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

add more llava-next series template

Browse Source 
Former-commit-id: 93f64f2aebf41582d39aa8a2c6059e562ca694b0
This commit is contained in:
BUAADreamer
2024-09-29 21:29:29 +08:00
parent 4d43317aa1
commit ab5d042cd3
6 changed files with 141 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
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", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
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")
@@ -108,12 +108,7 @@ def configure_visual_model(config: "PretrainedConfig") -> None:
Patches VLMs before loading them.
"""
model_type = getattr(config, "model_type", None)
if model_type in [
"llava",
"llava_next",
"llava_next_video",
"video_llava",
]: # required for ds zero3 and valuehead models
if model_type in ["llava", "video_llava"] or "llava_next" in model_type: # required for ds zero3 and valuehead models
setattr(config, "hidden_size", getattr(config.text_config, "hidden_size", None))
if getattr(config, "is_yi_vl_derived_model", None):
@@ -127,7 +122,7 @@ def get_forbidden_modules(config: "PretrainedConfig", finetuning_args: "Finetuni
"""
model_type = getattr(config, "model_type", None)
forbidden_modules = set()
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
if finetuning_args.freeze_vision_tower:
forbidden_modules.add("vision_tower")
@@ -185,7 +180,7 @@ def patch_target_modules(
"""
model_type = getattr(config, "model_type", None)
if finetuning_args.freeze_vision_tower:
if model_type in ["llava", "llava_next", "llava_next_video", "paligemma", "video_llava"]:
if model_type in ["llava", "paligemma", "video_llava"] or "llava_next" in model_type:
return "^(?!.*vision_tower).*(?:{}).*".format("|".join(target_modules))
elif model_type == "qwen2_vl":
return "^(?!.*visual).*(?:{}).*".format("|".join(target_modules))