Spaces:

---

google

/

TIPSv2

like 25

Paused

Fixes two requirements.txt bugs that currently break the Space:

1. ZeroGPU AttributeError (worker error)
--extra-index-url https://download.pytorch.org/whl/cpu installs a CPU-only torch, overwriting the CUDA build the ZeroGPU base image ships. ZeroGPU's spaces runtime calls CUDA-specific torch APIs that don't exist in the CPU build → worker dies with AttributeError.

2. Tokenizer AttributeError on every inference
Unpinned sentencepiece resolves to 0.2.2, which removed SentencePieceProcessor.SetEncodeExtraOptions. text_encoder.py calls it during tokenizer init, so model loading crashes:

AttributeError: 'SentencePieceProcessor' object has no attribute 'SetEncodeExtraOptions'

Fix: drop the CPU index and pin sentencepiece==0.2.0 (still exposes the method). Verified end-to-end on a duplicate ZeroGPU Space — all tabs (PCA, depth/normals, supervised + zero-shot segmentation) run.