ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Bug fixes

Browse source
This commit is contained in:
Daniel Han 2024-08-15 15:04:46 -07:00
commit 2b7f769e13
2 changed files with 26 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

15
unsloth/__init__.py
View file

@ -11,10 +11,8 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import warnings
import importlib
import sys
import warnings, importlib, sys
from packaging.version import Version
# # Define a list of modules to check
@ -60,9 +58,8 @@ except:
"We have some installation instructions on our Github page.")
pass
import os, re
import os, re, subprocess, inspect
import numpy as np
import subprocess
# Hugging Face Hub faster downloads (only enable during Colab and Kaggle sessions)
keynames = "\n" + "\n".join(os.environ.keys())
@ -83,12 +80,12 @@ elif (major_torch == 2) and (minor_torch < 2):
del os.environ["PYTORCH_CUDA_ALLOC_CONF"]
pass
# Torch 2.5 has including_emulation
# Torch 2.4 has including_emulation
major_version, minor_version = torch.cuda.get_device_capability()
SUPPORTS_BFLOAT16 = (major_version >= 8)
if (major_torch == 2) and (minor_torch >= 5):
old_is_bf16_supported = torch.cuda.is_bf16_supported
old_is_bf16_supported = torch.cuda.is_bf16_supported
if "including_emulation" in str(inspect.signature(old_is_bf16_supported)):
def is_bf16_supported(including_emulation = False):
return old_is_bf16_supported(including_emulation)
torch.cuda.is_bf16_supported = is_bf16_supported

24
unsloth/models/loader.py
View file

@ -169,13 +169,23 @@ class FastLanguageModel(FastLlamaModel):
autoconfig_error = None
peft_error = None
try:
model_config = AutoConfig.from_pretrained(model_name, token = token, revision = revision)
model_config = AutoConfig.from_pretrained(
model_name,
token = token,
revision = revision,
trust_remote_code = trust_remote_code,
)
is_model = True
except Exception as error:
autoconfig_error = str(error)
is_model = False
try:
peft_config = PeftConfig .from_pretrained(model_name, token = token, revision = revision)
peft_config = PeftConfig.from_pretrained(
model_name,
token = token,
revision = revision,
trust_remote_code = trust_remote_code,
)
is_peft = True
except Exception as error:
peft_error = str(error)
@ -207,7 +217,12 @@ class FastLanguageModel(FastLlamaModel):
if is_peft:
# Check base model again for PEFT
model_name = get_model_name(peft_config.base_model_name_or_path, load_in_4bit)
model_config = AutoConfig.from_pretrained(model_name, token = token, revision = revision)
model_config = AutoConfig.from_pretrained(
model_name,
token = token,
revision = revision,
trust_remote_code = trust_remote_code,
)
pass
if not was_disabled: enable_progress_bars()
@ -340,10 +355,11 @@ class FastLanguageModel(FastLlamaModel):
token = token,
revision = revision,
is_trainable = True,
trust_remote_code = trust_remote_code,
)
# Patch it as well!
model = dispatch_model.patch_peft_model(model, use_gradient_checkpointing)
pass
return model, tokenizer
pass
pass
pass