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 #8331] Llama3.3:70b-intstruct-q5_K_M got EOF on MacMini M4 pro 64GB RAM #5339

New Issue
Closed
opened 2026-04-12 16:32:22 -05:00 by GiteaMirror · 4 comments
GiteaMirror commented 2026-04-12 16:32:22 -05:00
Owner
Copy Link

Originally created by @simon4ddvd on GitHub (Jan 7, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/8331

What is the issue?

Mac OS: Sequoia 15.2
Llama3.3:70b-instruct-q4_K_m is running OK.
Llama3.3:70b-instruct-q6_K is running OK.
But Llama3.3:70b-instruct-q5_K_m got EOF as below:

% ollama run llama3.3:70b-instruct-q5_k_m
>>> hi
Error: POST predict: Post "http://127.0.0.1:53684/completion": EOF

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.5.4

Originally created by @simon4ddvd on GitHub (Jan 7, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/8331 ### What is the issue? Mac OS: Sequoia 15.2 Llama3.3:70b-instruct-q4_K_m is running OK. Llama3.3:70b-instruct-q6_K is running OK. But Llama3.3:70b-instruct-q5_K_m got EOF as below: ``` % ollama run llama3.3:70b-instruct-q5_k_m >>> hi Error: POST predict: Post "http://127.0.0.1:53684/completion": EOF ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.5.4
GiteaMirror added the bug label 2026-04-12 16:32:22 -05:00
GiteaMirror commented 2026-04-12 16:32:23 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 7, 2025):

Server logs may aid in debugging.

<!-- gh-comment-id:2574555013 --> @rick-github commented on GitHub (Jan 7, 2025): [Server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) may aid in debugging.
GiteaMirror commented 2026-04-12 16:32:23 -05:00
Author
Owner
Copy Link

@simon4ddvd commented on GitHub (Jan 7, 2025):

server.20250107.log
Please refer to the uploaded log. The error messages as below:
ggml_metal_graph_compute: command buffer 0 failed with status 5
error: Insufficient Memory (00000008:kIOGPUCommandBufferCallbackErrorOutOfMemory)
llama_graph_compute: ggml_backend_sched_graph_compute_async failed with error -1
llama_decode: failed to decode, ret = -3
panic: failed to decode batch: llama_decode failed with code -3

<!-- gh-comment-id:2575099324 --> @simon4ddvd commented on GitHub (Jan 7, 2025): [server.20250107.log](https://github.com/user-attachments/files/18332112/server.20250107.log) Please refer to the uploaded log. The error messages as below: ggml_metal_graph_compute: command buffer 0 failed with status 5 error: Insufficient Memory (00000008:kIOGPUCommandBufferCallbackErrorOutOfMemory) llama_graph_compute: ggml_backend_sched_graph_compute_async failed with error -1 llama_decode: failed to decode, ret = -3 panic: failed to decode batch: llama_decode failed with code -3
GiteaMirror commented 2026-04-12 16:32:24 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 7, 2025):

time=2025-01-07T18:10:10.780+08:00 level=INFO source=memory.go:356 msg="offload to metal" layers.requested=-1 layers.model=81 layers.offload=81 layers.split="" memory.available="[48.0 GiB]" memory.gpu_overhead="0 B" memory.required.full="47.9 GiB" memory.required.partial="47.9 GiB" memory.required.kv="640.0 MiB" memory.required.allocations="[47.9 GiB]" memory.weights.total="45.7 GiB" memory.weights.repeating="44.9 GiB" memory.weights.nonrepeating="822.0 MiB" memory.graph.full="324.0 MiB" memory.graph.partial="324.0 MiB"

The GPU has 48G available and ollama wants to use 47.9G, so it's a really tight squeeze. I suspect that during the inference llama.cpp is making a temporary allocation which exceeds the available spare VRAM and the process dies with an OOM error. The llama3.3:70b-instruct-q6_K model, which is larger, is probably overflowing in to system RAM and leaving enough room for the transient memory requirerments.

I'm not an Apple user but it's my understanding that you can increase the amount of VRAM your GPU has access to. If you can do that on your system, that will solve your problem. Otherwise, the only solution is to reduce the number of layers (num_gpu) that is being loaded into VRAM. See here for ways set num_gpu. Currently ollama is offloading 81 layers, try 75 and see how it goes.

<!-- gh-comment-id:2575437516 --> @rick-github commented on GitHub (Jan 7, 2025): ``` time=2025-01-07T18:10:10.780+08:00 level=INFO source=memory.go:356 msg="offload to metal" layers.requested=-1 layers.model=81 layers.offload=81 layers.split="" memory.available="[48.0 GiB]" memory.gpu_overhead="0 B" memory.required.full="47.9 GiB" memory.required.partial="47.9 GiB" memory.required.kv="640.0 MiB" memory.required.allocations="[47.9 GiB]" memory.weights.total="45.7 GiB" memory.weights.repeating="44.9 GiB" memory.weights.nonrepeating="822.0 MiB" memory.graph.full="324.0 MiB" memory.graph.partial="324.0 MiB" ``` The GPU has 48G available and ollama wants to use 47.9G, so it's a really tight squeeze. I suspect that during the inference llama.cpp is making a temporary allocation which exceeds the available spare VRAM and the process dies with an OOM error. The llama3.3:70b-instruct-q6_K model, which is larger, is probably overflowing in to system RAM and leaving enough room for the transient memory requirerments. I'm not an Apple user but it's my understanding that you can [increase](https://github.com/ggerganov/llama.cpp/discussions/2182#discussioncomment-7698315) the amount of VRAM your GPU has access to. If you can do that on your system, that will solve your problem. Otherwise, the only solution is to reduce the number of layers (`num_gpu`) that is being loaded into VRAM. See [here](https://github.com/ollama/ollama/issues/6950#issuecomment-2373663650) for ways set `num_gpu`. Currently ollama is offloading 81 layers, try 75 and see how it goes.
GiteaMirror commented 2026-04-12 16:32:24 -05:00
Author
Owner
Copy Link

@simon4ddvd commented on GitHub (Jan 8, 2025):

Thank you so much.
After I set iogpu.wired_limit_mb=50176 (49GB at least), it's working fine.

sudo sysctl iogpu.wired_limit_mb=50176
<!-- gh-comment-id:2576776867 --> @simon4ddvd commented on GitHub (Jan 8, 2025): Thank you so much. After I set iogpu.wired_limit_mb=50176 (49GB at least), it's working fine. ``` sudo sysctl iogpu.wired_limit_mb=50176 ```
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#5339
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?