Looks like the updated code requires trl>=0.9.3, which isn't compatiable with unsloth

[YifanDengWHU]

Reminder

• I have read the README and searched the existing issues.

System Info

• llamafactory version: 0.7.2.dev0
• Platform: Linux-4.18.0-517.el8.x86_64-x86_64-with-glibc2.28
• Python version: 3.9.0
• PyTorch version: 2.0.0+cu117 (GPU)
• Transformers version: 4.41.2
• Datasets version: 2.19.2
• Accelerate version: 0.30.1
• PEFT version: 0.11.1
• TRL version: 0.9.4
• GPU type: NVIDIA L40
• Bitsandbytes version: 0.43.1

Reproduction

git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e .[metrics]

pip install "unsloth[cu118-torch230] @ git+https://github.com/unslothai/unsloth.git"

Yesterday everthing worked well, but today I got error:

Traceback (most recent call last):
File "/var/lib/condor/execute/slot1/dir_677837/mm_molecule/bin/llamafactory-cli", line 5, in
from llamafactory.cli import main
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/init.py", line 3, in
from .cli import VERSION
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/cli.py", line 7, in
from . import launcher
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/launcher.py", line 1, in
from llamafactory.train.tuner import run_exp
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/train/tuner.py", line 9, in
from ..hparams import get_infer_args, get_train_args
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/hparams/init.py", line 6, in
from .parser import get_eval_args, get_infer_args, get_train_args
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/hparams/parser.py", line 27, in
check_dependencies()
File "/var/lib/condor/execute/slot1/dir_677837/LLaMA-Factory/src/llamafactory/extras/misc.py", line 68, in check_dependencies
require_version("trl>=0.9.3", "To fix: pip install trl>=0.9.3")
File "/var/lib/condor/execute/slot1/dir_677837/mm_molecule/lib/python3.9/site-packages/transformers/utils/versions.py", line 111, in require_version
_compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
File "/var/lib/condor/execute/slot1/dir_677837/mm_molecule/lib/python3.9/site-packages/transformers/utils/versions.py", line 44, in _compare_versions
raise ImportError(
ImportError: trl>=0.9.3 is required for a normal functioning of this module, but found trl==0.8.6.
To fix: pip install trl>=0.9.3

Even though I can install trl==0.9.4, the unsloth will override it again:

Collecting trl<0.9.0,>=0.7.9 (from unsloth[cu118-torch230]@ git+https://github.com/unslothai/unsloth.git)
Downloading trl-0.8.6-py3-none-any.whl.metadata (11 kB)

After I delete the pip install "unsloth[cu118-torch230] @ git+https://github.com/unslothai/unsloth.git", it works fine. But it also means I can't use unsloth for training.

Expected behavior

It seems that updated LLAMA-Factory code add the constraint for trl version>=0.9.3. Is it necessary?

Others

No response

[hiyouga]

fixed