[Docs] Two non-fatal warnings on every model load + small setup-friction note

Three small documentation items from running the standard install + inference path. All non-blocking — model works correctly — but each adds a bit of noise / friction for first-time users.

1. preprocessor_config.json not found warning on every model load

Loading microsoft/VibeVoice-ASR from HuggingFace via the standard demo script emits:

Could not load preprocessor_config.json: microsoft/VibeVoice-ASR does not appear to have
a file named preprocessor_config.json. Checkout 'https://huggingface.co/microsoft/VibeVoice-ASR/tree/main'
for available files.
Using default configuration

The default fallback works fine, but the warning fires on every load.

Suggestion: either ship preprocessor_config.json in the HF repo, or add a note to the model card explaining the warning is expected behavior.

2. Tokenizer class mismatch warning

Same load path emits:

The tokenizer class you load from this checkpoint is not the same type as the class
this function is called from. It may result in unexpected tokenization.

Tokenization output looks correct, but again — noise on every load.

Suggestion: align the tokenizer-config so the warning doesn't fire, or note in the model card that this is expected.

3. Setup friction outside NVIDIA Deep Learning Container

The README recommends nvcr.io/nvidia/pytorch:25.12-py3 as the base. Running on alternatives (we used runpod/pytorch:1.0.3-cu1281-torch291-ubuntu2404, comparable PyTorch base) hit two install issues:

• Ubuntu 24.04 PEP 668 blocks pip install -e . against system Python without --break-system-packages or a venv
• Debian's cryptography package conflicts with the version pip install -e . wants to upgrade to (no RECORD file → can't uninstall via pip)

Workaround: install into a clean Python venv inside the repo dir:

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

This works cleanly. Suggesting one-line addition to the install docs:

If using a non-NVIDIA Deep Learning Container base image, install in a clean Python venv (python3 -m venv .venv && source .venv/bin/activate && pip install -e .) to avoid system-Python conflicts on Ubuntu 24.04 / Debian-based images.

Setup details for context

• runpod/pytorch:1.0.3-cu1281-torch291-ubuntu2404
• Python 3.12, transformers 4.57.6, torch 2.11.0+cu130
• RTX 4090 / cuda 13

None of these are blocking — model works correctly — but addressing them would smooth the first-run experience for users adopting outside the recommended container.

Thanks for the open-source release.

Start by reviewing the README installation guidance and the microsoft/VibeVoice-ASR model card, then reproduce the standard demo load to confirm the two warnings and the non-NVIDIA setup friction. Document the clean venv path and clarify the expected warning behavior, or determine whether the referenced model configuration can be aligned; done means the first-run guidance covers all three reported items.