Skip to content

Added support for SFTTrainer checkpoint models and adapter models containing some non-LoRA weights #9778

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 2 commits into from
Closed

Added support for SFTTrainer checkpoint models and adapter models containing some non-LoRA weights #9778

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c6396aa
Added support for SFTTrainer checkpoint models and adapter models con…
Victoran0 Oct 8, 2024
c2c2626
Added support for SFTTrainer checkpoint models and adapter models con…
Victoran0 Oct 8, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
51 changes: 31 additions & 20 deletions convert_lora_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -338,28 +338,39 @@ def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
tensor_map: dict[str, PartialLoraTensor] = {}

# The following edits will enable conversion for: SFTTrainer checkpoint adapter models and other adapter models that contain weights besides LoRA weights

# Here, we first get the items with the 'lora_' substring
lora_model_items_name = [name for name,_ in lora_model.items()]
lora_model_items_with_lora_tensors = [name for name in lora_model_items_name if 'lora_' in name]

for name, tensor in lora_model.items():
if self.lazy:
tensor = LazyTorchTensor.from_eager(tensor)
base_name = get_base_tensor_name(name)
is_lora_a = ".lora_A.weight" in name
is_lora_b = ".lora_B.weight" in name
if not is_lora_a and not is_lora_b:
if ".base_layer.weight" in name:
continue
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
sys.exit(1)

if base_name in tensor_map:
if is_lora_a:
tensor_map[base_name].A = tensor
else:
tensor_map[base_name].B = tensor
else:
if is_lora_a:
tensor_map[base_name] = PartialLoraTensor(A=tensor)

# Check for only LoRA finetuned weights and base layer weights
if (name in lora_model_items_with_lora_tensors) or (".base_layer.weight" in name):
if self.lazy:
tensor = LazyTorchTensor.from_eager(tensor)
base_name = get_base_tensor_name(name)
is_lora_a = ".lora_A.weight" in name
is_lora_b = ".lora_B.weight" in name
if not is_lora_a and not is_lora_b:
if ".base_layer.weight" in name:
continue

# we will either have a lora weight or a base layer weight, this error becomes trivial
# logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
# sys.exit(1)

if base_name in tensor_map:
if is_lora_a:
tensor_map[base_name].A = tensor
else:
tensor_map[base_name].B = tensor
else:
tensor_map[base_name] = PartialLoraTensor(B=tensor)
if is_lora_a:
tensor_map[base_name] = PartialLoraTensor(A=tensor)
else:
tensor_map[base_name] = PartialLoraTensor(B=tensor)

for name, tensor in tensor_map.items():
assert tensor.A is not None
Expand Down