ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add llava-next/llava-next-video/video-llava

Browse Source 
Former-commit-id: a4e4239931b0b0e3fd12c9f9bbfd2c201cbc78ca
This commit is contained in:
BUAADreamer
2024-09-28 00:57:03 +08:00
parent 86c84972c8
commit b6fb00e046
5 changed files with 134 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/llamafactory/model/model_utils/visual.py
View File

@@ -92,7 +92,7 @@ def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArgumen
if getattr(model, "quantization_method", None):
model_type = getattr(model.config, "model_type", None)
if model_type in ["llava", "paligemma"]:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
elif model_type == "qwen2_vl":
mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
@@ -111,9 +111,8 @@ def configure_visual_model(config: "PretrainedConfig") -> None:
if model_type in [
"llava",
"llava_next",
"video_llava",
"idefics2",
"llava_next_video",
"video_llava",
]: # required for ds zero3 and valuehead models
setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
@@ -128,7 +127,7 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
"""
model_type = getattr(config, "model_type", None)
forbidden_modules = set()
if model_type in ["llava", "paligemma"]:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vision_tower")
@@ -170,7 +169,7 @@ def patch_target_modules(
"""
model_type = getattr(config, "model_type", None)
if finetuning_args.freeze_vision_tower:
if model_type in ["llava", "paligemma"]:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
elif model_type == "qwen2_vl":
return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))