Spaces:

---

HugMinjun

/

LLPSense

like 1

Running on Zero

Hello,

The precomputed examples work normally, but live ProtT5 feature extraction always returns “GPU task aborted”.

I tested both:

• a 420-aa protein sequence
• a shortened 100-aa standard amino-acid sequence

Both produced the same error.

I checked the current code and noticed that app.py uses:

@spaces.GPU
def extract_t5_feature(sequence: str) -> np.ndarray:
return _extract_t5_feature_core(sequence)

Then _extract_t5_feature_core() calls load_t5() in t5_utils.py. On the first request, load_t5() runs:

T5EncoderModel.from_pretrained(...)
.to(dev)

Therefore, the ProtT5 model appears to be loaded and transferred to the GPU during the first GPU-decorated request. Since @spaces.GPU does not specify a duration, the cold-loading step may exceed the default runtime and cause the task to be aborted.

Could you please check whether the ProtT5 model can be preloaded, or increase the GPU duration, for example:

@spaces.GPU(duration=180)

Thank you for providing LLPSense.