Quick fixes (#106)
* Fix tokenizer, dropout, bias for LoRA * Update loader.py * Fix LoRA downcasting * Update _utils.py * Saving to GGUF * fix * colab_quantize_to_gguf * move save modules * save module * Update __init__.py * Update save.py * Temp downgrade due to TRL issue * Fix up bugs * Faster saving + other changes * Update llama.py * Saving modules * spelling * Update llama.py * Update save.py * Update save.py * Update loader.py * Update llama.py * patch saving * Update save.py * Update save.py * Update save.py * patch saving * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * original_model * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * saving to RAM leakage? * Update save.py * new_save_directory * Update save.py * Update save.py * Update save.py * Update save.py * Update pyproject.toml * Update pyproject.toml * Update pyproject.toml * Quick fixes * Update llama.py * Update llama.py * Update dpo.py * Update dpo.py * Update llama.py * Update save.py * getattr * RSLoRA and LoftQ direct support * Update llama.py * Update llama.py * Update llama.py * Fix DPO + GGUF
This commit is contained in:
Daniel Han
GitHub
parent
52b5ef31e0
commit
12e75c93d0
4 changed files with 33 additions and 26 deletions
|
|
@ -242,7 +242,17 @@ def LoraLayer_update_layer(self, adapter_name, r, lora_alpha, lora_dropout, init
|
|||
self.scaling[adapter_name] = lora_alpha / r
|
||||
|
||||
if init_lora_weights == "loftq":
|
||||
# We manually check for PEFT
|
||||
if not hasattr(self, "loftq_init"):
|
||||
import peft
|
||||
raise RuntimeError(
|
||||
f"Unsloth: Your PEFT version of {peft.__version__} does not support LoftQ init.\n"\
|
||||
"Please install PEFT 0.7.2 or higher.\n"\
|
||||
"You can also install from source: `pip install git+https://github.com/huggingface/peft.git"
|
||||
)
|
||||
pass
|
||||
self.loftq_init(adapter_name)
|
||||
|
||||
elif init_lora_weights:
|
||||
self.reset_lora_parameters(adapter_name, init_lora_weights)
|
||||
|
||||
|
|
|
|||
|
|
@ -28,6 +28,7 @@ DPOTrainer_metrics = [
|
|||
"logits/rejected",
|
||||
"logits/chosen",
|
||||
]
|
||||
set_DPOTrainer_metrics = frozenset(DPOTrainer_metrics)
|
||||
|
||||
|
||||
def NotebookProgressCallback_on_train_begin(self, args, state, control, **kwargs):
|
||||
|
|
@ -47,16 +48,7 @@ def NotebookProgressCallback_on_log(self, args, state, control, logs=None, **kwa
|
|||
if args.evaluation_strategy == IntervalStrategy.NO and "loss" in logs:
|
||||
values = {"Training Loss": logs["loss"]}
|
||||
for metric in DPOTrainer_metrics:
|
||||
if metric in logs:
|
||||
values[metric.replace("/", " / ")] = logs[metric]
|
||||
else:
|
||||
# Maybe not a DPO Trainer anymore? Redo the tracker
|
||||
column_names = [self.first_column] + ["Training Loss"]
|
||||
if args.evaluation_strategy != IntervalStrategy.NO:
|
||||
column_names.append("Validation Loss")
|
||||
self.training_tracker = NotebookTrainingTracker(state.max_steps, column_names)
|
||||
break
|
||||
pass
|
||||
values[metric.replace("/", " / ")] = logs[metric]
|
||||
pass
|
||||
# First column is necessarily Step since we're not in epoch eval strategy
|
||||
values["Step"] = state.global_step
|
||||
|
|
@ -76,10 +68,16 @@ def NotebookTrainingTracker_write_line(self, values):
|
|||
self.inner_table = [list(values.keys()), list(values.values())]
|
||||
else:
|
||||
columns = self.inner_table[0]
|
||||
print(columns)
|
||||
for key in values.keys():
|
||||
if key not in columns:
|
||||
columns.append(key)
|
||||
new_values = {}
|
||||
for key, value in values.items():
|
||||
lowered = key.lower()
|
||||
if lowered in set_DPOTrainer_metrics:
|
||||
new_values[lowered.replace("/", " / ")] = value
|
||||
else:
|
||||
new_values[key] = value
|
||||
pass
|
||||
values = new_values
|
||||
|
||||
self.inner_table[0] = columns
|
||||
if len(self.inner_table) > 1:
|
||||
last_values = self.inner_table[-1]
|
||||
|
|
@ -104,7 +102,7 @@ pass
|
|||
|
||||
def PatchDPOTrainer():
|
||||
# Patch DPO notebook printing
|
||||
# NotebookTrainingTracker.write_line = NotebookTrainingTracker_write_line
|
||||
NotebookTrainingTracker.write_line = NotebookTrainingTracker_write_line
|
||||
from transformers.trainer import DEFAULT_PROGRESS_CALLBACK
|
||||
DEFAULT_PROGRESS_CALLBACK.on_train_begin = NotebookProgressCallback_on_train_begin
|
||||
DEFAULT_PROGRESS_CALLBACK.on_log = NotebookProgressCallback_on_log
|
||||
|
|
|
|||
|
|
@ -912,14 +912,13 @@ class FastLlamaModel:
|
|||
assert(type(use_rslora) is bool)
|
||||
if use_rslora:
|
||||
if not SUPPORTS_RSLORA:
|
||||
# We do it ourselves!
|
||||
new_alpha = lora_alpha / (r**0.5)
|
||||
# We manually check for PEFT
|
||||
import peft
|
||||
logger.warning_once(
|
||||
f"Unsloth: Your PEFT version of {peft.__version__} (0.7.2 needed) does not support `use_rslora` natively.\n"\
|
||||
f"But, we do it ourselves by setting `alpha = {new_alpha}.`"
|
||||
raise RuntimeError(
|
||||
f"Unsloth: Your PEFT version of {peft.__version__} does not support `use_rslora`.\n"\
|
||||
"Please install PEFT 0.7.2 or higher.\n"\
|
||||
"You can also install from source: `pip install git+https://github.com/huggingface/peft.git"
|
||||
)
|
||||
lora_alpha = new_alpha
|
||||
pass
|
||||
pass
|
||||
|
||||
|
|
|
|||
|
|
@ -555,7 +555,7 @@ def unsloth_push_to_hub_merged(
|
|||
self,
|
||||
repo_id : str,
|
||||
tokenizer = None,
|
||||
save_method : str = "merged_16bit", # ["lora", "merged_16bit", "merged_4bit"]
|
||||
save_method : str = "merged_16bit", # ["lora", "merged_16bit", "merged_4bit"]
|
||||
use_temp_dir : Optional[bool] = None,
|
||||
commit_message : Optional[str] = None,
|
||||
private : Optional[bool] = None,
|
||||
|
|
@ -601,7 +601,7 @@ def unsloth_save_pretrained_gguf(
|
|||
self,
|
||||
save_directory : Union[str, os.PathLike],
|
||||
tokenizer = None,
|
||||
quantization_method : str = "fast_quantized",
|
||||
quantization_method : str = "fast_quantized",
|
||||
push_to_hub : bool = False,
|
||||
token : Optional[Union[str, bool]] = None,
|
||||
is_main_process : bool = True,
|
||||
|
|
@ -649,7 +649,7 @@ def unsloth_save_pretrained_gguf(
|
|||
arguments["push_to_hub"] = False # We save ourselves
|
||||
arguments["save_method"] = "merged_16bit" # Must be 16bit
|
||||
del arguments["self"]
|
||||
del arguments["quantization"]
|
||||
del arguments["quantization_method"]
|
||||
|
||||
# Non blocking install GGUF first
|
||||
git_clone = install_llama_cpp_clone_non_blocking()
|
||||
|
|
@ -699,7 +699,7 @@ def unsloth_push_to_hub_gguf(
|
|||
self,
|
||||
repo_id : str,
|
||||
tokenizer = None,
|
||||
quantization_method : str = "fast_quantized",
|
||||
quantization_method : str = "fast_quantized",
|
||||
use_temp_dir : Optional[bool] = None,
|
||||
commit_message : Optional[str] = None,
|
||||
private : Optional[bool] = None,
|
||||
|
|
@ -749,7 +749,7 @@ def unsloth_push_to_hub_gguf(
|
|||
arguments["save_method"] = "merged_16bit" # Must be 16bit
|
||||
del arguments["self"]
|
||||
del arguments["repo_id"]
|
||||
del arguments["quantization"]
|
||||
del arguments["quantization_method"]
|
||||
|
||||
# Non blocking install GGUF first
|
||||
git_clone = install_llama_cpp_clone_non_blocking()
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue