Hugging Face's logo

Spaces:
ACE-Step
/
Ace-Step-v1.5
Running on Zero

App Files Community
20

̶S̶i̶m̶p̶l̶y̶ ̶d̶o̶e̶s̶n̶'̶t̶ ̶w̶o̶r̶k̶ - UPDATE: fixed (`acestep_handler.py` & `acestep_llm_inference.py`)

#20
by Blursed - opened
Discussion
Blursed

Cover -> upload 16-bit .wav file -> Generate

Runs for ~320 seconds, then returns "Error".

Same behavior if the "Analyze" button is clicked.
Blursed changed discussion title from Simply doesn't work to Simply doesn't work - UPDATE: fixed (`acestep_handler.py` & `acestep_llm_inference.py`)
Blursed

Seems to be working now: https://huggingface.co/spaces/Blursed/Ace-Step-v1.5

Two files patched:
handler.py
llm_inference.py

  • get_best_attn_implementation() now queries torch.cuda.get_device_capability() and only selects flash-attn3 on Hopper (sm_9x) and flash_attention_2 on sm_8x/9x; everything else (including Blackwell sm_120) gets sdpa, which is PyTorch-native and works fine on sm_120 with the torch 2.9.1/cu128 build. If the capability can't be queried, it fails safe to sdpa.

  • Fixed the text-encoder loader in handler.py, which was unconditionally retrying flash_attention_2 in its fallback list regardless of hardware.

  • Added an escape hatch: set the ACESTEP_ATTN_IMPLEMENTATION env var (Space settings → Variables) to force a specific implementation if we ever want to override the auto-detection.
Blursed changed discussion title from Simply doesn't work - UPDATE: fixed (`acestep_handler.py` & `acestep_llm_inference.py`) to ~~Simply doesn't work~~ - UPDATE: fixed (`acestep_handler.py` & `acestep_llm_inference.py`)
Blursed changed discussion title from ~~Simply doesn't work~~ - UPDATE: fixed (`acestep_handler.py` & `acestep_llm_inference.py`) to ̶S̶i̶m̶p̶l̶y̶ ̶d̶o̶e̶s̶n̶'̶t̶ ̶w̶o̶r̶k̶ - UPDATE: fixed (`acestep_handler.py` & `acestep_llm_inference.py`)

Sign up or log in to comment