github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #1007] Context Shifting To Increase Speed Dramatically #490

New Issue
Closed
opened 2026-04-12 10:10:12 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-12 10:10:12 -05:00
Owner
Copy Link

Originally created by @chigkim on GitHub (Nov 5, 2023).
Original GitHub issue: https://github.com/ollama/ollama/issues/1007

koboldcpp v1.48.1 now has implemented this feature, and apparently it increased the speed dramatically because it doesn't have to reprocess previous context to generate a new response.

Could you look into implement this feature to Ollama as well?

https://github.com/LostRuins/koboldcpp/releases/tag/v1.48.1

"Context Shifting (A.K.A. EvenSmarterContext) - This feature utilizes KV cache shifting to automatically remove old tokens from context and add new ones without requiring any reprocessing. So long as you use no memory/fixed memory and don't use world info, you should be able to avoid almost all reprocessing between consecutive generations even at max context. This does not consume any additional context space, making it superior to SmartContext."

Thanks!

Originally created by @chigkim on GitHub (Nov 5, 2023). Original GitHub issue: https://github.com/ollama/ollama/issues/1007 koboldcpp v1.48.1 now has implemented this feature, and apparently it increased the speed dramatically because it doesn't have to reprocess previous context to generate a new response. Could you look into implement this feature to Ollama as well? https://github.com/LostRuins/koboldcpp/releases/tag/v1.48.1 "Context Shifting (A.K.A. EvenSmarterContext) - This feature utilizes KV cache shifting to automatically remove old tokens from context and add new ones without requiring any reprocessing. So long as you use no memory/fixed memory and don't use world info, you should be able to avoid almost all reprocessing between consecutive generations even at max context. This does not consume any additional context space, making it superior to SmartContext." Thanks!
GiteaMirror added the feature request label 2026-04-12 10:10:12 -05:00
GiteaMirror commented 2026-04-12 10:10:13 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (May 6, 2024):

Hi there, thank you so much for the issue. At this point we've seen quite a few quality concerns with context shifting and will be focused on helping understand context utilization (e.g. API will return a field when the context limit is hit) vs shifting the context.

We may re-introduce context shifting later on once we can do so between bos/eos tokens but currently context shifting happens at arbitrary points in the context which causes infinite and/or bad generations

<!-- gh-comment-id:2097088114 --> @jmorganca commented on GitHub (May 6, 2024): Hi there, thank you so much for the issue. At this point we've seen quite a few quality concerns with context shifting and will be focused on helping understand context utilization (e.g. API will return a field when the context limit is hit) vs shifting the context. We may re-introduce context shifting later on once we can do so between bos/eos tokens but currently context shifting happens at arbitrary points in the context which causes infinite and/or bad generations
GiteaMirror commented 2026-04-12 10:10:14 -05:00
Author
Owner
Copy Link

@chigkim commented on GitHub (May 7, 2024):

Could you expose tokenize and detokenize feature of llama.cpp api? Then we can know the token count before feeding text for generation. It seems there are a lot of interest in this.

<!-- gh-comment-id:2097133747 --> @chigkim commented on GitHub (May 7, 2024): Could you expose tokenize and detokenize feature of llama.cpp api? Then we can know the token count before feeding text for generation. It seems there are a lot of interest in this.
GiteaMirror commented 2026-04-12 10:10:14 -05:00
Author
Owner
Copy Link

@chigkim commented on GitHub (May 7, 2024):

https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md

https://github.com/ollama/ollama/issues/4186

https://github.com/ollama/ollama/issues/1716

https://github.com/ollama/ollama/issues/3582

https://github.com/ollama/ollama/issues/1345

<!-- gh-comment-id:2097135703 --> @chigkim commented on GitHub (May 7, 2024): https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md https://github.com/ollama/ollama/issues/4186 https://github.com/ollama/ollama/issues/1716 https://github.com/ollama/ollama/issues/3582 https://github.com/ollama/ollama/issues/1345
GiteaMirror commented 2026-04-12 10:10:15 -05:00
Author
Owner
Copy Link

@joewinke commented on GitHub (May 22, 2024):

Could you expose tokenize and detokenize feature of llama.cpp api? Then we can know the token count before feeding text for generation. It seems there are a lot of interest in this.

This would allow for UX that shows % of context window used, which would be helpful in many contexts.

<!-- gh-comment-id:2125546786 --> @joewinke commented on GitHub (May 22, 2024): > Could you expose tokenize and detokenize feature of llama.cpp api? Then we can know the token count before feeding text for generation. It seems there are a lot of interest in this. This would allow for UX that shows % of context window used, which would be helpful in many contexts.
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 feature request
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#490
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?