github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #1303] Memory required to run differs from expectation #47185

New Issue
Closed
opened 2026-04-28 03:25:05 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-28 03:25:05 -05:00
Owner
Copy Link

Originally created by @technovangelist on GitHub (Nov 28, 2023).
Original GitHub issue: https://github.com/ollama/ollama/issues/1303

After discussing internally, it was suggested that as long as we have enough total memory across ram and vram, the model should load. Layers are loaded into main memory then offloaded into vram. So I tried with different memory sizes and number of attached T4 cards with 16-ish GB vram each.

When there is 16 GB RAM and 4x T4, adding up to 76GB, I get a timeout: Error: timed out waiting for llama runner to start.

I get the same error with 30GB RAM, and 60GB RAM. Its not until i go to the next threshold (100 GB) with the 4x T4 that it loads correctly.

We need to clarify how much memory is required to run models. This is easy where we started on Apple Silicon because there is one number. But more complicated on Nvidia.

Originally created by @technovangelist on GitHub (Nov 28, 2023). Original GitHub issue: https://github.com/ollama/ollama/issues/1303 After discussing internally, it was suggested that as long as we have enough total memory across ram and vram, the model should load. Layers are loaded into main memory then offloaded into vram. So I tried with different memory sizes and number of attached T4 cards with 16-ish GB vram each. When there is 16 GB RAM and 4x T4, adding up to 76GB, I get a timeout: `Error: timed out waiting for llama runner to start`. I get the same error with 30GB RAM, and 60GB RAM. Its not until i go to the next threshold (100 GB) with the 4x T4 that it loads correctly. We need to clarify how much memory is required to run models. This is easy where we started on Apple Silicon because there is one number. But more complicated on Nvidia.
GiteaMirror commented 2026-04-28 03:25:06 -05:00
Author
Owner
Copy Link

@phalexo commented on GitHub (Dec 14, 2023):

@technovangelist The function that computes NVIDIA GPU VRAM ignores CUDA_VISIBLE_DEVICES variable and considers VRAM from all GPUs, even if some of them cannot be used.

<!-- gh-comment-id:1856338876 --> @phalexo commented on GitHub (Dec 14, 2023): @technovangelist The function that computes NVIDIA GPU VRAM ignores CUDA_VISIBLE_DEVICES variable and considers VRAM from all GPUs, even if some of them cannot be used.
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#47185
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?