[GH-ISSUE #78] Can't use the DeepSeekVL2 model with A6000 #61

Open
opened 2026-10-26 21:11:08 +00:00 by bsakaguchi · 4 comments
Owner

Originally created by @Baiqi-Li on GitHub (Feb 13, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/78

Image

Originally created by @Baiqi-Li on GitHub (Feb 13, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-VL2/issues/78 ![Image](https://github.com/user-attachments/assets/3473fbe7-6b21-47d7-81ac-84be94ad583c)
Author
Owner

@HubHop commented on GitHub (Feb 13, 2025):

Hi @Baiqi-Li, with the limited context, I guess the issue is due to SigLip-400M not supporting FlashAttention. Please disable it and try to rerun your demo.

<!-- gh-comment-id:2656058993 --> @HubHop commented on GitHub (Feb 13, 2025): Hi @Baiqi-Li, with the limited context, I guess the issue is due to SigLip-400M not supporting FlashAttention. Please disable it and try to rerun your demo.
Author
Owner

@Baiqi-Li commented on GitHub (Feb 15, 2025):

Thank you for your kind assistance!
Here is the complete error message, what should I do next?

Image

Image

<!-- gh-comment-id:2660753246 --> @Baiqi-Li commented on GitHub (Feb 15, 2025): Thank you for your kind assistance! Here is the complete error message, what should I do next? ![Image](https://github.com/user-attachments/assets/2ee5f00c-b52f-43c5-9c6b-68ffa9876852) ![Image](https://github.com/user-attachments/assets/71dc67c2-ccd6-4912-abf8-7b2e1ccbd9e7)
Author
Owner

@aysangh commented on GitHub (Apr 20, 2025):

i get the same error

<!-- gh-comment-id:2817321352 --> @aysangh commented on GitHub (Apr 20, 2025): i get the same error
Author
Owner

@aysangh commented on GitHub (Apr 24, 2025):

change all torch.bfloat16 to torch.float32

<!-- gh-comment-id:2828389529 --> @aysangh commented on GitHub (Apr 24, 2025): change all torch.bfloat16 to torch.float32
Sign in to join this conversation.
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-VL2#61
No description provided.