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 #6654] Multi-instance seems not working #50701

New Issue
Closed
opened 2026-04-28 16:48:35 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-28 16:48:35 -05:00
Owner
Copy Link

Originally created by @bigsausage on GitHub (Sep 5, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/6654

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

i want to use multi-process to increase the concurrency of my server

i use the follow command to start the server first
CUDA_VISIBLE_DEVICES=3 OLLAMA_NUM_PARALLEL=3 OLLAMA_MAX_LOADED_MODELS=3 /usr/bin/ollama serve

then i created 3 different copy of the model
ollama create my_lama3_1 -f ./Modelfile1
ollama create my_lama3_2 -f ./Modelfile2
ollama create my_lama3_2 -f ./Modelfile2

and the Modelfile1 Modelfile2 Modelfile3 points to the llama3_quantize_1.gguf llama3_quantize_2.gguf` llama3_quantize_3.gguf` which actually is the same int4 gguf (about 5GB).

ollama list shows that i got 3 different instance model but got the same id
image

nvidia-smi shows that the gpu only uses 6g
image

ollama ps shows that i got only one instance running.
image

i wrote a script to randomly distribute the request to one of the three models(my_lama3_1 / my_lama3_2 / my_lama3_3 ) to test the concurrency ability of my server , but the gpu still keeps at the range of 6GB, which i expect to be 6GB * 3 = 18GB to release all its ability ..

is there any way to get the 3 instance load into the gpu so i can get a better qps of my server?
or i was not using the command correctly and please tell me the correct way...
all i want is to improve the qps of my server, any advice is welcome!!

thanks ~

OS

Linux

GPU

Intel

CPU

Intel

Ollama version

0.2.5

Originally created by @bigsausage on GitHub (Sep 5, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/6654 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? i want to use multi-process to increase the concurrency of my server i use the follow command to start the server first `CUDA_VISIBLE_DEVICES=3 OLLAMA_NUM_PARALLEL=3 OLLAMA_MAX_LOADED_MODELS=3 /usr/bin/ollama serve` then i created 3 different copy of the model `ollama create my_lama3_1 -f ./Modelfile1` `ollama create my_lama3_2 -f ./Modelfile2` `ollama create my_lama3_2 -f ./Modelfile2` and the `Modelfile1` `Modelfile2` `Modelfile3` points to the `llama3_quantize_1.gguf` ``llama3_quantize_2.gguf` ``llama3_quantize_3.gguf` which actually is the same int4 gguf (about 5GB). `ollama list` shows that i got 3 different instance model but got the same id ![image](https://github.com/user-attachments/assets/ff66289d-cfd0-4b21-9d7e-faa7dabd20ee) `nvidia-smi` shows that the gpu only uses 6g ![image](https://github.com/user-attachments/assets/47bfad3e-e140-4d84-98bb-cbe76e5bcf16) `ollama ps` shows that i got only one instance running. ![image](https://github.com/user-attachments/assets/fb1c0420-19a8-4982-ba2a-0fa521d88b6a) i wrote a script to randomly distribute the request to one of the three models(my_lama3_1 / my_lama3_2 / my_lama3_3 ) to test the concurrency ability of my server , but the gpu still keeps at the range of 6GB, which i expect to be 6GB * 3 = 18GB to release all its ability .. is there any way to get the 3 instance load into the gpu so i can get a better qps of my server? or i was not using the command correctly and please tell me the correct way... all i want is to improve the qps of my server, any advice is welcome!! thanks ~ ### OS Linux ### GPU Intel ### CPU Intel ### Ollama version 0.2.5
GiteaMirror added the feature request label 2026-04-28 16:48:35 -05:00
GiteaMirror commented 2026-04-28 16:48:37 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 5, 2024):

What you have is the same model with three names, so ollama just loads it once. ollama doesn't currently support loading the same model more than once. I say currently because I'm sure I read something some time ago that the developers were considering this, but I can't find a reference so consider that a possible hallucination.

If you want to run multiple copies of the same model, the easiest way at the moment is to start an ollama server on a different port for each GPU, using CUDA_VISIBLE_DEVICES to bind each GPU to it's server, and then use a reverse proxy (eg nginx or ollama_proxy_server) to distribute the requests to the ollama servers.

Another way would be to use slightly different quants for your custom models, eg q4_0, q4_1, q4_K_M.

<!-- gh-comment-id:2331521028 --> @rick-github commented on GitHub (Sep 5, 2024): What you have is the same model with three names, so ollama just loads it once. ollama doesn't currently support loading the same model more than once. I say currently because I'm sure I read something some time ago that the developers were considering this, but I can't find a reference so consider that a possible hallucination. If you want to run multiple copies of the same model, the easiest way at the moment is to start an ollama server on a different port for each GPU, using `CUDA_VISIBLE_DEVICES` to bind each GPU to it's server, and then use a reverse proxy (eg [nginx](https://docs.nginx.com/nginx/admin-guide/load-balancer/http-load-balancer/) or [ollama_proxy_server](https://github.com/ParisNeo/ollama_proxy_server)) to distribute the requests to the ollama servers. Another way would be to use slightly different quants for your custom models, eg q4_0, q4_1, q4_K_M.
GiteaMirror commented 2026-04-28 16:48:38 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Sep 5, 2024):

We're tracking adding the ability to load the same model multiple times in issue #3902

<!-- gh-comment-id:2332145470 --> @dhiltgen commented on GitHub (Sep 5, 2024): We're tracking adding the ability to load the same model multiple times in issue #3902
GiteaMirror commented 2026-04-28 16:48:39 -05:00
Author
Owner
Copy Link

@bigsausage commented on GitHub (Sep 6, 2024):

What you have is the same model with three names, so ollama just loads it once. ollama doesn't currently support loading the same model more than once. I say currently because I'm sure I read something some time ago that the developers were considering this, but I can't find a reference so consider that a possible hallucination.

If you want to run multiple copies of the same model, the easiest way at the moment is to start an ollama server on a different port for each GPU, using CUDA_VISIBLE_DEVICES to bind each GPU to it's server, and then use a reverse proxy (eg nginx or ollama_proxy_server) to distribute the requests to the ollama servers.

Another way would be to use slightly different quants for your custom models, eg q4_0, q4_1, q4_K_M.

thanks, that helps a lot

<!-- gh-comment-id:2333000582 --> @bigsausage commented on GitHub (Sep 6, 2024): > What you have is the same model with three names, so ollama just loads it once. ollama doesn't currently support loading the same model more than once. I say currently because I'm sure I read something some time ago that the developers were considering this, but I can't find a reference so consider that a possible hallucination. > > If you want to run multiple copies of the same model, the easiest way at the moment is to start an ollama server on a different port for each GPU, using `CUDA_VISIBLE_DEVICES` to bind each GPU to it's server, and then use a reverse proxy (eg [nginx](https://docs.nginx.com/nginx/admin-guide/load-balancer/http-load-balancer/) or [ollama_proxy_server](https://github.com/ParisNeo/ollama_proxy_server)) to distribute the requests to the ollama servers. > > Another way would be to use slightly different quants for your custom models, eg q4_0, q4_1, q4_K_M. thanks, that helps a lot
GiteaMirror commented 2026-04-28 16:48:40 -05:00
Author
Owner
Copy Link

@bigsausage commented on GitHub (Sep 6, 2024):

We're tracking adding the ability to load the same model multiple times in issue #3902

ok ~ look forward to the new feature~~

<!-- gh-comment-id:2333027755 --> @bigsausage commented on GitHub (Sep 6, 2024): > We're tracking adding the ability to load the same model multiple times in issue #3902 ok ~ look forward to the new feature~~
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#50701
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?