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 #1578] Ollama order of magnitude slower on Apple M1 vs Llama.cpp #47379

New Issue
Closed
opened 2026-04-28 03:39:09 -05:00 by GiteaMirror · 7 comments
GiteaMirror commented 2026-04-28 03:39:09 -05:00
Owner
Copy Link

Originally created by @svilupp on GitHub (Dec 18, 2023).
Original GitHub issue: https://github.com/ollama/ollama/issues/1578

First of all, thank you for the amazing app!

Observation: When I run the same prompt via latest Ollama vs Llama.cpp I get order of magnitude slower generation on Ollama.

  • With Ollama in generation, GPU usage is 0% and from time to time it jumps to 40%
  • With llama.cpp in generation, GPU usage constantly sits at ~99%

Setup:

  • Device: Apple M1 Pro, 32GB ram, shifted memory limit for mixtral to work
  • System: Ventura 13.6
  • Model: dolphin-mixtral:8x7b-v2.5-q4_K_M

Prompt: "Count to 5 and say hi"

Ollama: ollama run dolphin-mixtral:8x7b-v2.5-q4_K_M "Count to 5 then say hi." --verbose

First, I will start by counting from 1 to 5.

  1. One
  2. Two
  3. Three
  4. Four
  5. Five

Now that I have counted to 5, let me say hi! Hi there!

total duration: 5m3.16583525s
load duration: 33.760953875s
prompt eval count: 35 token(s)
prompt eval duration: 24.710485s
prompt eval rate: 1.42 tokens/s
eval count: 54 token(s)
eval duration: 4m4.681389s
eval rate: 0.22 tokens/s

Llama.cpp: ./main -m .ollama/models/blobs/sha256:34855d29fd5901f6ed6fe8112a80dc137bafdeb135d89bf75f9b171e62980ac2 --prompt "[INST] Count to 5 and then say hi. [INST]"

1
2
3
4
5
Hi!
<...it goes on about something else for a bit...it has some stopping issues>

llama_print_timings: load time = 5242.30 ms
llama_print_timings: sample time = 38.25 ms / 425 runs ( 0.09 ms per token, 11109.95 tokens per second)
llama_print_timings: prompt eval time = 800.60 ms / 17 tokens ( 47.09 ms per token, 21.23 tokens per second)
llama_print_timings: eval time = 25695.06 ms / 424 runs ( 60.60 ms per token, 16.50 tokens per second)
llama_print_timings: total time = 26599.97 ms
ggml_metal_free: deallocating
Log end

Any idea what I could doing wrong?

Originally created by @svilupp on GitHub (Dec 18, 2023). Original GitHub issue: https://github.com/ollama/ollama/issues/1578 First of all, thank you for the amazing app! **Observation**: When I run the same prompt via latest Ollama vs Llama.cpp I get order of magnitude slower generation on Ollama. - With Ollama in generation, GPU usage is 0% and from time to time it jumps to 40% - With llama.cpp in generation, GPU usage constantly sits at ~99% **Setup**: - Device: Apple M1 Pro, 32GB ram, shifted memory limit for mixtral to work - System: Ventura 13.6 - Model: dolphin-mixtral:8x7b-v2.5-q4_K_M **Prompt**: "Count to 5 and say hi" **Ollama**: `ollama run dolphin-mixtral:8x7b-v2.5-q4_K_M "Count to 5 then say hi." --verbose` > First, I will start by counting from 1 to 5. > > 1. One > 2. Two > 3. Three > 4. Four > 5. Five > > Now that I have counted to 5, let me say hi! Hi there! > > total duration: 5m3.16583525s > load duration: 33.760953875s > prompt eval count: 35 token(s) > prompt eval duration: 24.710485s > prompt eval rate: 1.42 tokens/s > eval count: 54 token(s) > eval duration: 4m4.681389s > eval rate: 0.22 tokens/s **Llama.cpp**: `./main -m .ollama/models/blobs/sha256:34855d29fd5901f6ed6fe8112a80dc137bafdeb135d89bf75f9b171e62980ac2 --prompt "[INST] Count to 5 and then say hi. [INST]"` > 1 > 2 > 3 > 4 > 5 > Hi! > <...it goes on about something else for a bit...it has some stopping issues> > > llama_print_timings: load time = 5242.30 ms > llama_print_timings: sample time = 38.25 ms / 425 runs ( 0.09 ms per token, 11109.95 tokens per second) > llama_print_timings: prompt eval time = 800.60 ms / 17 tokens ( 47.09 ms per token, 21.23 tokens per second) > llama_print_timings: eval time = 25695.06 ms / 424 runs ( 60.60 ms per token, 16.50 tokens per second) > llama_print_timings: total time = 26599.97 ms > ggml_metal_free: deallocating > Log end Any idea what I could doing wrong?
GiteaMirror commented 2026-04-28 03:39:10 -05:00
Author
Owner
Copy Link

@heresyrj commented on GitHub (Dec 18, 2023):

curious too. how do I turn on Apple Metal for Ollama inference?
I tried:
sudo sysctl iogpu.wired_limit_mb=26624
which helped somewhat.

<!-- gh-comment-id:1860083889 --> @heresyrj commented on GitHub (Dec 18, 2023): curious too. how do I turn on Apple Metal for Ollama inference? I tried: sudo sysctl iogpu.wired_limit_mb=26624 which helped somewhat.
GiteaMirror commented 2026-04-28 03:39:10 -05:00
Author
Owner
Copy Link

@svilupp commented on GitHub (Dec 18, 2023):

Apologies, but it seems to be a duplicate of:

I searched for M1 and llama.cpp but clearly missed these :(

Closing!

<!-- gh-comment-id:1860692459 --> @svilupp commented on GitHub (Dec 18, 2023): Apologies, but it seems to be a duplicate of: - https://github.com/jmorganca/ollama/issues/1556 - https://github.com/jmorganca/ollama/issues/1557 I searched for M1 and llama.cpp but clearly missed these :( Closing!
GiteaMirror commented 2026-04-28 03:39:11 -05:00
Author
Owner
Copy Link

@svilupp commented on GitHub (Dec 20, 2023):

So I've resolved the issue -- it's because Ollama by default offloads only one GPU layer (see: 23dc179350/docs/modelfile.md (L141))

If you set num_gpu=99, you get similar performance as llama.cpp.

Not sure what the rationale is for this default

<!-- gh-comment-id:1864095588 --> @svilupp commented on GitHub (Dec 20, 2023): So I've resolved the issue -- it's because Ollama by default offloads only one GPU layer (see: https://github.com/jmorganca/ollama/blob/23dc1793500c1e8d9709fb6ed57537f9010a0b84/docs/modelfile.md?plain=1#L141) If you set `num_gpu=99`, you get similar performance as llama.cpp. Not sure what the rationale is for this default
GiteaMirror commented 2026-04-28 03:39:12 -05:00
Author
Owner
Copy Link

@vbcoach commented on GitHub (Dec 6, 2024):

@svilupp can you tell us where you set num_gpu=99?

<!-- gh-comment-id:2524234654 --> @vbcoach commented on GitHub (Dec 6, 2024): @svilupp can you tell us where you set num_gpu=99?
GiteaMirror commented 2026-04-28 03:39:12 -05:00
Author
Owner
Copy Link

@svilupp commented on GitHub (Dec 6, 2024):

I just passed it as a kwarg to the model when making the request, eg, https://siml.earth/PromptingTools.jl/dev/examples/working_with_ollama#Local-models-with-Ollama.ai

But it's quite old so maybe the API has changed!

EDIT: Sorry, as per the docs it was under the key "options"! Not directly at the top-level of the request.

<!-- gh-comment-id:2524240702 --> @svilupp commented on GitHub (Dec 6, 2024): I just passed it as a kwarg to the model when making the request, eg, https://siml.earth/PromptingTools.jl/dev/examples/working_with_ollama#Local-models-with-Ollama.ai But it's quite old so maybe the API has changed! EDIT: Sorry, as per the docs it was under the key "options"! Not directly at the top-level of the request.
GiteaMirror commented 2026-04-28 03:39:13 -05:00
Author
Owner
Copy Link

@vbcoach commented on GitHub (Dec 6, 2024):

got it, thanks. I'm going to try to recompile with a model file with this setting and see if things are faster on my M3/64gb

<!-- gh-comment-id:2524251931 --> @vbcoach commented on GitHub (Dec 6, 2024): got it, thanks. I'm going to try to recompile with a model file with this setting and see if things are faster on my M3/64gb
GiteaMirror commented 2026-04-28 03:39:13 -05:00
Author
Owner
Copy Link

@vbcoach commented on GitHub (Dec 6, 2024):

actually, running qwen2.5-coder:32b for a couple simple questions used 94% of my GPU. So, I think Olllama might be optimized already.

<!-- gh-comment-id:2524261055 --> @vbcoach commented on GitHub (Dec 6, 2024): actually, running qwen2.5-coder:32b for a couple simple questions used 94% of my GPU. So, I think Olllama might be optimized already.
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#47379
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?