Install correct version nexfort is hard

[xiongqisong]

📚 The doc issue

I can't easy install correct version nexfort after my environment need to restart.
I face some problem like:

• ModuleNotFoundError: No module named 'nexfort.inductor.codegen'
• nexfort compile env is not compatible with current run env

If i follow the install instructions in README, it will always install the newest version nexfort which can't run simply.

Suggest a potential alternative/fix

No response

[marigoold]

hi @xiongqisong
could you please show me how you installed nexfort and your environment (including torch version and cuda version)?

[gedasv]

hi @marigoold
I am having the same issue.

for all tests, python==3.11

onediff is installed through source, like so: git clone https://github.com/siliconflow/onediff.git && cd onediff && python3 -m pip install -e .
I have tried with various version combinations, including:
cuda==12.1
torch==2.4.0+cu121
nexfort==0.1.dev330 (tried a lower version too)
onediff==1.2.0 (I have also tried other versions: newest dev versions, and older versions)

and also tried downloading nexfort like this:
pip uninstall nexfort -y && pip --no-cache-dir install --pre nexfort -f https://nexfort-whl.oss-cn-beijing.aliyuncs.com/torch2.5.1/cu124/
with
cuda=12.4
torch==2.5.1+cu124
onediff==1.2.0

I have also tried to downgrade to
cuda==12.1
torch==2.2.0

but the error is the same -> ModuleNotFoundError: No module named 'nexfort.inductor.codegen' when it gets to the pipe = compile_pipe(pipe, ...) part.

Key points that may cause the issue:

• I am building the image with Cog
• I am applying compile_pipe() onto StableDiffusionXLInstantIDPipeline (tried with different parameters, none worked), here is the full code snippet:

self.instant_id_pipe = compile_pipe(
    self.instant_id_pipe, 
    backend="nexfort", 
    options={"mode": "max-autotune:cudagraphs", "dynamic": True},
    fuse_qkv_projections=True
)

PS Importing and using EulerDiscreteScheduler from onediff seems to work.
LMK if you need extra details.

[dimentary]

Having the same problem.

[chaochaobar]

Having the same problem.

[chaochaobar]

Sharing my solution here:
pip install nexfort==0.1.dev268 -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
downgrade the version of the package, it works

[dimentary]

Can this be fixed? nexfort==0.1.dev268 doesn't work in my cause because I need CUDA 12.4 support

[631068264]

Same as #1189

@dimentary you can try sth like that to match your torch version and cuda version
https://github.com/siliconflow/nexfort_releases/releases/expanded_assets/torch2.4.1_cu121