Downloading: 100%|██████████| 1.99G/1.99G [00:18<00:00, 113MiB/s]zstd: /*stdin*\: unsupported format tar: Child returned status 1tar: Error is not recoverable: exiting nowTraceback (most recent call last): File "/api/download.py", line 62, in <module> download_model() File "/api/download.py", line 27, in download_model subprocess.run( File "/opt/conda/envs/xformers/lib/python3.9/subprocess.py", line 528, in run raise CalledProcessError(retcode, process.args,subprocess.CalledProcessError: Command '['tar', '--use-compress-program=unzstd', '-C', '_arimiu', '-xvf', '_arimiu.ckpt']' returned non-zero exit status 2.ERROR conda.cli.main_run:execute(47): conda run /bin/bash -c if [ “$RUNTIME_DOWNLOADS” = “0” ] ; then python3 download.py ; fi failed. (See above for error)error building image: error building stage: failed to execute command: waiting for process to exit: exit status 1
Thanks for the report… can you provide all the call/modelinputs (for runtime download setup) or all the build vars (for serverless build-downloads) so I can try reproduce? You can mask out the full url if the model is private.
At first glance though, it looks like the .ckpt is coming in through the code that’s expecting an optimized .tar.zst of a diffusers model. Is it possible you specified the checkpoint url with MODEL_URL instead of CHECKPOINT_URL? If not, definitely a bug and I’ll look into it as soon as I have the full inputs.
I’m kind of tempted to start accepting .ckpt files even via MODEL_URL, for ease-of-use, but hesitate as it’s not necessarily a 1:1 conversion… and can perform badly if converted with the wrong config. But might be worth doing so regardless… so thanks for the idea
