8000 Add flag for disabling flash attention by Maxusmusti · Pull Request #15 · instructlab/training · GitHub
[go: up one dir, main page]

Skip to content

Add flag for disabling flash attention - #15

Closed
Maxusmusti wants to merge 1 commit into
mainfrom
flash-attn-toggle
Closed

Maxusmusti wants to merge 1 commit into
mainfrom
flash-attn-toggle

Conversation

@Maxusmusti
Copy link
Copy Markdown
Collaborator

For hardware scenarios where flash attention is potentially not supported.

Signed-off-by: Mustafa Eyceoz <meyceoz@redhat.com>
@aldopareja
Copy link
Copy Markdown
Contributor

are we sure that not specifying the attention mechanism is enough for this not to fail?. We need to setup the 3090 machine from Mo, I guess we can spinup docker containers there just to not mess up with her drivers?

Comment thread main_ds.py
attn_implementation="flash_attention_2",
torch_dtype=torch.bfloat16,
**base_model_args,
use_padding_free_transformer=True,
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if you want to disable FA2, you also need to disable padding free, since it uses FA2 under the hood

@RobotSail RobotSail left a comment
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will need to test this but otherwise it looks good.

@ktam3 ktam3 linked an issue Jun 27, 2024 that may be closed by this pull request
@Maxusmusti Maxusmusti closed this Jul 3, 2024
@Maxusmusti
Maxusmusti deleted the flash-attn-toggle branch July 9, 2024 17:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

FlashAttention disable support

4 participants

0