOKEN, "#444444"))
+        display_token = MASK_TOKEN; color = "#444444"
+        vis_data_initial.append((display_token, color))
+
     previous_tokens_vis = initial_generated_tokens
-    # Yield the *current* history (with None for last bot msg)
-    yield history_for_yield, vis_data_initial, ""
+    # Yield the history (which now includes the empty assistant message) and initial vis
+    yield history, vis_data_initial
     time.sleep(visualization_delay)
 
     # --- 5. Step-by-Step Diffusion Loop ---
     try:
         start_time = time.time()
-        current_response_text = "" # Store intermediate text
-
         for i in range(steps):
             mask_index = (x == MASK_ID)
-            if not mask_index.any():
-                 print(f"No mask tokens left at step {i}. Stopping early.")
-                 break
-
-            outputs = model(
-                input_ids=x,
-                attention_mask=attention_mask_for_model,
-                position_ids=None, use_cache=False, return_dict=True
-            )
+            if not mask_index.any(): break # Stop early
+
+            outputs = model(input_ids=x, attention_mask=attention_mask_for_model, return_dict=True)
             logits = outputs.logits
-            logits = torch.cat([logits[:,:1], logits[:, :-1]], dim=1)
+            logits = torch.cat([logits[:,:1], logits[:, :-1]], dim=1) # Align logits
 
             mask_logits = logits[mask_index]
-            if mask_logits.numel() == 0:
-                 print(f"No masked tokens found for logit selection at step {i}. Stopping.")
-                 break
+            if mask_logits.numel() == 0: break # Stop early
 
             t = timesteps[i]; s = timesteps[i + 1]
             x_new_masked_part = torch.full_like(x[mask_index], MASK_ID, device=device, dtype=torch.long)
 
-            # [Sampling logic remains the same as previous working version]
+            # [Keep sampling/remasking logic ('origin' and confidence-based) exactly the same]
             if alg == 'origin':
                 p_transfer = (1.0 - s / t) if i < steps - 1 else 1.0
                 num_masked = mask_logits.shape[0]
                 transfer_indices_relative = torch.rand(num_masked, device=device) < p_transfer
                 logits_to_sample = mask_logits[transfer_indices_relative]
-                if logits_to_sample.numel() > 0:
-                    _, sampled_tokens = sample_tokens(logits_to_sample, temperature=temperature, top_p=top_p_val, top_k=top_k_val)
-                    x_new_masked_part[transfer_indices_relative] = sampled_tokens
-            else: # Confidence-based
-                use_margin = (alg == 'topk_margin'); use_entropy = (alg == 'entropy')
-                confidence, x0_candidates = sample_tokens(
-                    mask_logits, temperature=temperature, top_p=top_p_val, top_k=top_k_val,
-                    margin_confidence=use_margin, neg_entropy=use_entropy
-                )
+                if logits_to_sample.numel() > 0: _, sampled_tokens = sample_tokens(logits_to_sample, temperature=temperature, top_p=top_p_val, top_k=top_k_val); x_new_masked_part[transfer_indices_relative] = sampled_tokens
+            else:
+                use_margin=(alg == 'topk_margin'); use_entropy=(alg == 'entropy')
+                confidence, x0_candidates = sample_tokens(mask_logits, temperature=temperature, top_p=top_p_val, top_k=top_k_val, margin_confidence=use_margin, neg_entropy=use_entropy)
                 num_mask_token = mask_logits.shape[0]
                 target_num_revealed_float = num_mask_token * (1.0 - s / t)
                 number_transfer_tokens = int(target_num_revealed_float) if i < steps - 1 else num_mask_token
                 if number_transfer_tokens > 0:
                     num_samples = min(number_transfer_tokens, num_mask_token)
                     if num_samples > 0:
-                        transfer_indices_relative = torch.tensor([], dtype=torch.long, device=device) # Init empty
-                        if alg_temp_val is None or alg_temp_val <= 0: # Top-k
+                        transfer_indices_relative = torch.tensor([], dtype=torch.long, device=device)
+                        if alg_temp_val is None or alg_temp_val <= 0: # Top-k confidence
                             sort_metric = confidence if alg != 'entropy' else -confidence
                             k_topk = min(num_samples, sort_metric.numel())
                             if k_topk > 0: _, transfer_indices_relative = torch.topk(sort_metric, k=k_topk)
-                        else: # Sampled
+                        else: # Sample based on confidence temperature
                             if confidence.numel() > 0:
-                                conf_probs = confidence / alg_temp_val
-                                conf_probs = torch.nan_to_num(conf_probs, nan=0.0, posinf=1e9, neginf=-1e9)
-                                conf_probs = torch.clamp(conf_probs - conf_probs.max(), min=-30)
-                                conf_probs = F.softmax(conf_probs, dim=-1)
-                                conf_probs = torch.clamp(conf_probs, min=0.0)
-                                conf_probs = torch.nan_to_num(conf_probs, nan=0.0)
-                                prob_sum = conf_probs.sum()
-                                target_sum_tensor = torch.tensor(1.0, device=device, dtype=prob_sum.dtype)
-                                if not torch.isclose(prob_sum, target_sum_tensor, atol=1e-4) and prob_sum > 0:
-                                    safe_prob_sum = torch.max(prob_sum, torch.tensor(1e-12, device=device, dtype=prob_sum.dtype))
-                                    conf_probs = conf_probs / safe_prob_sum
+                                conf_probs = confidence / alg_temp_val; conf_probs = torch.nan_to_num(conf_probs, nan=0.0, posinf=1e9, neginf=-1e9); conf_probs = torch.clamp(conf_probs - conf_probs.max(), min=-30); conf_probs = F.softmax(conf_probs, dim=-1); conf_probs = torch.clamp(conf_probs, min=0.0); conf_probs = torch.nan_to_num(conf_probs, nan=0.0)
+                                prob_sum = conf_probs.sum(); target_sum_tensor = torch.tensor(1.0, device=device, dtype=prob_sum.dtype)
+                                if not torch.isclose(prob_sum, target_sum_tensor, atol=1e-4) and prob_sum > 0: safe_prob_sum = torch.max(prob_sum, torch.tensor(1e-12, device=device, dtype=prob_sum.dtype)); conf_probs = conf_probs / safe_prob_sum
                                 final_prob_sum_check = conf_probs.sum()
                                 if conf_probs.numel() > 0 and num_samples > 0 and torch.all(conf_probs >= 0) and torch.isclose(final_prob_sum_check, target_sum_tensor, atol=1e-4):
                                     try: transfer_indices_relative = torch.multinomial(conf_probs, num_samples=num_samples, replacement=False)
-                                    except RuntimeError as e: print(f"W{i}: Multinomial failed ('{e}'). Fallback.") # Fallback handled below
-                                if transfer_indices_relative.numel() == 0: # Fallback if sampling failed or wasn't possible
-                                    sort_metric = confidence if alg != 'entropy' else -confidence
-                                    k_fallback = min(num_samples, sort_metric.numel())
-                                    if k_fallback > 0: _, transfer_indices_relative = torch.topk(sort_metric, k=k_fallback)
+                                    except RuntimeError as e: print(f"Warning step {i}: Multinomial failed ('{e}'). Fallback."); sort_metric = confidence if alg != 'entropy' else -confidence; k_fallback = min(num_samples, sort_metric.numel()); if k_fallback > 0: _, transfer_indices_relative = torch.topk(sort_metric, k=k_fallback)
+                                else: sort_metric = confidence if alg != 'entropy' else -confidence; k_fallback = min(num_samples, sort_metric.numel()); if k_fallback > 0: _, transfer_indices_relative = torch.topk(sort_metric, k=k_fallback)
                         # Apply transfer
                         if transfer_indices_relative.numel() > 0:
-                            valid_indices = transfer_indices_relative < x0_candidates.shape[0]
-                            valid_transfer_indices = transfer_indices_relative[valid_indices]
-                            if valid_transfer_indices.numel() > 0 and valid_transfer_indices.max() < x_new_masked_part.shape[0]:
-                                x_new_masked_part[valid_transfer_indices] = x0_candidates[valid_transfer_indices].clone()
+                             valid_indices = transfer_indices_relative < x0_candidates.shape[0]; valid_transfer_indices = transfer_indices_relative[valid_indices]
+                             if valid_transfer_indices.numel() > 0:
+                                  if valid_transfer_indices.max() < x_new_masked_part.shape[0]: x_new_masked_part[valid_transfer_indices] = x0_candidates[valid_transfer_indices].clone()
+                                  else: print(f"Warning step {i}: transfer_indices OOB for x_new_masked_part.")
 
+            x[mask_index] = x_new_masked_part # Update state
 
-            x[mask_index] = x_new_masked_part
+            # --- Apply Constraints ---
+            # Remember prompt_length now includes the assistant prompt turn
             x = apply_constraints_to_state(x, prompt_length, total_length, parsed_constraints, current_step=i)
 
             # --- Yield Visualization ---
             current_generated_tokens = x[0, prompt_length:].cpu()
             vis_data = []
-            # [Visualization formatting logic remains the same]
+            # [Keep visualization formatting logic the same]
             for j in range(gen_length):
                 current_tok_id = current_generated_tokens[j].item()
                 previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None and j < len(previous_tokens_vis) else MASK_ID
-                try:
-                    decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False, clean_up_tokenization_spaces=False)
-                    display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
+                try: decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False, clean_up_tokenization_spaces=False); display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
                 except Exception: display_token = f"[ID:{current_tok_id}]"
                 color = None; token_to_display = display_token
                 if current_tok_id == MASK_ID: color = "#444444"
@@ -361,27 +275,17 @@ def generate_dream_response(
                 should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or (EOS_ID is not None and current_tok_id == EOS_ID)
                 if should_hide and previous_tok_id == current_tok_id: token_to_display = ""; color = None
                 if token_to_display: vis_data.append((token_to_display, color))
-            # ---
 
             previous_tokens_vis = current_generated_tokens
 
-            # --- Update intermediate response text ---
+            # MODIFIED: Update the *content* of the last history item
             intermediate_response_tokens = x[0, prompt_length:]
-            current_response_text = tokenizer.decode(
-                intermediate_response_tokens,
-                skip_special_tokens=True,
-                clean_up_tokenization_spaces=True
-            ).strip()
-
-            # --- Update history for yield ---
-            # Update the placeholder in the *last turn* of the history list
-            if history_for_yield and history_for_yield[-1]:
-                 history_for_yield[-1][1] = current_response_text + "..." # Indicate streaming
-
-            # --- Yield current state ---
-            yield history_for_yield, vis_data, current_response_text
+            intermediate_response_text = tokenizer.decode(intermediate_response_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=True).strip()
+            history[-1]["content"] = intermediate_response_text # Update last dict entry
+
+            # Yield the updated history list and current vis data
+            yield history, vis_data
             time.sleep(visualization_delay)
-            # --- End loop iteration ---
 
         end_time = time.time()
         print(f"Dream generation finished in {end_time - start_time:.2f} seconds.")
@@ -389,49 +293,38 @@ def generate_dream_response(
         # --- 6. Final Processing & Yield ---
         final_sequence = x[0]
         response_tokens = final_sequence[prompt_length:]
-        final_response_text = tokenizer.decode(
-            response_tokens,
-            skip_special_tokens=True,
-            clean_up_tokenization_spaces=True
-        ).strip()
+        final_response_text = tokenizer.decode(response_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=True).strip()
+        # Update the final content in the history object
+        history[-1]["content"] = final_response_text
 
-        # Update the history definitively with the final text
-        if history_for_yield and history_for_yield[-1]:
-            history_for_yield[-1][1] = final_response_text
-
-        # Format final visualization
         final_generated_tokens = x[0, prompt_length:].cpu()
         vis_data_final = []
-        # [Final visualization formatting logic remains the same]
+        # [Keep final visualization formatting logic the same]
         for j in range(gen_length):
-             current_tok_id = final_generated_tokens[j].item()
-             previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None and j < len(previous_tokens_vis) else MASK_ID
-             try:
-                 decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False, clean_up_tokenization_spaces=False)
-                 display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
-             except Exception: display_token = f"[ID:{current_tok_id}]"
-             color = None; token_to_display = display_token
-             if current_tok_id == MASK_ID: color = "#444444"
-             elif previous_tok_id == MASK_ID: color = "#66CC66"
-             else: color = "#6699CC"
-             should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or (EOS_ID is not None and current_tok_id == EOS_ID)
-             if should_hide and previous_tok_id == current_tok_id: token_to_display = ""; color = None
-             if token_to_display: vis_data_final.append((token_to_display, color))
-        # ---
-
-        # Yield the final state
-        yield history_for_yield, vis_data_final, final_response_text
+            current_tok_id = final_generated_tokens[j].item()
+            previous_tok_id = previous_tokens_vis[j].item() if previous_tokens_vis is not None and j < len(previous_tokens_vis) else MASK_ID
+            try: decoded_token = tokenizer.decode([current_tok_id], skip_special_tokens=False, clean_up_tokenization_spaces=False); display_token = MASK_TOKEN if current_tok_id == MASK_ID else decoded_token
+            except Exception: display_token = f"[ID:{current_tok_id}]"
+            color = None; token_to_display = display_token
+            if current_tok_id == MASK_ID: color = "#444444"
+            elif previous_tok_id == MASK_ID: color = "#66CC66"
+            else: color = "#6699CC"
+            should_hide = (PAD_ID is not None and current_tok_id == PAD_ID) or (EOS_ID is not None and current_tok_id == EOS_ID)
+            if should_hide and previous_tok_id == current_tok_id: token_to_display = ""; color = None
+            if token_to_display: vis_data_final.append((token_to_display, color))
+
+        # Yield final history and visualization
+        yield history, vis_data_final
         print("Visualization streaming complete.")
 
     except Exception as e:
         print(f"Error during generation or processing: {e}")
         import traceback
         traceback.print_exc()
-        # Ensure the history state reflects the error somehow? Or just yield error vis.
-        # Yield the history *as it was* when the error occurred.
-        if history_for_yield and history_for_yield[-1]:
-             history_for_yield[-1][1] = f"<Error: {e}>" # Put error in bot response
-        yield history_for_yield, [("Error during generation.", "red")], ""
+        # Set error message in the last history item? Or yield separate error?
+        # Let's just yield the current history and error vis
+        history[-1]["content"] = f"Error: {e}" # Put error in assistant message
+        yield history, [("Error during generation.", "red")]
         return
 
 
@@ -448,17 +341,17 @@ def create_chatbot_demo():
             "[[Blog](https://hkunlp.github.io/blog/2025/dream/)]"
         )
 
-        # Use a single state variable for the history list
-        chat_history_state = gr.State([])
+        # STATE: No explicit state needed if chatbot manages it via input/output
 
         with gr.Row():
             with gr.Column(scale=3):
+                # MODIFIED: Use type="messages"
                 chatbot_ui = gr.Chatbot(
                     label="Conversation",
+                    type="messages", # Use dictionary format
                     height=500,
                     show_copy_button=True,
                     bubble_full_width=False,
-                    # value=[] # Initial value set by state binding later
                 )
                 with gr.Group():
                     with gr.Row():
@@ -474,12 +367,10 @@ def create_chatbot_demo():
                 )
             with gr.Column(scale=2):
                 output_vis = gr.HighlightedText(
-                    label="Denoising Process Visualization", combine_adjacent=True,
-                    show_legend=False, interactive=False
-                )
-                response_text_display = gr.Textbox(
-                    label="Generated Response (Live)", interactive=False, lines=5
+                    label="Denoising Process Visualization",
+                    combine_adjacent=True, show_legend=False, interactive=False
                 )
+                # REMOVED: Separate response text display
 
         with gr.Accordion("Generation Settings", open=False):
              # [Settings sliders remain the same]
@@ -497,88 +388,75 @@ def create_chatbot_demo():
              with gr.Row():
                 visualization_delay = gr.Slider(minimum=0.0, maximum=0.5, value=0.03, step=0.01, label="Visualization Delay (seconds)")
 
-
         clear_btn = gr.Button("Clear Conversation")
 
-        # --- Event Handler Functions ---
+        # --- Event Handlers ---
 
-        def add_user_message(message: str, history: List[List[Optional[str]]]):
-            """
-            Adds the user message to the history state and prepares the UI
-            for the bot's response (clearing previous outputs).
-            """
+        # MODIFIED: add_user_message uses dictionary format
+        def add_user_message(message: str, history: List[Dict[str, str]]):
+            """Adds user message in dictionary format, clears input."""
             if not message.strip():
                 gr.Warning("Please enter a message.")
-                # Return unchanged history and empty outputs
-                return history, history, "", [], ""
-            # Append new turn with user message and None placeholder for bot response
-            history.append([message, None])
-            # Return updated history (for state), history (for immediate UI update),
-            # empty input, empty vis, empty response text.
-            return history, history, "", [], ""
+                return history, "" # Return unchanged history, don't clear input here
+            # Append user message as a dictionary
+            history.append({"role": "user", "content": message})
+            # Return updated history, clear input box
+            return history, ""
 
         def clear_all():
-            """Clears state and all relevant UI components."""
-            return [], [], "", [], "" # state, chatbot, input, vis, response text
+            """Clears chatbot, visualization, and input."""
+            return [], [], "" # Chatbot, Vis, Input
 
         # --- Connect UI elements ---
 
-        # Define inputs/outputs for the generator
+        # Define the inputs for the generation function
+        # MODIFIED: Input is chatbot_ui (provides List[Dict])
         generation_inputs = [
-            chat_history_state, gen_length, steps, constraints_input,
+            chatbot_ui, # Get history directly from chatbot component
+            gen_length, steps, constraints_input,
             temperature, top_p, top_k, remasking_strategy, alg_temp,
             visualization_delay
         ]
-        # Generator yields: history_list, vis_data, response_text
-        generation_outputs = [chatbot_ui, output_vis, response_text_display]
-
-        # Chain the actions: Submit/Click -> add_user_message -> generate_dream_response
-
-        # 1. User submits message (Enter or Button)
-        user_interaction = [user_input, chat_history_state]
-        outputs_after_user_add = [
-            chat_history_state, # Update the state
-            chatbot_ui,         # Update chatbot UI immediately
-            user_input,         # Clear user input box
-            output_vis,         # Clear visualization
-            response_text_display # Clear response text box
-        ]
+        # Define the outputs for the generation function
+        # MODIFIED: Output history (List[Dict]) to chatbot_ui, vis_data to output_vis
+        generation_outputs = [chatbot_ui, output_vis]
 
+        # Handle Textbox Submission (Enter key)
         submit_listener = user_input.submit(
-            fn=add_user_message,
-            inputs=user_interaction,
-            outputs=outputs_after_user_add
-        ).then( # 2. Trigger generation AFTER user message is added and UI cleared
+            fn=add_user_message, # Use modified function
+            inputs=[user_input, chatbot_ui], # Pass chatbot state
+            outputs=[chatbot_ui, user_input], # Update chatbot state, clear input
+            queue=False # User message add should be quick
+        ).then(
             fn=generate_dream_response,
-            inputs=generation_inputs, # Pass the updated state and parameters
-            outputs=generation_outputs, # Stream updates to chatbot, vis, text
+            inputs=generation_inputs,
+            outputs=generation_outputs, # Stream history to chatbot, vis to output_vis
             show_progress="hidden"
         )
 
+        # Handle Send Button Click
         click_listener = send_btn.click(
-            fn=add_user_message,
-            inputs=user_interaction,
-            outputs=outputs_after_user_add
-        ).then( # 2. Trigger generation AFTER user message is added and UI cleared
+            fn=add_user_message, # Use modified function
+            inputs=[user_input, chatbot_ui], # Pass chatbot state
+            outputs=[chatbot_ui, user_input], # Update chatbot state, clear input
+            queue=False # User message add should be quick
+        ).then(
             fn=generate_dream_response,
             inputs=generation_inputs,
-            outputs=generation_outputs,
+            outputs=generation_outputs, # Stream history to chatbot, vis to output_vis
             show_progress="hidden"
         )
 
-        # 3. Clear Button
+        # Clear Button Action
         clear_btn.click(
-            clear_all,
+            clear_all, # Use modified clear function
             inputs=[],
-            outputs=[
-                chat_history_state, chatbot_ui, user_input,
-                output_vis, response_text_display
-            ]
+            outputs=[chatbot_ui, output_vis, user_input], # Clear chatbot, vis, input
+            queue=False
         )
 
     return demo
 
-
 # --- Launch ---
 if __name__ == "__main__":
     demo = create_chatbot_demo()