Duplicated Space Not Working As Original

[Ratnesh-dev]

I duplicated this space: Parakeet TDT 0.6b V3 - a Hugging Face Space by pratyushmittal
to here: Parakeet TDT 0.6b V3 - a Hugging Face Space by Ratnesh-dev

Which gave me a build error in the requirements files (it was using the #egg= syntax).
Old line in requirements:
git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]
Updated to:
nemo_toolkit[asr] @ git+https://github.com/NVIDIA/NeMo.git@main
(You can see the commit history on the duplicated space)

Now, the space builds and is in a running state, but I can’t seem to get any transcription output out of it. On the original space, I can upload any file (5 min, 15 min, 30 min) and get the output within 20 seconds. But on my duplicated space, I get an error even on a small 5 minute file.

I don’t see any specific errors in the container logs.
What is going on here? Both spaces are using the same ZeroGPU hardware. One works, the other doesn’t.
Can you please help me with this?

[John6666]

After cloning the repository and pinning the dependencies to older versions, transcription now completes successfully. So I’ve sent PR.

The cause was likely that PyTorch on Zero GPU defaulted to version 2.9, combined with the fact that PyTorch 2.9 and the new NeMo don’t seem to play well together.
Also, I’m still not quite sure how to stabilize this Space with the new PyTorch and NeMo…

PR

• Ratnesh-dev/parakeet-tdt-0.6b-v3 · Fix version drift

Related issue

1. [NeMo 2.6] RNNT ASR inference fails on A100 (CUDA 12.8, PyTorch 2.9) with CUDA Graphs error CUDA failure! 35 · Issue #15145 · NVIDIA-NeMo/NeMo · GitHub
2. TDT Inference uses a lot of CPU and slow down with `use_cuda_graph_decoder: True` · Issue #15164 · NVIDIA-NeMo/NeMo · GitHub

Edit:
You maybe need to do Factory rebuild to fix the Space after merged.

[Ratnesh-dev]

Tried with a 30 min audio after merging your fix.
Works like a charm!

Thank you so much. You really saved me from this 2-day long headache.