You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I tried to run Whisper with torch-directml. It can run with either --fp16 flag or --use_dml_attn flag. However, I want to run with both flags. It failed with the following information:
The text was updated successfully, but these errors were encountered:
@XciciciX Thank you for reporting this issue. Could you please try with the latest Whisper sample? Multilingual models preserve past_key_value tensors when detecting language which causes the parameter to be incorrect when running transcription inference on audio. The sample has been updated to remove DML attention in AudioEncoder as there's no significant perf gains in the encoder
I tried to run Whisper with torch-directml. It can run with either --fp16 flag or --use_dml_attn flag. However, I want to run with both flags. It failed with the following information:
The text was updated successfully, but these errors were encountered: