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 #11298] Why would max token limit affect prompt eval time? #69511

New Issue
Closed
opened 2026-05-04 18:17:32 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-05-04 18:17:32 -05:00
Owner
Copy Link

Originally created by @negaralizadeh on GitHub (Jul 4, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/11298

I was running some experiments with codellama:7b-instruct-q4_K_M using the Python library (ollama.generate), and I noticed that changing the max generated tokens limit (num_predict) affects the prompt evaluation time(prompt_eval_duration), even though the input and all other options remain the same. In each round, I queried the model with 330 different queries. Here are the results:

Metric 1024 Output Tokens 1 Output Token
prompt_eval_count 190,760 190,760
prompt_eval_duration 122,118,342,000 84,885,213,000
output_tokens 74,891 330
eval_duration 2,107,317,930,000 4,357,000
total_duration 2,265,230,752,632 121,807,018,567

As you can see, there is a 30% drop in the total prompt evaluation time, around a 112ms drop ((122118342000 − 84885213000) / 330) per prompt. I'd appreciate some explanation for that.
The model fits 100% on my GPU (RTX3070) and my OS is Ubuntu 20.04.
Thank you!

Originally created by @negaralizadeh on GitHub (Jul 4, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/11298 I was running some experiments with codellama:7b-instruct-q4_K_M using the Python library (ollama.generate), and I noticed that changing the max generated tokens limit (num_predict) affects the prompt evaluation time(prompt_eval_duration), even though the input and all other options remain the same. In each round, I queried the model with 330 different queries. Here are the results: | Metric | 1024 Output Tokens | 1 Output Token | |:--------------------:|:-----------------:|:---------------:| | prompt_eval_count | 190,760 | 190,760 | | prompt_eval_duration | 122,118,342,000 | 84,885,213,000 | | output_tokens | 74,891 | 330 | | eval_duration | 2,107,317,930,000 | 4,357,000 | | total_duration | 2,265,230,752,632 | 121,807,018,567 | As you can see, there is a 30% drop in the total prompt evaluation time, around a 112ms drop ((122118342000 − 84885213000) / 330) per prompt. I'd appreciate some explanation for that. The model fits 100% on my GPU (RTX3070) and my OS is Ubuntu 20.04. Thank you!
GiteaMirror commented 2026-05-04 18:17:36 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jul 4, 2025):

Part of the process of prompt processing is finding and preparing a cache slot for inference. If a cache slot was previously used, it needs to have its corresponding KV cache entries cleared. The more tokens that were generated, the more time spent clearing the KV cache.

<!-- gh-comment-id:3036870960 --> @rick-github commented on GitHub (Jul 4, 2025): Part of the process of prompt processing is finding and preparing a cache slot for inference. If a cache slot was previously used, it needs to have its corresponding KV cache entries cleared. The more tokens that were generated, the more time spent clearing the KV cache.
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#69511
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?