[GH-ISSUE #1385] [Feature] Flash Attention Alternative for Older GPU's #928

New issue

Open

opened 2026-10-29 14:46:17 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-29 14:46:17 +00:00

Owner

Originally created by @apcameron on GitHub (Jun 12, 2025).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1385

Checklist

1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/kvcache-ai/ktransformers/discussions. Otherwise, it will be closed.
2. To help the community, I will use Chinese/English or attach an Chinese/English translation if using another language. Non-English/Chinese content without translation may be closed.

Motivation

Many older GPU's have enough VRAM but do not support flash attention
EG the Tesla P40

Please consider enabling the Use of vanilla attention as an alternative for those with older GPU's

See this example for another product https://github.com/HiDream-ai/HiDream-I1/pull/27

Originally created by @apcameron on GitHub (Jun 12, 2025). Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1385 ### Checklist - [ ] 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/kvcache-ai/ktransformers/discussions. Otherwise, it will be closed. - [ ] 2. To help the community, I will use Chinese/English or attach an Chinese/English translation if using another language. Non-English/Chinese content without translation may be closed. ### Motivation Many older GPU's have enough VRAM but do not support flash attention EG the Tesla P40 ### Related resources Please consider enabling the Use of vanilla attention as an alternative for those with older GPU's See this example for another product https://github.com/HiDream-ai/HiDream-I1/pull/27

bsakaguchi commented

2026-10-29 14:46:18 +00:00

Author

Owner

@Atream commented on GitHub (Jun 15, 2025):

Thanks for your suggestion, we will implement it as soon as possible.

@Atream commented on GitHub (Jun 15, 2025): Thanks for your suggestion, we will implement it as soon as possible.