423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-23 22:32:54 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix visual patch

Browse Source 
Former-commit-id: 75b586c31acf47d1bd28e04566ffd1d954e45596
This commit is contained in:
hoshi-hiyouga 2024-11-25 20:06:06 +08:00 committed by GitHub
parent f48d07cd6c
commit a489f10986
1 changed files with 14 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
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): if getattr(model, "quantization_method", None):
model_type = getattr(model.config, "model_type", None) model_type = getattr(model.config, "model_type", None)
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "pixtral", "video_llava"]: if model_type in ["llava", "llava_next", "llava_next_video", "mllama", "paligemma", "video_llava"]:
mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector") mm_projector: "torch.nn.Module" = getattr(model, "multi_modal_projector")
elif model_type == "qwen2_vl": elif model_type == "qwen2_vl":
mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger") mm_projector: "torch.nn.Module" = getattr(getattr(model, "visual"), "merger")
@ -108,15 +108,8 @@ def configure_visual_model(config: "PretrainedConfig") -> None:
Patches VLMs before loading them. Patches VLMs before loading them.
""" """
model_type = getattr(config, "model_type", None) model_type = getattr(config, "model_type", None)
if model_type in [ if model_type in ["llava", "llava_next", "llava_next_video", "mllama", "paligemma", "video_llava"]:
"llava", # required for ds zero3 and valuehead models
"llava_next",
"llava_next_video",
"mllama",
"paligemma",
"pixtral",
"video_llava",
]: # required for ds zero3 and valuehead models
setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None)) setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
if getattr(config, "is_yi_vl_derived_model", None): if getattr(config, "is_yi_vl_derived_model", None):
@ -130,13 +123,20 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
""" """
model_type = getattr(config, "model_type", None) model_type = getattr(config, "model_type", None)
forbidden_modules = set() forbidden_modules = set()
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "pixtral", "video_llava"]: if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if finetuning_args.freeze_vision_tower: if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vision_tower") forbidden_modules.add("vision_tower")
if finetuning_args.train_mm_proj_only: if finetuning_args.train_mm_proj_only:
forbidden_modules.add("language_model") forbidden_modules.add("language_model")
elif model_type == "mllama":
if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vision_model")
if finetuning_args.train_mm_proj_only:
forbidden_modules.add("language_model")
elif model_type == "qwen2_vl": elif model_type == "qwen2_vl":
if finetuning_args.freeze_vision_tower: if finetuning_args.freeze_vision_tower:
forbidden_modules.add("visual") forbidden_modules.add("visual")
@ -189,8 +189,9 @@ def patch_target_modules(
Freezes vision tower for VLM LoRA tuning. Freezes vision tower for VLM LoRA tuning.
""" """
model_type = getattr(config, "model_type", None) model_type = getattr(config, "model_type", None)
vit_model_type = getattr(getattr(config, "vision_config", None), "model_type", None)
if finetuning_args.freeze_vision_tower: if finetuning_args.freeze_vision_tower:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "pixtral", "video_llava"]: if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
elif model_type == "mllama": elif model_type == "mllama":
return "^(?!.*vision_model).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*vision_model).*(?:{}).*".format("|".join(target_modules))
@ -201,7 +202,7 @@ def patch_target_modules(
else: else:
if model_type == "qwen2_vl": if model_type == "qwen2_vl":
return "^(?!.*patch_embed).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*patch_embed).*(?:{}).*".format("|".join(target_modules))
elif model_type == "pixtral": elif vit_model_type == "pixtral":
return "^(?!.*patch_conv).*(?:{}).*".format("|".join(target_modules)) return "^(?!.*patch_conv).*(?:{}).*".format("|".join(target_modules))
else: else:
return target_modules return target_modules