github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-16 13:19:40 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #11260] deepseek-r1:70b is running about 1/3rd to 1/4 the speed it used to run, pre-update #53934

New Issue
Closed
opened 2026-04-29 04:58:27 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-04-29 04:58:27 -05:00
Owner
Copy Link

Originally created by @alpha754293 on GitHub (Jul 2, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/11260

What is the issue?

I have 6700K system with 64 GB of RAM and a pair of 3090s.

I am not really sure what happened exactly, but after updating to the latest version of ollama, I noticed that the deepseek-r1:70b model, which used to run on my pair of 3090s, between 15-20 tokens/s, is now running at around 5 tokens/s.

I am running this through Proxmox 8.4.1, running an Ubuntu 22.04 LXC.

When I run the deepseek-r1:8b model, that is running normally at around 92 tokens/s.

The deepseek-r1:70b model should be able to just fit into the VRAM of my pair of 3090s (being that the model is about 44 GB in size and between the two 3090s, I have a total of 48 GB of VRAM available).

I'm currently running Nvidia (proprietary) Linux driver version 570.133.20.

Your help is greatly appreciated.

Thank you.

Relevant log output

OS

No response

GPU

No response

CPU

No response

Ollama version

No response

Originally created by @alpha754293 on GitHub (Jul 2, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/11260 ### What is the issue? I have 6700K system with 64 GB of RAM and a pair of 3090s. I am not really sure _what_ happened exactly, but after updating to the latest version of ollama, I noticed that the `deepseek-r1:70b` model, which used to run on my pair of 3090s, between 15-20 tokens/s, is now running at around 5 tokens/s. I am running this through Proxmox 8.4.1, running an Ubuntu 22.04 LXC. When I run the `deepseek-r1:8b` model, that is running normally at around 92 tokens/s. The `deepseek-r1:70b` model should be able to _just_ fit into the VRAM of my pair of 3090s (being that the model is about 44 GB in size and between the two 3090s, I have a total of 48 GB of VRAM available). I'm currently running Nvidia (proprietary) Linux driver version 570.133.20. Your help is greatly appreciated. Thank you. ### Relevant log output ```shell ``` ### OS _No response_ ### GPU _No response_ ### CPU _No response_ ### Ollama version _No response_
GiteaMirror added the bug label 2026-04-29 04:58:28 -05:00
GiteaMirror commented 2026-04-29 04:58:30 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jul 2, 2025):

Perhaps layers offloading to the CPU. Server logs will aid in debugging.

<!-- gh-comment-id:3026073659 --> @rick-github commented on GitHub (Jul 2, 2025): Perhaps layers offloading to the CPU. [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) will aid in debugging.
GiteaMirror commented 2026-04-29 04:58:31 -05:00
Author
Owner
Copy Link

@alpha754293 commented on GitHub (Jul 2, 2025):

Perhaps layers offloading to the CPU. Server logs will aid in debugging.

load_tensors: loading model tensors, this can take a while... (mmap = false)
load_tensors: offloading 74 repeating layers to GPU
load_tensors: offloaded 74/81 layers to GPU
load_tensors:    CUDA_Host model buffer size =  3935.23 MiB
load_tensors:        CUDA0 model buffer size = 17904.50 MiB
load_tensors:        CUDA1 model buffer size = 18139.75 MiB
load_tensors:          CPU model buffer size =   563.62 MiB

It would appear that you are correct.

Is there a way to "force" it to offload all 81/81 layers onto the GPU?

<!-- gh-comment-id:3026094132 --> @alpha754293 commented on GitHub (Jul 2, 2025): > Perhaps layers offloading to the CPU. [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) will aid in debugging. ``` load_tensors: loading model tensors, this can take a while... (mmap = false) load_tensors: offloading 74 repeating layers to GPU load_tensors: offloaded 74/81 layers to GPU load_tensors: CUDA_Host model buffer size = 3935.23 MiB load_tensors: CUDA0 model buffer size = 17904.50 MiB load_tensors: CUDA1 model buffer size = 18139.75 MiB load_tensors: CPU model buffer size = 563.62 MiB ``` It would appear that you are correct. Is there a way to "force" it to offload all 81/81 layers onto the GPU?
GiteaMirror commented 2026-04-29 04:58:31 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jul 2, 2025):

Server logs would give some insight to memory allocation. In general, reduce context size (OLLAMA_CONTEXT_LENGTH), reduce parallelism (OLLAMA_NUM_PARALLEL), enable flash attention (OLLAMA_FLASH_ATTENTON), quantize the cache (OLLAMA_KV_CACHE_TYPE), override num_gpu.

<!-- gh-comment-id:3026107145 --> @rick-github commented on GitHub (Jul 2, 2025): [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) would give some insight to memory allocation. In general, reduce context size ([`OLLAMA_CONTEXT_LENGTH`](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-specify-the-context-window-size)), reduce parallelism ([`OLLAMA_NUM_PARALLEL`](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-does-ollama-handle-concurrent-requests)), enable flash attention ([`OLLAMA_FLASH_ATTENTON`](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-enable-flash-attention)), quantize the cache ([`OLLAMA_KV_CACHE_TYPE`](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-can-i-set-the-quantization-type-for-the-kv-cache)), override [`num_gpu`](https://github.com/ollama/ollama/issues/6950#issuecomment-2373663650).
GiteaMirror commented 2026-04-29 04:58:32 -05:00
Author
Owner
Copy Link

@alpha754293 commented on GitHub (Jul 2, 2025):

Thank you.

This has actually been immensely helpful.

If I am using a docker container to run this, what would be the ollama environment variable that I would need to set for num_gpu?

<!-- gh-comment-id:3026119675 --> @alpha754293 commented on GitHub (Jul 2, 2025): Thank you. This has actually been immensely helpful. If I am using a docker container to run this, what would be the ollama environment variable that I would need to set for `num_gpu`?
GiteaMirror commented 2026-04-29 04:58:32 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jul 2, 2025):

There is no environment variable for num_gpu, see here.

<!-- gh-comment-id:3026129263 --> @rick-github commented on GitHub (Jul 2, 2025): There is no environment variable for `num_gpu`, see [here](https://github.com/ollama/ollama/issues/6950#issuecomment-2373663650).
GiteaMirror commented 2026-04-29 04:58:33 -05:00
Author
Owner
Copy Link

@alpha754293 commented on GitHub (Jul 2, 2025):

Thank you. I will close this issue now. I appreciate your help.

<!-- gh-comment-id:3026132048 --> @alpha754293 commented on GitHub (Jul 2, 2025): Thank you. I will close this issue now. I appreciate your help.
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#53934
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?