github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 08:30:05 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #8270] Incorrect NUMA detection logic, fails for AMD Threadripper 1950X #31048

New Issue
Open
opened 2026-04-22 11:10:19 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-22 11:10:19 -05:00
Owner
Copy Link

Originally created by @lukedd on GitHub (Dec 30, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/8270

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

On my AMD Threadripper 1950X CPU with NUMA mode enabled in the BIOS, ollama does not detect that I am running on a NUMA system due to flawed logic in its detection code here: 459d822b51/discover/cpu_common.go (L10-L24)

I can "trick" ollama into detecting NUMA by setting up fake information in /sys/devices/system/cpu/cpu*/topology/physical_package_id using overlayfs, which gives me a ~20% speedup for CPU-only eval-rate (tested with gemma2:27b).

The problem in the logic is that it counts how many physical CPU packages are in the system, but my system has a single CPU package containing 2 dies each with their own memory controller.

A naïve fix would be to look at die_id rather than physical_package_id: this would work for me but I fear there may exist other hardware which has multiple dies sharing a single memory controller. Also even on my system I can disable NUMA in the BIOS so that memory access appears to be uniform - under the hood this interleaves memory access across both NUMA nodes. So in this mode looking at die_id would give the wrong answer.

A better fix would be to look at the actual NUMA node information presented by the kernel under /sys/devices/system/node, e.g. on my system the file /sys/devices/system/node/online contains 0-1 whereas on a uniform memory system it contains 0.

OS

Linux

GPU

Nvidia

CPU

AMD

Ollama version

0.5.4

Originally created by @lukedd on GitHub (Dec 30, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/8270 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? On my AMD Threadripper 1950X CPU with NUMA mode enabled in the BIOS, ollama does not detect that I am running on a NUMA system due to flawed logic in its detection code here: https://github.com/ollama/ollama/blob/459d822b5188dba051e21dfd15b6552543a4bbcf/discover/cpu_common.go#L10-L24 I can "trick" ollama into detecting NUMA by setting up fake information in `/sys/devices/system/cpu/cpu*/topology/physical_package_id` using `overlayfs`, which gives me a ~20% speedup for CPU-only eval-rate (tested with gemma2:27b). The problem in the logic is that it counts how many physical CPU packages are in the system, but my system has a single CPU package containing 2 dies each with their own memory controller. A naïve fix would be to look at `die_id` rather than `physical_package_id`: this would work for me but I fear there may exist other hardware which has multiple dies sharing a single memory controller. Also even on my system I can disable NUMA in the BIOS so that memory access appears to be uniform - under the hood this interleaves memory access across both NUMA nodes. So in this mode looking at `die_id` would give the wrong answer. A better fix would be to look at the actual NUMA node information presented by the kernel under `/sys/devices/system/node`, e.g. on my system the file `/sys/devices/system/node/online` contains `0-1` whereas on a uniform memory system it contains `0`. ### OS Linux ### GPU Nvidia ### CPU AMD ### Ollama version 0.5.4
GiteaMirror added the bug label 2026-04-22 11:10:19 -05:00
GiteaMirror commented 2026-04-22 11:10:20 -05:00
Author
Owner
Copy Link

@jmorganca commented on GitHub (Dec 30, 2024):

@dhiltgen possible to take a look?

<!-- gh-comment-id:2565963191 --> @jmorganca commented on GitHub (Dec 30, 2024): @dhiltgen possible to take a look?
GiteaMirror commented 2026-04-22 11:10:21 -05:00
Author
Owner
Copy Link

@chhu commented on GitHub (Apr 25, 2025):

I'm also using Threadripper with active NUMA and see even a larger boost, almost 80% faster with llama.cpp when using the same model with --numa=distribute . Started experimenting when the numbers did not add up, a test program I wrote in c++ gives 90GByte/s so I expect a token output close to model size / realistic mem bandwidth, which llama.cpp was able to reproduce... (pure cpu)

<!-- gh-comment-id:2831287537 --> @chhu commented on GitHub (Apr 25, 2025): I'm also using Threadripper with active NUMA and see even a larger boost, almost 80% faster with llama.cpp when using the same model with --numa=distribute . Started experimenting when the numbers did not add up, a test program I wrote in c++ gives 90GByte/s so I expect a token output close to model size / realistic mem bandwidth, which llama.cpp was able to reproduce... (pure cpu)
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#31048
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?