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 #12637] Wrap llama.cpp #8389

New Issue
Closed
opened 2026-04-12 21:02:10 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-12 21:02:10 -05:00
Owner
Copy Link

Originally created by @iplayfast on GitHub (Oct 15, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12637

At one time ollama wrapped llama.cpp, now apparently it doesn't.
This thread on twitter/x https://x.com/DFinsterwalder/status/1978372050239516989
shows that benchmarking on ollama vs llama.cpp ollama loses by a large amount.

I was quite surprised by this as I had thought that ollama wrapped llama.cpp
I think maybe ollama needs to go back to it's roots.

Originally created by @iplayfast on GitHub (Oct 15, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12637 At one time ollama wrapped llama.cpp, now apparently it doesn't. This thread on twitter/x https://x.com/DFinsterwalder/status/1978372050239516989 shows that benchmarking on ollama vs llama.cpp ollama loses by a large amount. I was quite surprised by this as I had thought that ollama wrapped llama.cpp I think maybe ollama needs to go back to it's roots.
GiteaMirror added the feature request label 2026-04-12 21:02:10 -05:00
GiteaMirror commented 2026-04-12 21:02:12 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Oct 15, 2025):

The twit doesn't indicate where the table came from and it's not mentioned in the llama.cpp discussion, so I did some tests to get actual data. Since I don't have access to a DGX Spark, these tests were done on a RTX 3080, using the gpt-oss:20b GGUF published by gglm.org.

ollama:

$ for i in {1..5} ; do ollama run gglm-org/gpt-oss:20b --verbose hello 2>&1 ; done | sed -ne 's/^eval rate: *\([^ ]*\) .*/\1/p' | gnuplot -e 'stats "-" nooutput; print STATS_mean,STATS_stddev'
117.6 1.73046814475159

llama.cpp:

$ llama-server -m gpt-oss-20b.gguf -ngl 25 --port 8000 &
$ for i in {1..5} ; do curl -s localhost:8000/v1/chat/completions -d '{"model":"gpt-oss:20b","messages":[{"role":"user","content":"hello"}]}' | jq .timings.predicted_per_second ; done | gnuplot -e 'stats "-" nooutput; print STATS_mean,STATS_stddev'
121.319669439504 1.74871199165077

So llama.cpp is about 3% faster.

<!-- gh-comment-id:3408098461 --> @rick-github commented on GitHub (Oct 15, 2025): The twit doesn't indicate where the table came from and it's not mentioned in the llama.cpp discussion, so I did some tests to get actual data. Since I don't have access to a DGX Spark, these tests were done on a RTX 3080, using the gpt-oss:20b GGUF published by gglm.org. ollama: ```console $ for i in {1..5} ; do ollama run gglm-org/gpt-oss:20b --verbose hello 2>&1 ; done | sed -ne 's/^eval rate: *\([^ ]*\) .*/\1/p' | gnuplot -e 'stats "-" nooutput; print STATS_mean,STATS_stddev' 117.6 1.73046814475159 ``` llama.cpp: ```console $ llama-server -m gpt-oss-20b.gguf -ngl 25 --port 8000 & $ for i in {1..5} ; do curl -s localhost:8000/v1/chat/completions -d '{"model":"gpt-oss:20b","messages":[{"role":"user","content":"hello"}]}' | jq .timings.predicted_per_second ; done | gnuplot -e 'stats "-" nooutput; print STATS_mean,STATS_stddev' 121.319669439504 1.74871199165077 ``` So llama.cpp is about 3% faster.
GiteaMirror commented 2026-04-12 21:02:12 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Oct 16, 2025):

Here are the 120b full precision (BF16 qkv tensor) numbers I'm seeing on 0.11.6rc1:

total duration:       17.1862498s
load duration:        111.612653ms
prompt eval count:    142 token(s)
prompt eval duration: 167.295603ms
prompt eval rate:     848.80 tokens/s
eval count:           709 token(s)
eval duration:        16.713867189s
eval rate:            42.42 tokens/s

Which is faster than what Georgi was posting for the half precision (q8_0 qkv tensor) weights.

I'm going to close this as unhelpful.

<!-- gh-comment-id:3408788920 --> @pdevine commented on GitHub (Oct 16, 2025): Here are the 120b full precision (BF16 qkv tensor) numbers I'm seeing on `0.11.6rc1`: ``` total duration: 17.1862498s load duration: 111.612653ms prompt eval count: 142 token(s) prompt eval duration: 167.295603ms prompt eval rate: 848.80 tokens/s eval count: 709 token(s) eval duration: 16.713867189s eval rate: 42.42 tokens/s ``` Which is faster than what Georgi was posting for the half precision (q8_0 qkv tensor) weights. I'm going to close this as unhelpful.
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#8389
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?