github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-08 00:51:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #10680] tried the new 0.7 and my t/s speed dropped from 2.9 to 0.5 on a 24b model #7019

New Issue
Closed
opened 2026-04-12 18:54:54 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-12 18:54:54 -05:00
Owner
Copy Link

Originally created by @AncientMystic on GitHub (May 13, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10680

What is the issue?

i just tried the new ollama 0.7 and i had been playing around with a 24B model that is 14GB and loads 75% cpu/25% GPU on a install of ubuntu with 4GB vram via vGPU profile.

i was seeing 2.5-2.9+t/s on 0.6.8 but i tried 0.7 and it dropped to a max of 0.5 (tried generating 10 responses) where it is generating, very, very slowly, i went back to 0.6.8 and i am back to 2.9t/s, so for some reason ollama is 5-6x slower in the latest pre-release.

hardware:
CPU: i7-7820X
GPU: Tesla P4 8GB using 4GB vGPU profile
Driver version: 550.144
cuda version: 12.4
compute version: 6.1

OS

Linux

GPU

Nvidia

CPU

Intel

Originally created by @AncientMystic on GitHub (May 13, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10680 ### What is the issue? i just tried the new ollama 0.7 and i had been playing around with a 24B model that is 14GB and loads 75% cpu/25% GPU on a install of ubuntu with 4GB vram via vGPU profile. i was seeing 2.5-2.9+t/s on 0.6.8 but i tried 0.7 and it dropped to a max of 0.5 (tried generating 10 responses) where it is generating, very, very slowly, i went back to 0.6.8 and i am back to 2.9t/s, so for some reason ollama is 5-6x slower in the latest pre-release. hardware: CPU: i7-7820X GPU: Tesla P4 8GB using 4GB vGPU profile Driver version: 550.144 cuda version: 12.4 compute version: 6.1 ### OS Linux ### GPU Nvidia ### CPU Intel
GiteaMirror added the bug label 2026-04-12 18:54:54 -05:00
GiteaMirror commented 2026-04-12 18:54:58 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (May 13, 2025):

Server logs from both versions may aid in debugging.

<!-- gh-comment-id:2875322827 --> @rick-github commented on GitHub (May 13, 2025): [Server logs]( https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) from both versions may aid in debugging.
GiteaMirror commented 2026-04-12 18:54:59 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (May 13, 2025):

Can confirm that models that spill to system RAM take a significant performance hit. Affects both old and new engines.

$ for m in qwen2.5:14b-instruct-q4_K_M phi4-reasoning:14b-q4_K_M gemma3:12b-it-q4_K_M qwen3:14b-q4_K_M ; do for v in 0.6.8 0.7.0-rc0 ; do OLLAMA_DEBUG=1 OLLAMA_DOCKER_TAG=$v OLLAMA_KEEP_ALIVE=-1 OLLAMA_NUM_PARALLEL=1 docker compose up -d ollama 2>&- >&- ; sleep 2 ; for c in 4096 65536 ; do echo $v $m $c ; curl -s localhost:11434/api/generate -d '{"model":"'$m'","prompt":"hello","stream":false,"options":{"num_ctx":'$c'}}' | jq -r '" prompt tps: \(.prompt_eval_count/(.prompt_eval_duration/1000000000))"," eval tps  : \(.eval_count/(.eval_duration/1000000000))"' ; ollama ps | grep $m ; echo ; done ; done ; done
0.6.8 qwen2.5:14b-instruct-q4_K_M 4096
 prompt tps: 228.23505444398873
 eval tps  : 40.94157748537369
qwen2.5:14b-instruct-q4_K_M    7cdf5a0187d5    10 GB    100% GPU     Forever    

0.6.8 qwen2.5:14b-instruct-q4_K_M 65536
 prompt tps: 22.840987952794233
 eval tps  : 7.886128381153751
qwen2.5:14b-instruct-q4_K_M    7cdf5a0187d5    29 GB    44%/56% CPU/GPU    Forever    

0.7.0-rc0 qwen2.5:14b-instruct-q4_K_M 4096
 prompt tps: 220.25571394714981
 eval tps  : 40.93492084414357
qwen2.5:14b-instruct-q4_K_M    7cdf5a0187d5    10 GB    100% GPU     Forever    

0.7.0-rc0 qwen2.5:14b-instruct-q4_K_M 65536
 prompt tps: 1.7845719102765394
 eval tps  : 1.715184921837344
qwen2.5:14b-instruct-q4_K_M    7cdf5a0187d5    29 GB    44%/56% CPU/GPU    Forever    

...
model engine 0.6.8-gpu 0.7.0-gpu 0.6.8-hybrid 0.7.0-hybrid 0.6.8-cpu 0.7.0-cpu
qwen2.5:14b-instruct-q4_K_M old 40.9 40.9 7.8 1.7 6.5 1.1
phi4-reasoning:14b-q4_K_M old 33.1 33.1 5.1 1.1 5.2 0.9
qwen3:14b-q4_K_M old 40.8 40.8 6.6 1.3 6.1 0.9
gemma3:12b-it-q4_K_M new 45.8 45.6 25.7 10.5 7.6 1.1
gemma2:9b-instruct-q4_K_M new 59.1 58.0 11.5 2.5 9.7 1.6
<!-- gh-comment-id:2876271530 --> @rick-github commented on GitHub (May 13, 2025): Can confirm that models that spill to system RAM take a significant performance hit. Affects both old and new engines. ```console $ for m in qwen2.5:14b-instruct-q4_K_M phi4-reasoning:14b-q4_K_M gemma3:12b-it-q4_K_M qwen3:14b-q4_K_M ; do for v in 0.6.8 0.7.0-rc0 ; do OLLAMA_DEBUG=1 OLLAMA_DOCKER_TAG=$v OLLAMA_KEEP_ALIVE=-1 OLLAMA_NUM_PARALLEL=1 docker compose up -d ollama 2>&- >&- ; sleep 2 ; for c in 4096 65536 ; do echo $v $m $c ; curl -s localhost:11434/api/generate -d '{"model":"'$m'","prompt":"hello","stream":false,"options":{"num_ctx":'$c'}}' | jq -r '" prompt tps: \(.prompt_eval_count/(.prompt_eval_duration/1000000000))"," eval tps : \(.eval_count/(.eval_duration/1000000000))"' ; ollama ps | grep $m ; echo ; done ; done ; done 0.6.8 qwen2.5:14b-instruct-q4_K_M 4096 prompt tps: 228.23505444398873 eval tps : 40.94157748537369 qwen2.5:14b-instruct-q4_K_M 7cdf5a0187d5 10 GB 100% GPU Forever 0.6.8 qwen2.5:14b-instruct-q4_K_M 65536 prompt tps: 22.840987952794233 eval tps : 7.886128381153751 qwen2.5:14b-instruct-q4_K_M 7cdf5a0187d5 29 GB 44%/56% CPU/GPU Forever 0.7.0-rc0 qwen2.5:14b-instruct-q4_K_M 4096 prompt tps: 220.25571394714981 eval tps : 40.93492084414357 qwen2.5:14b-instruct-q4_K_M 7cdf5a0187d5 10 GB 100% GPU Forever 0.7.0-rc0 qwen2.5:14b-instruct-q4_K_M 65536 prompt tps: 1.7845719102765394 eval tps : 1.715184921837344 qwen2.5:14b-instruct-q4_K_M 7cdf5a0187d5 29 GB 44%/56% CPU/GPU Forever ... ``` | model | engine | 0.6.8-gpu | 0.7.0-gpu | 0.6.8-hybrid | 0.7.0-hybrid | 0.6.8-cpu | 0.7.0-cpu | -- | -- | -- | -- | -- | -- | -- | -- | | qwen2.5:14b-instruct-q4_K_M | old | 40.9 | 40.9 | 7.8 | 1.7 | 6.5 | 1.1 | | phi4-reasoning:14b-q4_K_M | old | 33.1 | 33.1 | 5.1 | 1.1 | 5.2 | 0.9 | | qwen3:14b-q4_K_M | old | 40.8 | 40.8 | 6.6 | 1.3 | 6.1 | 0.9 | | gemma3:12b-it-q4_K_M | new | 45.8 | 45.6 | 25.7 | 10.5 | 7.6 | 1.1 | | gemma2:9b-instruct-q4_K_M | new | 59.1 | 58.0 | 11.5 | 2.5 | 9.7 | 1.6 |
GiteaMirror commented 2026-04-12 18:55:01 -05:00
Author
Owner
Copy Link

@AncientMystic commented on GitHub (May 13, 2025):

thank you for adding all of this, this is very well summarized,

<!-- gh-comment-id:2877645353 --> @AncientMystic commented on GitHub (May 13, 2025): thank you for adding all of this, this is very well summarized,
GiteaMirror commented 2026-04-12 18:55:02 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (May 14, 2025):

Looks like this is fixed in 0.7.0-rc1.

model 0.6.8-cpu 0.7.0-rc0-cpu 0.7.0-rc1-cpu
qwen2.5:14b-instruct-q4_K_M 7.1 1.1 6.5
phi4-reasoning:14b-q4_K_M 5.1 0.9 5.1
qwen3:14b-q4_K_M 6.2 0.9 6.1
gemma3:12b-it-q4_K_M 7.6 1.1 7.6
gemma2:9b-instruct-q4_K_M 9.4 1.6 9.6
<!-- gh-comment-id:2880321366 --> @rick-github commented on GitHub (May 14, 2025): Looks like this is fixed in 0.7.0-rc1. | model | 0.6.8-cpu | 0.7.0-rc0-cpu | 0.7.0-rc1-cpu | | -- | -- | -- | -- | | qwen2.5:14b-instruct-q4_K_M | 7.1| 1.1 | 6.5 | | phi4-reasoning:14b-q4_K_M | 5.1 | 0.9 | 5.1 | | qwen3:14b-q4_K_M | 6.2 | 0.9 | 6.1 | | gemma3:12b-it-q4_K_M | 7.6 | 1.1 | 7.6 | | gemma2:9b-instruct-q4_K_M | 9.4 | 1.6 | 9.6 |
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 bug
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#7019
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?