Fixed saving! (#113)
* 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 * Fix quantization_method * Fix quantization_config * patch model * Update llama.py * Update llama.py * Update llama.py * Update save.py * Update save.py * tokenizer_save_settings * Update save.py * quantization and loftq * Update save.py * Update llama.py * Update save.py * upload_to_huggingface * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py * Update save.py
This commit is contained in:
Daniel Han
GitHub
parent
bb05d6b6e2
commit
3d05a74b12
1 changed files with 103 additions and 53 deletions
156
unsloth/save.py
156
unsloth/save.py
|
|
@ -208,6 +208,15 @@ def unsloth_save_model(
|
|||
return save_directory
|
||||
pass
|
||||
|
||||
# Update model tag
|
||||
username = ""
|
||||
if push_to_hub:
|
||||
username = upload_to_huggingface(
|
||||
model, save_directory, token,
|
||||
"finetuned", "trl", file_location = None,
|
||||
)
|
||||
pass
|
||||
|
||||
# If push_to_hub, we must remove the .../ part of a repo
|
||||
if push_to_hub and "/" in save_directory:
|
||||
|
||||
|
|
@ -331,6 +340,7 @@ def unsloth_save_model(
|
|||
if (torch.cuda.memory_allocated() + W.nbytes) < max_vram:
|
||||
# Save to GPU memory
|
||||
state_dict[name] = W
|
||||
# [TODO] Saving to RAM seems to leak memory???
|
||||
# elif (max_ram - W.nbytes) > 0:
|
||||
# # Save to CPU memory
|
||||
# logger.warning_once(f"We will save to RAM and not VRAM now.")
|
||||
|
|
@ -401,27 +411,32 @@ def unsloth_save_model(
|
|||
model.config = old_config
|
||||
print("Done.")
|
||||
|
||||
# Print location
|
||||
if push_to_hub:
|
||||
print(f"Saved to https://huggingface.co/{username}/{save_directory.lstrip('/')}")
|
||||
pass
|
||||
|
||||
save_pretrained_settings["state_dict"] = None
|
||||
|
||||
# for j, (key, value) in enumerate(state_dict.items()):
|
||||
# state_dict[key] = None
|
||||
# if j % 10 == 0:
|
||||
# torch.cuda.empty_cache()
|
||||
# gc.collect()
|
||||
# pass
|
||||
# pass
|
||||
# state_dict = None
|
||||
# del state_dict
|
||||
# torch.cuda.empty_cache()
|
||||
# gc.collect()
|
||||
for j, (key, value) in enumerate(state_dict.items()):
|
||||
state_dict[key] = None
|
||||
if j % 10 == 0:
|
||||
torch.cuda.empty_cache()
|
||||
gc.collect()
|
||||
pass
|
||||
pass
|
||||
state_dict = None
|
||||
del state_dict
|
||||
torch.cuda.empty_cache()
|
||||
gc.collect()
|
||||
|
||||
# Remove temporary location
|
||||
import shutil
|
||||
shutil.rmtree(temporary_location)
|
||||
|
||||
# for _ in range(3):
|
||||
# torch.cuda.empty_cache()
|
||||
# gc.collect()
|
||||
for _ in range(3):
|
||||
torch.cuda.empty_cache()
|
||||
gc.collect()
|
||||
return save_directory
|
||||
pass
|
||||
|
||||
|
|
@ -629,14 +644,44 @@ def unsloth_push_to_hub_merged(
|
|||
pass
|
||||
|
||||
|
||||
def upload_gguf_to_huggingface(save_directory, file_location, token, model_type):
|
||||
print("Unsloth: Uploading GGUF to Huggingface Hub...")
|
||||
MODEL_CARD = \
|
||||
"""---
|
||||
base_model: {base_model}
|
||||
tags:
|
||||
- text-generation-inference
|
||||
- transformers
|
||||
- unsloth
|
||||
- {model_type}
|
||||
- {extra}
|
||||
license: apache-2.0
|
||||
language:
|
||||
- en
|
||||
---
|
||||
|
||||
# Uploaded {method} model
|
||||
|
||||
- **Developed by:** {username}
|
||||
- **License:** apache-2.0
|
||||
- **Finetuned from model :** {base_model}
|
||||
|
||||
This {model_type} model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
|
||||
|
||||
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
|
||||
"""
|
||||
|
||||
|
||||
def upload_to_huggingface(model, save_directory, token, method, extra = "", file_location = None):
|
||||
# Check for username
|
||||
username = ""
|
||||
if "/" not in save_directory:
|
||||
from huggingface_hub import whoami
|
||||
try: save_directory = f"{save_directory}/{whoami()['name']}"
|
||||
except: pass
|
||||
try:
|
||||
username = whoami()['name']
|
||||
save_directory = f"{save_directory}/{username}"
|
||||
except:
|
||||
raise RuntimeError(f"Unsloth: {save_directory} is not a Huggingface directory.")
|
||||
else:
|
||||
username = save_directory.split("/")[0]
|
||||
pass
|
||||
|
||||
from huggingface_hub import create_repo
|
||||
|
|
@ -648,43 +693,36 @@ def upload_gguf_to_huggingface(save_directory, file_location, token, model_type)
|
|||
)
|
||||
|
||||
# Create model card
|
||||
from huggingface_hub import ModelCard, ModelCardData
|
||||
card_data = ModelCardData(
|
||||
language = "en",
|
||||
license = "apache-2.0",
|
||||
library = "unsloth",
|
||||
tags = ["gguf", "unsloth", "text-generation-inference", "transformers",],
|
||||
from huggingface_hub import ModelCard
|
||||
content = MODEL_CARD.format(
|
||||
username = username,
|
||||
base_model = model.config._name_or_path,
|
||||
model_type = model.config.model_type,
|
||||
method = "",
|
||||
extra = extra,
|
||||
)
|
||||
|
||||
content = f"\n"\
|
||||
f"---\n"\
|
||||
f"{ card_data.to_yaml() }\n"\
|
||||
f"---\n"\
|
||||
f"\n"\
|
||||
f"# My Model Card for {file_location}\n"\
|
||||
f"\n"\
|
||||
f"\nThis {model_type.title()} model was trained by [Unsloth](https://github.com/unslothai/unsloth) then saved to GGUF.\n"\
|
||||
f"\n"
|
||||
|
||||
card = ModelCard(content)
|
||||
card.push_to_hub(save_directory, token = token)
|
||||
|
||||
# Now upload file
|
||||
from huggingface_hub import HfApi
|
||||
hf_api = HfApi(token = token)
|
||||
if file_location is not None:
|
||||
# Now upload file
|
||||
from huggingface_hub import HfApi
|
||||
hf_api = HfApi(token = token)
|
||||
|
||||
if "/" in file_location:
|
||||
uploaded_location = file_location[file_location.rfind("/")+1:]
|
||||
else:
|
||||
uploaded_location = file_location
|
||||
if "/" in file_location:
|
||||
uploaded_location = file_location[file_location.rfind("/")+1:]
|
||||
else:
|
||||
uploaded_location = file_location
|
||||
pass
|
||||
|
||||
hf_api.upload_file(
|
||||
path_or_fileobj = file_location,
|
||||
path_in_repo = uploaded_location,
|
||||
repo_id = save_directory,
|
||||
repo_type = "model",
|
||||
)
|
||||
pass
|
||||
|
||||
hf_api.upload_file(
|
||||
path_or_fileobj = file_location,
|
||||
path_in_repo = uploaded_location,
|
||||
repo_id = save_directory,
|
||||
repo_type = "model",
|
||||
)
|
||||
return username
|
||||
pass
|
||||
|
||||
|
||||
|
|
@ -754,8 +792,15 @@ def unsloth_save_pretrained_gguf(
|
|||
gc.collect()
|
||||
|
||||
file_location = save_to_gguf(new_save_directory, quantization_method, makefile)
|
||||
model_type = self.config.model_type
|
||||
if push_to_hub: upload_gguf_to_huggingface(new_save_directory, file_location, token, model_type)
|
||||
|
||||
if push_to_hub:
|
||||
print("Unsloth: Uploading GGUF to Huggingface Hub...")
|
||||
username = upload_to_huggingface(
|
||||
self, model, new_save_directory, token,
|
||||
"GGUF converted", "gguf", file_location,
|
||||
)
|
||||
print(f"Saved to https://huggingface.co/{username}/{new_save_directory.lstrip('/')}")
|
||||
pass
|
||||
pass
|
||||
|
||||
|
||||
|
|
@ -827,8 +872,13 @@ def unsloth_push_to_hub_gguf(
|
|||
|
||||
python_install.wait()
|
||||
file_location = save_to_gguf(new_save_directory, quantization_method, makefile)
|
||||
model_type = self.config.model_type
|
||||
upload_gguf_to_huggingface(new_save_directory, file_location, token, model_type)
|
||||
|
||||
print("Unsloth: Uploading GGUF to Huggingface Hub...")
|
||||
username = upload_to_huggingface(
|
||||
self, model, new_save_directory, token,
|
||||
"GGUF converted", "gguf", file_location,
|
||||
)
|
||||
print(f"Saved to https://huggingface.co/{username}/{new_save_directory.lstrip('/')}")
|
||||
pass
|
||||
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue