github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #7768] Model not loaded on all GPUs for load balancing #51473

New Issue
Closed
opened 2026-04-28 20:18:37 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-28 20:18:37 -05:00
Owner
Copy Link

Originally created by @brauliobo on GitHub (Nov 20, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/7768

What is the issue?

I expect that on a Multi GPU system it would load the model on all GPUs with the docker container loaded with --gpus all to balance the requests load between them.

Output of docker logs ollama:

ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes

Output of docker exec -it ollama nvidia-smi:

Wed Nov 20 20:00:03 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 565.57.01              Driver Version: 565.57.01      CUDA Version: 12.7     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3060        On  |   00000000:01:00.0  On |                  N/A |
| 54%   67C    P2             99W /  100W |    7203MiB /  12288MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 3060        On  |   00000000:05:00.0 Off |                  N/A |
| 41%   63C    P2             99W /  100W |    8057MiB /  12288MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    1   N/A  N/A        52      C   ...unners/cuda_v12/ollama_llama_server       3344MiB |
+-----------------------------------------------------------------------------------------+

OS

Linux, Docker

GPU

Nvidia

CPU

AMD

Ollama version

0.4.2 from docker

Originally created by @brauliobo on GitHub (Nov 20, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/7768 ### What is the issue? I expect that on a Multi GPU system it would load the model on all GPUs with the docker container loaded with `--gpus all` to balance the requests load between them. Output of `docker logs ollama`: ``` ggml_cuda_init: found 1 CUDA devices: Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes ``` Output of `docker exec -it ollama nvidia-smi`: ``` Wed Nov 20 20:00:03 2024 +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 565.57.01 Driver Version: 565.57.01 CUDA Version: 12.7 | |-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| | 0 NVIDIA GeForce RTX 3060 On | 00000000:01:00.0 On | N/A | | 54% 67C P2 99W / 100W | 7203MiB / 12288MiB | 100% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ | 1 NVIDIA GeForce RTX 3060 On | 00000000:05:00.0 Off | N/A | | 41% 63C P2 99W / 100W | 8057MiB / 12288MiB | 100% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ +-----------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| | 1 N/A N/A 52 C ...unners/cuda_v12/ollama_llama_server 3344MiB | +-----------------------------------------------------------------------------------------+ ``` ### OS Linux, Docker ### GPU Nvidia ### CPU AMD ### Ollama version 0.4.2 from docker
GiteaMirror added the bug label 2026-04-28 20:18:37 -05:00
GiteaMirror commented 2026-04-28 20:18:50 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Nov 20, 2024):

https://github.com/ollama/ollama/issues/7648#issuecomment-2473561990

If you want ollama to spread the model across all GPUs, set OLLAMA_SCHED_SPREAD: ecf41eed05/envconfig/config.go (L249)

<!-- gh-comment-id:2489470778 --> @rick-github commented on GitHub (Nov 20, 2024): https://github.com/ollama/ollama/issues/7648#issuecomment-2473561990 If you want ollama to spread the model across all GPUs, set `OLLAMA_SCHED_SPREAD`: https://github.com/ollama/ollama/blob/ecf41eed0595fb031f1addc179f6abb86d8405f8/envconfig/config.go#L249
GiteaMirror commented 2026-04-28 20:18:52 -05:00
Author
Owner
Copy Link

@brauliobo commented on GitHub (Nov 20, 2024):

thanks! it looks split into the 2 GPUs:

braulio @ whitebeast ➜  ollama  docker exec -it ollama nvidia-smi
Wed Nov 20 20:29:05 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 565.57.01              Driver Version: 565.57.01      CUDA Version: 12.7     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3060        On  |   00000000:01:00.0 Off |                  N/A |
| 52%   65C    P2             99W /  100W |    9298MiB /  12288MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 3060        On  |   00000000:05:00.0 Off |                  N/A |
| 40%   63C    P2             99W /  100W |    6647MiB /  12288MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A        58      C   ...unners/cuda_v12/ollama_llama_server       2082MiB |
|    1   N/A  N/A        58      C   ...unners/cuda_v12/ollama_llama_server       1934MiB |
+-----------------------------------------------------------------------------------------+

I expected it instead to load the whole model into each GPU to better load balance multiple requests.

Also still the bottleneck will be the CPU due to https://github.com/ollama/ollama/issues/6913

Thanks again and that should be what is possible for now

<!-- gh-comment-id:2489491295 --> @brauliobo commented on GitHub (Nov 20, 2024): thanks! it looks split into the 2 GPUs: ``` braulio @ whitebeast ➜ ollama docker exec -it ollama nvidia-smi Wed Nov 20 20:29:05 2024 +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 565.57.01 Driver Version: 565.57.01 CUDA Version: 12.7 | |-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| | 0 NVIDIA GeForce RTX 3060 On | 00000000:01:00.0 Off | N/A | | 52% 65C P2 99W / 100W | 9298MiB / 12288MiB | 100% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ | 1 NVIDIA GeForce RTX 3060 On | 00000000:05:00.0 Off | N/A | | 40% 63C P2 99W / 100W | 6647MiB / 12288MiB | 100% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ +-----------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| | 0 N/A N/A 58 C ...unners/cuda_v12/ollama_llama_server 2082MiB | | 1 N/A N/A 58 C ...unners/cuda_v12/ollama_llama_server 1934MiB | +-----------------------------------------------------------------------------------------+ ``` I expected it instead to load the whole model into each GPU to better load balance multiple requests. Also still the bottleneck will be the CPU due to https://github.com/ollama/ollama/issues/6913 Thanks again and that should be what is possible for now
GiteaMirror commented 2026-04-28 20:18:55 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Nov 20, 2024):

https://github.com/ollama/ollama/issues/7648#issuecomment-2473561990 for ways to load balance multiple requests.

<!-- gh-comment-id:2489496287 --> @rick-github commented on GitHub (Nov 20, 2024): https://github.com/ollama/ollama/issues/7648#issuecomment-2473561990 for ways to load balance multiple requests.
GiteaMirror commented 2026-04-28 20:18:58 -05:00
Author
Owner
Copy Link

@brauliobo commented on GitHub (Nov 20, 2024):

thanks again, I've set -e OLLAMA_NUM_PARALLEL=2, but I wonder how much slower it is compared to running 2 servers separately with one GPU set with CUDA_VISIBLE_DEVICES for each, as its setup is much more complicated.

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 565.57.01              Driver Version: 565.57.01      CUDA Version: 12.7     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 3060        On  |   00000000:01:00.0 Off |                  N/A |
| 52%   65C    P2             99W /  100W |    8996MiB /  12288MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA GeForce RTX 3060        On  |   00000000:05:00.0 Off |                  N/A |
| 40%   63C    P2             99W /  100W |    6213MiB /  12288MiB |    100%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A        59      C   ...unners/cuda_v12/ollama_llama_server       1720MiB |
|    1   N/A  N/A        59      C   ...unners/cuda_v12/ollama_llama_server       1500MiB |
+-----------------------------------------------------------------------------------------+
<!-- gh-comment-id:2489510738 --> @brauliobo commented on GitHub (Nov 20, 2024): thanks again, I've set `-e OLLAMA_NUM_PARALLEL=2`, but I wonder how much slower it is compared to running 2 servers separately with one GPU set with `CUDA_VISIBLE_DEVICES` for each, as its setup is much more complicated. ``` +-----------------------------------------------------------------------------------------+ | NVIDIA-SMI 565.57.01 Driver Version: 565.57.01 CUDA Version: 12.7 | |-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| | 0 NVIDIA GeForce RTX 3060 On | 00000000:01:00.0 Off | N/A | | 52% 65C P2 99W / 100W | 8996MiB / 12288MiB | 100% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ | 1 NVIDIA GeForce RTX 3060 On | 00000000:05:00.0 Off | N/A | | 40% 63C P2 99W / 100W | 6213MiB / 12288MiB | 100% Default | | | | N/A | +-----------------------------------------+------------------------+----------------------+ +-----------------------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| | 0 N/A N/A 59 C ...unners/cuda_v12/ollama_llama_server 1720MiB | | 1 N/A N/A 59 C ...unners/cuda_v12/ollama_llama_server 1500MiB | +-----------------------------------------------------------------------------------------+ ```
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#51473
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?