[Bugfix]Disable the post_norm layer of the vision encoder for LLaVA models (#9653)
This commit is contained in:
@@ -277,7 +277,8 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
self.multimodal_config = multimodal_config
|
||||
|
||||
# TODO: Optionally initializes this for supporting embeddings.
|
||||
self.vision_tower = init_vision_tower_for_llava(config, quant_config)
|
||||
self.vision_tower = init_vision_tower_for_llava(
|
||||
config, quant_config, require_post_norm=False)
|
||||
self.image_newline = nn.Parameter(
|
||||
torch.empty(config.text_config.hidden_size))
|
||||
self.multi_modal_projector = LlavaMultiModalProjector(
|
||||
|
||||
Reference in New Issue
Block a user