github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-05 23:53:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15881] Flash Attention gating for deepseek2 appears controlled by GGUF metadata #72179

New Issue
Open
opened 2026-05-05 03:35:52 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-05-05 03:35:52 -05:00
Owner
Copy Link

Originally created by @gotnochill815-web on GitHub (Apr 29, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15881

It looks like Flash Attention support for GLM4.7-flash (unsloth version) is not enabled, even though similar models (e.g. using glmmoelite arch) do support it.

From investigation in #15855:

  • Renaming the architecture does not enable FA
  • The difference appears to come from GGUF metadata rather than architecture
  • Models with deepseek2 arch do not trigger FA

This suggests FA eligibility is controlled by specific GGUF metadata fields.

Questions:

  • Which GGUF metadata keys determine whether Flash Attention is enabled?
  • Are there known constraints (attention type, head dims, KV layout, etc.) that block FA for deepseek2?

If this is just a metadata mismatch, aligning those fields might allow FA support without changing model weights.

Originally created by @gotnochill815-web on GitHub (Apr 29, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15881 It looks like Flash Attention support for GLM4.7-flash (unsloth version) is not enabled, even though similar models (e.g. using `glmmoelite` arch) do support it. From investigation in #15855: - Renaming the architecture does not enable FA - The difference appears to come from GGUF metadata rather than architecture - Models with `deepseek2` arch do not trigger FA This suggests FA eligibility is controlled by specific GGUF metadata fields. Questions: - Which GGUF metadata keys determine whether Flash Attention is enabled? - Are there known constraints (attention type, head dims, KV layout, etc.) that block FA for `deepseek2`? If this is just a metadata mismatch, aligning those fields might allow FA support without changing model weights.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#72179
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?