LLaVa: add cache class attribute (#32278)

cache class flag

src/transformers/models/llava/modeling_llava.py

@@ -126,6 +126,7 @@ class LlavaPreTrainedModel(PreTrainedModel):
     _no_split_modules = ["LlavaVisionAttention"]
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
+    _supports_cache_class = True
 
     def _init_weights(self, module):
         # important: this ported version of Llava isn't meant for training from scratch - only

src/transformers/models/llava_next/modeling_llava_next.py

@@ -232,6 +232,7 @@ class LlavaNextPreTrainedModel(PreTrainedModel):
     _no_split_modules = ["LlavaNextVisionAttention"]
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
+    _supports_cache_class = True
 
     def _init_weights(self, module):
         # important: this ported version of LlavaNext isn't meant for training from scratch - only

src/transformers/models/llava_next_video/modeling_llava_next_video.py

@@ -272,6 +272,7 @@ class LlavaNextVideoPreTrainedModel(PreTrainedModel):
     _no_split_modules = ["LlavaNextVideoVisionAttention"]
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
+    _supports_cache_class = True
 
     def _init_weights(self, module):
         # important: this ported version of LlavaNextVideo isn't meant for training from scratch - only

src/transformers/models/paligemma/modeling_paligemma.py

@@ -127,6 +127,7 @@ class PaliGemmaPreTrainedModel(PreTrainedModel):
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = False
     _supports_sdpa = True
+    _supports_cache_class = True
 
     def _init_weights(self, module):
         # important: this ported version of PaliGemmaisn't meant for training from scratch - only

src/transformers/models/video_llava/modeling_video_llava.py

@@ -126,6 +126,7 @@ class VideoLlavaPreTrainedModel(PreTrainedModel):
     _no_split_modules = ["VideoLlavaVisionAttention"]
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
+    _supports_cache_class = True
 
     def _init_weights(self, module):
         std = (

src/transformers/models/vipllava/modeling_vipllava.py

@@ -135,6 +135,7 @@ class VipLlavaPreTrainedModel(PreTrainedModel):
     _no_split_modules = ["VipLlavaVisionAttention"]
     _skip_keys_device_placement = "past_key_values"
     _supports_flash_attn_2 = True
+    _supports_cache_class = True
 
     def _init_weights(self, module):
         # important: this ported version of VipLlava isn't meant for training from scratch - only