修复了一个无关紧要的warning

indextts/accel/gpt2_accel.py

@@ -1,7 +1,9 @@
 import torch
 import torch.nn as nn
 from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions
-from transformers.models.gpt2.modeling_gpt2 import Conv1D, GPT2Block, GPT2Model
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.models.gpt2.modeling_gpt2 import Conv1D, GPT2Block
+from indextts.gpt.transformers_gpt2 import GPT2Model
 
 from .attention import Attention
 

indextts/gpt/model.py

@@ -5,7 +5,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 import transformers
-from transformers import GPT2Config, LogitsProcessorList
+from transformers import GPT2Config, LogitsProcessorList, DynamicCache
+from transformers.cache_utils import Cache
 from indextts.gpt.transformers_gpt2 import GPT2PreTrainedModel, GPT2Model
 
 # from transformers import GPT2Config, GPT2PreTrainedModel, LogitsProcessorList
@@ -91,8 +92,13 @@ class GPT2InferenceModel(GPT2PreTrainedModel):
         token_type_ids = kwargs.get("token_type_ids", None)  # usually None
         if not self.kv_cache:
             past_key_values = None
+        
+        # 将旧格式的元组转换为 DynamicCache（支持 v4.53.0+）
+        if past_key_values is not None and not isinstance(past_key_values, Cache):
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+        
         # only last token for inputs_ids if past is defined in kwargs
-        if past_key_values:
+        if past_key_values is not None:
             input_ids = input_ids[:, -1].unsqueeze(-1)
             if token_type_ids is not None:
                 token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
@@ -104,7 +110,7 @@ class GPT2InferenceModel(GPT2PreTrainedModel):
             # create position_ids on the fly for batch generation
             position_ids = attention_mask.long().cumsum(-1) - 1
             position_ids.masked_fill_(attention_mask == 0, 0)
-            if past_key_values:
+            if past_key_values is not None:
                 position_ids = position_ids[:, -1].unsqueeze(-1)
         else:
             position_ids = None

indextts/gpt/model_v2.py

@@ -5,7 +5,8 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 import transformers
-from transformers import GPT2Config, LogitsProcessorList
+from transformers import GPT2Config, LogitsProcessorList, DynamicCache
+from transformers.cache_utils import Cache
 from indextts.gpt.transformers_gpt2 import GPT2PreTrainedModel, GPT2Model
 
 # from transformers import GPT2Config, GPT2PreTrainedModel, LogitsProcessorList
@@ -91,8 +92,13 @@ class GPT2InferenceModel(GPT2PreTrainedModel):
         token_type_ids = kwargs.get("token_type_ids", None)  # usually None
         if not self.kv_cache:
             past_key_values = None
+        
+        # 将旧格式的元组转换为 DynamicCache（支持 v4.53.0+）
+        if past_key_values is not None and not isinstance(past_key_values, Cache):
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+        
         # only last token for inputs_ids if past is defined in kwargs
-        if past_key_values:
+        if past_key_values is not None:
             input_ids = input_ids[:, -1].unsqueeze(-1)
             if token_type_ids is not None:
                 token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
@@ -104,7 +110,7 @@ class GPT2InferenceModel(GPT2PreTrainedModel):
             # create position_ids on the fly for batch generation
             position_ids = attention_mask.long().cumsum(-1) - 1
             position_ids.masked_fill_(attention_mask == 0, 0)
-            if past_key_values:
+            if past_key_values is not None:
                 position_ids = position_ids[:, -1].unsqueeze(-1)
         else:
             position_ids = None

indextts/gpt/transformers_gpt2.py

@@ -29,6 +29,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from transformers.activations import ACT2FN
 import transformers
+from transformers.cache_utils import Cache, DynamicCache
 
 from indextts.gpt.transformers_generation_utils import GenerationMixin
 from indextts.gpt.transformers_modeling_utils import PreTrainedModel
@@ -1023,9 +1024,12 @@ class GPT2Model(GPT2PreTrainedModel):
         if token_type_ids is not None:
             token_type_ids = token_type_ids.view(-1, input_shape[-1])
 
+        # 支持 Cache 类（v4.53.0+）和旧版元组格式
         if past_key_values is None:
             past_length = 0
             past_key_values = tuple([None] * len(self.h))
+        elif isinstance(past_key_values, Cache):
+            past_length = past_key_values.get_seq_length()
         else:
             past_length = past_key_values[0][0].size(-2)
         if position_ids is None:
@@ -1107,7 +1111,15 @@ class GPT2Model(GPT2PreTrainedModel):
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
         all_hidden_states = () if output_hidden_states else None
-        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+        
+        # 处理 Cache 类的迭代
+        if isinstance(past_key_values, Cache):
+            # Cache 类通过 __getitem__ 访问各层
+            past_key_values_iter = [past_key_values[i] if i < len(self.h) else None for i in range(len(self.h))]
+        else:
+            past_key_values_iter = past_key_values
+        
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values_iter)):
             # Model parallel
             if self.model_parallel:
                 torch.cuda.set_device(hidden_states.device)
@@ -1175,6 +1187,11 @@ class GPT2Model(GPT2PreTrainedModel):
                 if v is not None
             )
 
+        # 如果输入是 Cache 类，输出也应该是 Cache 类
+        if use_cache and presents is not None:
+            if isinstance(past_key_values, Cache):
+                presents = DynamicCache.from_legacy_cache(presents)
+        
         return BaseModelOutputWithPastAndCrossAttentions(
             last_hidden_state=hidden_states,
             past_key_values=presents,