Update generation_utils.py

generation_utils.py

@@ -1,5 +1,7 @@
 # coding=utf-8
-# Copyright ...
+# Copyright 2024 The Dream team, HKUNLP Group and the HuggingFace
+# Licensed under the Apache License, Version 2.0
+
 import warnings
 import copy
 from dataclasses import dataclass
@@ -16,33 +18,38 @@ logger = logging.get_logger(__name__)
 
 
 def top_p_logits(logits, top_p=None):
+    if top_p is None or top_p >= 1:
+        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
+    # keep first token above threshold
     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
+    return logits.masked_fill(mask, torch.finfo(logits.dtype).min)
 
 
 def top_k_logits(logits, top_k=None):
+    if top_k is None:
+        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
+    thresh = torch.topk(logits, top_k)[0][..., -1, None]
+    indices_to_remove = logits < thresh
+    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,
+    temperature: float = 0.0,
+    top_p: Optional[float] = None,
+    top_k: Optional[int] = None,
+    margin_confidence: bool = False,
+    neg_entropy: bool = False,
 ):
-    if temperature > 0:
+    # 保持 dtype 与 logits 一致（包含 bf16/fp16）
+    if temperature and temperature > 0:
         logits = logits / temperature
     if top_p is not None and top_p < 1:
         logits = top_p_logits(logits, top_p)
@@ -51,25 +58,27 @@ def sample_tokens(
 
     probs = torch.softmax(logits, dim=-1)
 
-    if temperature > 0:
+    if temperature and temperature > 0:
+        # 采样
         try:
             x0 = dists.Categorical(probs=probs).sample()
             confidence = torch.gather(probs, -1, x0.unsqueeze(-1)).squeeze(-1)
         except Exception:
             confidence, x0 = probs.max(dim=-1)
     else:
+        # 贪心
         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]
+        top1_probs = sorted_probs[..., 0]
+        top2_probs = sorted_probs[..., 1]
         confidence = top1_probs - top2_probs
 
     if neg_entropy:
-        # 保持你原来的“熵”定义（注意它是负值；不改符号，避免影响 baseline）
-        epsilon = 1e-10
-        log_probs = torch.log(probs + epsilon)
+        eps = probs.new_tensor(1e-10)
+        log_probs = torch.log(probs + eps)
+        # 负熵（和为负数），数值上越大（绝对值越小）表示不确定；此处直接用于排序
         confidence = torch.sum(probs * log_probs, dim=-1)
 
     return confidence, x0
@@ -88,13 +97,14 @@ class DreamGenerationConfig(GenerationConfig):
         self.top_k: Optional[int] = kwargs.pop("top_k", None)
         self.max_length = kwargs.pop("max_length", 20)
         self.max_new_tokens = kwargs.pop("max_new_tokens", None)
-        # diffusion specific
+
+        # diffusion specific params
         self.eps: float = kwargs.pop("eps", 1e-3)
         self.steps: int = kwargs.pop("steps", 512)
         self.alg: str = kwargs.pop("alg", "origin")
         self.alg_temp: Optional[float] = kwargs.pop("alg_temp", None)
 
-        # RCR：默认关闭；开启后只做“选后回遮”，不动 baseline 行为
+        # RCR 参数（默认不生效）
         self.rcr: bool = kwargs.pop("rcr", False)
         self.conf_alg: str = kwargs.pop("conf_alg", "maskgit_plus")
 
@@ -111,6 +121,7 @@ class DreamGenerationConfig(GenerationConfig):
 
         self.generation_kwargs = kwargs.pop("generation_kwargs", {})
 
+        # hub meta
         self._from_model_config = kwargs.pop("_from_model_config", False)
         self._commit_hash = kwargs.pop("_commit_hash", None)
         self.transformers_version = kwargs.pop("transformers_version", __version__)
@@ -126,6 +137,7 @@ class DreamGenerationConfig(GenerationConfig):
         self.validate(is_init=True)
 
     def validate(self, is_init=False):
+        # 保留空实现，兼容 upstream
         pass
 
 
@@ -144,50 +156,70 @@ class DreamGenerationMixin:
             attention_mask = attention_mask.repeat_interleave(expand_size, dim=0)
         return input_ids, attention_mask
 
-    # 仅 rcr=True 使用；不改变 baseline 的选入逻辑
-    def _rcr_remask_after_selection(
+    def _apply_rcr_logic(
         self,
-        x,                        # [B, L] 当前序列
+        x: torch.LongTensor,
+        x0_sel: torch.LongTensor,
+        conf_sel: torch.Tensor,
+        mask_index: torch.Tensor,
+        overtime_confidence: torch.Tensor,
         mask_token_id: int,
         step: int,
-        steps: int,
+        total_steps: int,
         s: torch.Tensor,
         t: torch.Tensor,
-        is_fixed: torch.Tensor,   # [B, L] bool，已“确定”的位置
-        fixed_conf: torch.Tensor  # [B, L] float，已确定位置的置信度（其余为 -inf）
     ):
         """
-        在已经“按 baseline 完成选入”之后，按累计目标回遮最低置信度的超额 token。
-        —— 极小侵入：不改变 baseline 的挑选，只在其后做回遮。
+        Running Confidence Remasking (RCR)
+        - 按 Dream 原调度计算每步应转移的 token 数；
+        - 先把本步最高置信度的若干个位置从 [MASK] 转为预测；
+        - 再根据“截至本步的目标累计数量”，把最低置信度的多余部分回遮回 [MASK]。
+        仅在 rcr=True 时调用。
         """
-        B, L = x.shape
-        # 计算“批均值语义”的 num_mask_token（与 baseline 保持一致）
-        # 注意这里基于当前 x 的 [MASK] 数量计算
-        mask_index = (x == mask_token_id)
-        num_mask_token = (mask_index.sum() / mask_index.shape[0]).item()
+        device = x.device
+        dtype = overtime_confidence.dtype  # == logits.dtype
+        B = x.shape[0]
+
+        # 当前 batch 平均剩余 mask 数
+        num_mask_token = mask_index.sum() / mask_index.shape[0]
+        # 本步的转移数量（与 Dream 调度一致）
+        number_transfer_tokens = int(num_mask_token * (1 - s / t)) if step < total_steps - 1 else int(num_mask_token)
 
-        # Dream 原调度：到本步为止应累计确定的目标总量
-        target_cum = int(num_mask_token * (1 - (s / t).item())) if step < steps - 1 else int(num_mask_token)
+        # 构造“全长”置信度与候选 token（非 mask 位置分别设为 -inf / mask_token_id）
+        full_conf = torch.full(x.shape, float("-inf"), device=device, dtype=dtype)
+        x_temp = torch.full_like(x, fill_value=mask_token_id, dtype=torch.long, device=device)
+        full_conf[mask_index] = conf_sel
+        x_temp[mask_index] = x0_sel
 
         for j in range(B):
-            # 当前累计（已确定）的数量
-            fixed_j = is_fixed[j]
-            current_gen = int(fixed_j.sum().item())
-            # 如果超额，回遮最低置信度的那部分
-            to_remask = max(0, current_gen - target_cum)
-            if to_remask > 0:
-                cand_idx = torch.where(fixed_j)[0]
-                if cand_idx.numel() == 0:
-                    continue
-                conf_vals = fixed_conf[j, cand_idx]
-                # 取最小的 to_remask 个
-                k = min(to_remask, int(cand_idx.numel()))
-                _, local_low = torch.topk(conf_vals, k=k, largest=False)
-                low_global = cand_idx[local_low]
-                # 打回 [MASK]，并清空标记
-                x[j, low_global] = mask_token_id
-                is_fixed[j, low_global] = False
-                fixed_conf[j, low_global] = float("-inf")
+            masked_j = int(mask_index[j].sum().item())
+            if masked_j == 0:
+                continue
+            k_j = min(number_transfer_tokens, masked_j)
+
+            if k_j > 0:
+                # 选出本步 top-k_j 的位置
+                _, select_idx = torch.topk(full_conf[j], k=k_j, largest=True)
+                x[j, select_idx] = x_temp[j, select_idx]
+                # 记录这些位置的置信度，用于累计与回遮判断
+                overtime_confidence[j, select_idx] = full_conf[j, select_idx]
+
+            # 目标累计（与原 Dream 线性进度对齐）
+            if step < total_steps - 1:
+                target_cum = int(num_mask_token * (1 - s / t))  # 累计目标到当前步
+                gen_mask = overtime_confidence[j] > overtime_confidence.new_tensor(0)
+                current_gen = int(gen_mask.sum().item())
+                overflow = max(0, current_gen - target_cum)
+                if overflow > 0:
+                    gen_indices = torch.where(gen_mask)[0]
+                    if gen_indices.numel() > 0:
+                        gen_conf = overtime_confidence[j, gen_indices]
+                        overflow = min(overflow, int(gen_indices.numel()))
+                        # 选“最低置信度”的 overflow 个位置回遮
+                        _, low_local = torch.topk(gen_conf, k=overflow, largest=False)
+                        low_global = gen_indices[low_local]
+                        x[j, low_global] = mask_token_id
+                        overtime_confidence[j, low_global] = overtime_confidence.new_zeros(low_global.shape)
 
     def _validate_generated_length(self, generation_config, input_ids_length, has_default_max_length):
         if is_torchdynamo_compiling():
@@ -200,11 +232,9 @@ class DreamGenerationMixin:
                 UserWarning,
             )
         if input_ids_length >= generation_config.max_length:
-            input_ids_string = "input_ids"
             raise ValueError(
-                f"Input length of {input_ids_string} is {input_ids_length}, but `max_length` is set to"
-                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
-                " increasing `max_length` or, better yet, setting `max_new_tokens`."
+                f"Input length is {input_ids_length}, but `max_length` is {generation_config.max_length}. "
+                "Consider increasing `max_length` or setting `max_new_tokens`."
             )
 
     def _prepare_generated_length(self, generation_config, has_default_max_length, input_ids_length):
@@ -218,12 +248,14 @@ class DreamGenerationMixin:
         elif has_default_max_length:
             if generation_config.max_length == DreamGenerationConfig().max_length:
                 generation_config.max_length = generation_config.max_length + input_ids_length
-                max_position_embeddings = getattr(self.config, "max_position_embeddings", None)
-                if max_position_embeddings is not None:
-                    generation_config.max_length = min(generation_config.max_length, max_position_embeddings)
+                mpe = getattr(self.config, "max_position_embeddings", None)
+                if mpe is not None:
+                    generation_config.max_length = min(generation_config.max_length, mpe)
         return generation_config
 
-    def _prepare_generation_config(self, generation_config: Optional[DreamGenerationConfig], **kwargs: Dict) -> DreamGenerationConfig:
+    def _prepare_generation_config(
+        self, generation_config: Optional[DreamGenerationConfig], **kwargs: Dict
+    ) -> DreamGenerationConfig:
         using_model_generation_config = False
         if generation_config is None:
             generation_config = DreamGenerationConfig.from_model_config(self.config)
@@ -231,7 +263,7 @@ class DreamGenerationMixin:
 
         if not is_torchdynamo_compiling():
             generation_config = copy.deepcopy(generation_config)
-            _kwargs = generation_config.update(**kwargs)
+            _ = generation_config.update(**kwargs)
             if not using_model_generation_config:
                 if generation_config.bos_token_id is None:
                     generation_config.bos_token_id = self.generation_config.bos_token_id
@@ -243,7 +275,9 @@ class DreamGenerationMixin:
                     generation_config.mask_token_id = self.generation_config.mask_token_id
         return generation_config
 
-    def _prepare_special_tokens(self, generation_config: DreamGenerationConfig, device: Optional[Union[torch.device, str]] = None):
+    def _prepare_special_tokens(
+        self, generation_config: DreamGenerationConfig, device: Optional[Union[torch.device, str]] = None
+    ):
         def _tensor_or_none(token, device=None):
             if token is None:
                 return token
@@ -293,21 +327,24 @@ class DreamGenerationMixin:
             has_default_max_length=has_default_max_length,
             input_ids_length=input_ids_length,
         )
+
         self._validate_generated_length(generation_config, input_ids_length, has_default_max_length)
 
         if not is_torchdynamo_compiling() and self.device.type != input_ids.device.type:
             warnings.warn(
-                "You are calling .generate() with the `input_ids` being on a device type different"
-                f" than your model's device. `input_ids` is on {input_ids.device.type}, whereas the model"
-                f" is on {self.device.type}."
+                "You are calling .generate() with `input_ids` on a device type different than your model's device. "
+                f"`input_ids` is on {input_ids.device.type}, model is on {self.device.type}.",
+                UserWarning,
             )
+
         if (
             hasattr(generation_config, "pad_token_id")
             and torch.any(input_ids == generation_config.pad_token_id)
             and attention_mask is None
         ):
             warnings.warn(
-                "Padding was detected but no attention mask is passed here. For correct results, please set `attention_mask`."
+                "Padding was detected but no attention mask is passed. For correct results, set `attention_mask` when batch-padding inputs.",
+                UserWarning,
             )
 
         input_ids, attention_mask = self._expand_inputs_for_generation(
@@ -333,6 +370,7 @@ class DreamGenerationMixin:
         generation_tokens_hook_func,
         generation_logits_hook_func,
     ) -> Union[DreamModelOutput, torch.LongTensor]:
+        # 原变量
         output_history = generation_config.output_history
         return_dict_in_generate = generation_config.return_dict_in_generate
         max_length = generation_config.max_length
@@ -345,12 +383,13 @@ class DreamGenerationMixin:
         top_p = generation_config.top_p
         top_k = generation_config.top_k
 
+        # RCR 控制
         rcr = generation_config.rcr
         conf_alg = generation_config.conf_alg
 
         histories = [] if (return_dict_in_generate and output_history) else None
 
-        # pad to max_length
+        # pad 到 max_length
         x = F.pad(input_ids, (0, max_length - input_ids.shape[1]), value=mask_token_id)
 
         if attention_mask is not None and torch.any(attention_mask == 0.0):
@@ -367,107 +406,98 @@ class DreamGenerationMixin:
 
         timesteps = torch.linspace(1, eps, steps + 1, device=x.device)
 
-        # 仅 rcr=True：引入轻量跟踪，不影响 baseline
-        is_fixed = torch.zeros_like(x, dtype=torch.bool) if rcr else None
-        fixed_conf = torch.full(x.shape, float("-inf"), device=x.device, dtype=torch.float32) if rcr else None
-        # 存放已确定位置的置信度
+        # 置信度累计缓冲，延迟到拿到 logits.dtype 后再初始化，避免 dtype 错误
+        overtime_confidence = None  # dtype = logits.dtype（初始化时设置）
 
+        # 允许用户控制中间 tokens
         x = generation_tokens_hook_func(None, x, None)
+
         for i in range(steps):
             mask_index = (x == mask_token_id)
+
             logits = self(x, attention_mask, tok_idx).logits
             logits = torch.cat([logits[:, :1], logits[:, :-1]], dim=1)
 
+            # 允许用户控制中间 logits
             logits = generation_logits_hook_func(i, x, logits)
 
             mask_logits = logits[mask_index]
             t = timesteps[i]
             s = timesteps[i + 1]
 
+            # 首次根据 logits.dtype 初始化 overtime_confidence（避免 Float/BFloat16 冲突）
+            if rcr and overtime_confidence is None:
+                overtime_confidence = torch.zeros_like(x, dtype=logits.dtype, device=x.device)
+
             if alg == "origin":
-                # 完全保持原始
+                # 原始 Dream 逻辑（不动）
                 p_transfer = 1 - s / t if i < steps - 1 else 1
-                x0 = torch.zeros_like(x[mask_index], device=self.device, dtype=torch.long) + mask_token_id
+                x0 = torch.full_like(x[mask_index], fill_value=mask_token_id, dtype=torch.long, device=self.device)
                 transfer_index_t_s = torch.rand(*x0.shape, device=self.device) < p_transfer
                 _, x0[transfer_index_t_s] = sample_tokens(
-                    mask_logits[transfer_index_t_s],
-                    temperature=temperature,
-                    top_p=top_p,
-                    top_k=top_k,
+                    mask_logits[transfer_index_t_s], temperature=temperature, top_p=top_p, top_k=top_k
                 )
                 x[mask_index] = x0.clone()
 
-                # origin 分支不做 RCR（与原版一致）
             else:
-                # 置信度算法：rcr=False 用 alg；rcr=True 用 conf_alg（与之前一致）
-                if (not rcr and alg == "maskgit_plus") or (rcr and conf_alg == "maskgit_plus"):
+                # 选择置信度算法
+                use_alg = alg
+                if rcr:
+                    # rcr=True 时，置信度算法由 conf_alg 决定（不影响 baseline）
+                    use_alg = conf_alg
+
+                if use_alg == "maskgit_plus":
                     confidence, x0 = sample_tokens(mask_logits, temperature=temperature, top_p=top_p, top_k=top_k)
-                elif (not rcr and alg == "topk_margin") or (rcr and conf_alg == "topk_margin"):
+                elif use_alg == "topk_margin":
                     confidence, x0 = sample_tokens(
                         mask_logits, temperature=temperature, top_p=top_p, top_k=top_k, margin_confidence=True
                     )
-                elif (not rcr and alg == "entropy") or (rcr and conf_alg == "entropy"):
+                elif use_alg == "entropy":
                     confidence, x0 = sample_tokens(
                         mask_logits, temperature, top_p=top_p, top_k=top_k, neg_entropy=True
                     )
                 else:
-                    if rcr:
-                        if alg == "maskgit_plus":
-                            confidence, x0 = sample_tokens(mask_logits, temperature=temperature, top_p=top_p, top_k=top_k)
-                        elif alg == "topk_margin":
-                            confidence, x0 = sample_tokens(
-                                mask_logits, temperature=temperature, top_p=top_p, top_k=top_k, margin_confidence=True
-                            )
-                        elif alg == "entropy":
-                            confidence, x0 = sample_tokens(
-                                mask_logits, temperature, top_p=top_p, top_k=top_k, neg_entropy=True
-                            )
-                        else:
-                            raise RuntimeError(f"Unknown alg: {alg}")
-                    else:
-                        raise RuntimeError(f"Unknown alg: {alg}")
-
-                # ===== baseline 的“选入”逻辑：原样保留 =====
-                num_mask_token = mask_index.sum() / mask_index.shape[0]
-                number_transfer_tokens = (
-                    int(num_mask_token * (1 - s / t)) if i < steps - 1 else int(num_mask_token)
+                    raise RuntimeError(f"Unknown alg: {alg}")
+
+                # 统一 full_confidence 的 dtype = logits.dtype（避免 int/float 混合）
+                full_confidence = torch.full(
+                    x.shape, float("-inf"), device=self.device, dtype=logits.dtype
                 )
-                full_confidence = torch.full_like(x, -torch.inf, device=self.device, dtype=logits.dtype)
                 full_confidence[mask_index] = confidence
 
-                if number_transfer_tokens > 0:
-                    if alg_temp is None or alg_temp == 0:
-                        _, transfer_index = torch.topk(full_confidence, number_transfer_tokens)
-                    else:
-                        full_confidence = full_confidence / alg_temp
-                        full_confidence = F.softmax(full_confidence, dim=-1)
-                        transfer_index = torch.multinomial(full_confidence, num_samples=number_transfer_tokens)
-
-                    x_ = torch.zeros_like(x, device=self.device, dtype=torch.long) + mask_token_id
-                    x_[mask_index] = x0.clone()
-                    row_indices = torch.arange(x.size(0), device=self.device).unsqueeze(1).expand_as(transfer_index)
-                    x[row_indices, transfer_index] = x_[row_indices, transfer_index]
-
-                    # ===== 仅 rcr=True：记录“已确定”的位置与它们的置信度（用于后续回遮）=====
-                    if rcr:
-                        is_fixed[row_indices, transfer_index] = True
-                        # 注意：这里存 baseline 使用的 full_confidence（与 baseline 完全一致）
-                        fixed_conf[row_indices, transfer_index] = full_confidence[row_indices, transfer_index]
-
-                # ===== 仅 rcr=True：在“选入”之后按累计目标回遮最低置信度的超额部分 =====
                 if rcr:
-                    # 这一步只回遮，完全不改变 baseline 的选入行为
-                    self._rcr_remask_after_selection(
+                    # === RCR 分支：先转移 top-k，再根据累计目标回遮 ===
+                    self._apply_rcr_logic(
                         x=x,
+                        x0_sel=x0,
+                        conf_sel=confidence,
+                        mask_index=mask_index,
+                        overtime_confidence=overtime_confidence,
                         mask_token_id=mask_token_id,
                         step=i,
-                        steps=steps,
+                        total_steps=steps,
                         s=s,
                         t=t,
-                        is_fixed=is_fixed,
-                        fixed_conf=fixed_conf,
                     )
-
+                else:
+                    # === baseline 分支：保持 Dream 逻辑不变 ===
+                    num_mask_token = mask_index.sum() / mask_index.shape[0]
+                    number_transfer_tokens = (
+                        int(num_mask_token * (1 - s / t)) if i < steps - 1 else int(num_mask_token)
+                    )
+                    if number_transfer_tokens > 0:
+                        if alg_temp is None or alg_temp == 0:
+                            _, transfer_index = torch.topk(full_confidence, number_transfer_tokens)
+                        else:
+                            fc = full_confidence / alg_temp
+                            fc = F.softmax(fc, dim=-1)
+                            transfer_index = torch.multinomial(fc, num_samples=number_transfer_tokens)
+                        x_ = torch.full_like(x, fill_value=mask_token_id, dtype=torch.long, device=self.device)
+                        x_[mask_index] = x0.clone()
+                        row_indices = torch.arange(x.size(0), device=self.device).unsqueeze(1).expand_as(transfer_index)
+                        x[row_indices, transfer_index] = x_[row_indices, transfer_index]
+
+            # 允许用户控制中间 tokens
             x = generation_tokens_hook_func(i, x, logits)
 
             if histories is not None: