Sean Owen comments

Results 245 comments of


                                            Sean Owen

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name'

It's saying the directory doesn't contain what it expects. It should look something like https://huggingface.co/databricks/dolly-v2-12b/tree/main

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name'

Correct, you use /dbfs not dbfs: with tools that only know local files. But I am not clear what you are saying the root cause was

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name'

Not sure, you're not showing where that error arises

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name'

I'm not sure what you mean; that's a command line argument but you seem to be referring to code.

HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name'

Er, you're just typing those words on the command line! `--input-model` is a flag you pass to deepspeed

Added datasets in 5 languages

Oh that's neat. I'm not sure whether the other project folks want this hosted here vs just letting people host their own derived datasets. It's a great point though, you...

Added datasets in 5 languages

You can put them on HF too! I'm seeing things like https://huggingface.co/datasets/kunishou/databricks-dolly-15k-ja as well

OSError: databricks/dolly-v2-3b does not appear to have a file named pytorch_model.bin, tf_model.h5, model.ckpt or flax_model.msgpack.

Hm, that works for me. Clear your HF cache and run again or something? or add "model=" before the model name to be sure it's interpreting that correctly

RuntimeError: CUDA error: CUBLAS_STATUS_INVALID_VALUE when calling

I've seen cublas errors when: - It's not installed - It doesn't match the CUDA driver version - ... just out of memory in a weird way Check that first?

Running the train_dolly.py with `transformers[torch]==4.28.1`, `deepspeed==0.9.1` and V100 GPU gives error in defragment

Hm, that's strange. I wonder if 0.9.1 fixed it? I didn't observe this on A10s, FWIW. Offload is an interesting workaround, though shouldn't be needed with 32GB, yes.