pip install torchtune , version is 0.61 which does not contrain qwen3

[whk6688]

when run tune run lora_finetune_single_device --config 4B_lora_single_device.yaml,
error is following:
Exception: Error converting the state dict. Found unexpected key: "model.layers.0.self_attn.k_norm.weight". Please make sure you're loading a checkpoint with the right format.

[whk6688]

【’pip install -e .】 also not work
error is following：
tune help
Traceback (most recent call last):
File "/home/wang/anaconda3/envs/python310/bin/tune", line 5, in
from torchtune._cli.tune import main
File "/home/wang/code/torchtune/torchtune/_cli/tune.py", line 11, in
from torchtune._cli.cp import Copy
File "/home/wang/code/torchtune/torchtune/_cli/cp.py", line 15, in
ROOT = Path(torchtune.file).parent.parent
File "/home/wang/anaconda3/envs/python310/lib/python3.10/pathlib.py", line 960, in new
self = cls._from_parts(args)
File "/home/wang/anaconda3/envs/python310/lib/python3.10/pathlib.py", line 594, in _from_parts
drv, root, parts = self._parse_args(args)
File "/home/wang/anaconda3/envs/python310/lib/python3.10/pathlib.py", line 578, in _parse_args
a = os.fspath(a)
TypeError: expected str, bytes or os.PathLike object, not NoneType

[whk6688]

the third way:pip install --pre --upgrade torchtune --extra-index-url https://download.pytorch.org/whl/nightly/cpu

error is following:
tune run lora_finetune_single_device --config 4B_lora_single_device.yaml
Traceback (most recent call last):
File "/home/wang/anaconda3/envs/python310/bin/tune", line 5, in
from torchtune._cli.tune import main
File "/home/wang/anaconda3/envs/python310/lib/python3.10/site-packages/torchtune/_cli/tune.py", line 12, in
from torchtune._cli.download import Download
File "/home/wang/anaconda3/envs/python310/lib/python3.10/site-packages/torchtune/_cli/download.py", line 26, in
from torchtune.training.checkpointing._utils import REPO_ID_FNAME
File "/home/wang/anaconda3/envs/python310/lib/python3.10/site-packages/torchtune/training/init.py", line 12, in
from torchtune.training._distributed import (
File "/home/wang/anaconda3/envs/python310/lib/python3.10/site-packages/torchtune/training/_distributed.py", line 30, in
from torch.distributed.fsdp import FSDPModule, ShardingStrategy
ImportError: cannot import name 'FSDPModule' from 'torch.distributed.fsdp' (/home/wang/anaconda3/envs/python310/lib/python3.10/site-packages/torch/distributed/fsdp/init.py)

[SalmanMohammadi]

Hi @whk6688. Thank you for raising this. Firstly, to answer the title of your issue, Qwen3 is currently planned for an 0.7.0 release and is only available through torchtune nightly builds or through an installation from source.

Could you please confirm whether you are running these commands in a fresh python virtual environment? I can confirm the following steps work for me - could you try them out?

> uv venv # you may also use pyenv or conda here
> source .venv/bin/activate
# make sure you install torch *first*
# if you're not using uv here you may omit 'uv' from the below command and just use pip install
# make sure you target the correct backend here too by changing cu126 to your appropriate CUDA version/etc.
> uv pip install --pre --upgrade torch torchvision torchao --index-url https://download.pytorch.org/whl/nightly/cu126 # full options are cpu/cu118/cu121/cu124/cu126/xpu
# install the latest nightly version of torchtune - you don't need to modify 'cpu' here,  but omit uv as above if needed
> uv pip install --pre --upgrade torchtune --extra-index-url https://download.pytorch.org/whl/nightly/cpu
> tune help
usage: tune [-h] {download,ls,cp,run,validate,cat} ...
tune: error: argument {download,ls,cp,run,validate,cat}: invalid choice: 'help' (choose from download, ls, cp, run, validate, cat)
> tune cp qwen2_5/7B_lora ./
Copied file to ./7B_lora.yaml
> tune validate ./7B_lora.yaml
Config is well-formed!

[whk6688]

yes, i create a fresh python virtural via conda.
then
pip install --pre --upgrade torch torchvision torchao --index-url https://download.pytorch.org/whl/nightly/cu121
pip install --pre --upgrade torchtune --extra-index-url https://download.pytorch.org/whl/nightly/cpu

the same error

tune help
Traceback (most recent call last):
File "/home/wang/anaconda3/envs/python3102/bin/tune", line 5, in
from torchtune._cli.tune import main
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/_cli/tune.py", line 12, in
from torchtune._cli.download import Download
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/_cli/download.py", line 26, in
from torchtune.training.checkpointing._utils import REPO_ID_FNAME
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/training/init.py", line 12, in
from torchtune.training._distributed import (
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/training/_distributed.py", line 30, in
from torch.distributed.fsdp import FSDPModule, ShardingStrategy
ImportError: cannot import name 'FSDPModule' from 'torch.distributed.fsdp' (/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torch/distributed/fsdp/init.py)

[SalmanMohammadi]

@whk6688 it seems as though torchtune installed correctly, but the installed version of torch is incompatibe with torchtune. Could you please provide the output of the following commands?

which tune
which pip
which python
pip list | grep torch

[whk6688]

which tune
/home/wang/anaconda3/envs/python3102/bin/tune
which pip
/home/wang/anaconda3/envs/python3102/bin/pip
which python
/home/wang/anaconda3/envs/python3102/bin/python

pip list | grep torch
pytorch-triton 3.1.0+cf34004b8a
torch 2.6.0.dev20241112+cu121
torchao 0.7.0.dev20241112+cu121
torchdata 0.12.0.dev20250508+cpu
torchtune 0.7.0.dev20250529+cpu
torchvision 0.20.0.dev20241112+cu121

[whk6688]

python version is 310.

[whk6688]

[SalmanMohammadi]

@whk6688 This may seem strange but could you try install using nightlies for cu118?

> uv pip install --pre --upgrade torch torchvision torchao --index-url https://download.pytorch.org/whl/nightly/cu118
> pip list | grep torch
> tune help

[whk6688]

ok. but the same error.
(python3102) wang@LEGION-REN9000K-34IRZ-4090-01:/code$ pip list | grep torch
pytorch-triton 3.3.1+gitc8757738
torch 2.8.0.dev20250601+cu118
torchao 0.12.0.dev20250602+cu118
torchdata 0.12.0.dev20250508+cpu
torchtune 0.7.0.dev20250529+cpu
torchvision 0.23.0.dev20250602+cu118
(python3102) wang@LEGION-REN9000K-34IRZ-4090-01:/code$ tune help
Traceback (most recent call last):
File "/home/wang/anaconda3/envs/python3102/bin/tune", line 5, in
from torchtune._cli.tune import main
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/_cli/tune.py", line 12, in
from torchtune._cli.download import Download
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/_cli/download.py", line 26, in
from torchtune.training.checkpointing._utils import REPO_ID_FNAME
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/training/init.py", line 29, in
from torchtune.training.grad_scaler import scale_grads, scale_grads
File "/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torchtune/training/_grad_scaler.py", line 13, in
from torch.nn.utils.clip_grad import _no_grad, _tensor_or_tensors
ImportError: cannot import name '_tensor_or_tensors' from 'torch.nn.utils.clip_grad' (/home/wang/anaconda3/envs/python3102/lib/python3.10/site-packages/torch/nn/utils/clip_grad.py)

[SalmanMohammadi]

Thank you for your patience here @whk6688. This looks like a separate torch nightly issue which should be fixed by #2778. Until then you should be able to safely install torch 2.7 (not nightlies) using pip install torch==2.7 torchvision torchao --index-url https://download.pytorch.org/whl/cu118 and fingers crossed all should work well.

[whk6688]

I appreciate your help! it works

[ebsmothers]

Thanks @SalmanMohammadi! I'm gonna close this but @whk6688 please reopen if you're still seeing issues