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 #9697] Sending multiple images to Gemma3 on Mac causes an EOF #6331

New Issue
Closed
opened 2026-04-12 17:50:13 -05:00 by GiteaMirror · 9 comments
GiteaMirror commented 2026-04-12 17:50:13 -05:00
Owner
Copy Link

Originally created by @BruceMacD on GitHub (Mar 12, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9697

Originally assigned to: @jessegross on GitHub.

What is the issue?

Ollama crashes with a GGML assertion failure when attempting to run inference on Apple Silicon Mac using Metal GPU acceleration when given multiple images.

The issue appears to be a type mismatch in the GGML library when using the Metal framework. The code is expecting a tensor of type F32 (32-bit floating point), but received a different type, causing an assertion failure.

Relevant log output

ggml-metal.m:3253: GGML_ASSERT(src1->type == GGML_TYPE_F32) failed
SIGABRT: abort
PC=0x18cedea60 m=61 sigcode=0

github.com/ollama/ollama/ml/backend/ggml.Context.Compute({0x1400037c000, 0x600001cf4cc0, 0x141260e80, 0x0, 0x2000}, {0x1408d5131c0, 0x1, 0x141260e80?})
        /Users/runner/work/ollama/ollama/ml/backend/ggml/ggml.go:497 +0x9c fp=0x14000599b60 sp=0x14000599ad0 pc=0x10485919c

Steps to Reproduce

  1. Run Ollama with Metal GPU acceleration enabled (default on macOS)
  2. ollama run gemma3:27b "What is in these two images? ./photo.jpg ./other.jpg"
  3. Process crashes with the above error

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

0.6.0

Originally created by @BruceMacD on GitHub (Mar 12, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9697 Originally assigned to: @jessegross on GitHub. ### What is the issue? Ollama crashes with a GGML assertion failure when attempting to run inference on Apple Silicon Mac using Metal GPU acceleration when given multiple images. The issue appears to be a type mismatch in the GGML library when using the Metal framework. The code is expecting a tensor of type F32 (32-bit floating point), but received a different type, causing an assertion failure. ### Relevant log output ```shell ggml-metal.m:3253: GGML_ASSERT(src1->type == GGML_TYPE_F32) failed SIGABRT: abort PC=0x18cedea60 m=61 sigcode=0 ``` ``` github.com/ollama/ollama/ml/backend/ggml.Context.Compute({0x1400037c000, 0x600001cf4cc0, 0x141260e80, 0x0, 0x2000}, {0x1408d5131c0, 0x1, 0x141260e80?}) /Users/runner/work/ollama/ollama/ml/backend/ggml/ggml.go:497 +0x9c fp=0x14000599b60 sp=0x14000599ad0 pc=0x10485919c ``` ### Steps to Reproduce 1. Run Ollama with Metal GPU acceleration enabled (default on macOS) 2. `ollama run gemma3:27b "What is in these two images? ./photo.jpg ./other.jpg"` 3. Process crashes with the above error ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version 0.6.0
GiteaMirror added the bug label 2026-04-12 17:50:13 -05:00
GiteaMirror commented 2026-04-12 17:50:15 -05:00
Author
Owner
Copy Link

@powerpiggy commented on GitHub (Mar 13, 2025):

I have the same problem.

<!-- gh-comment-id:2719759421 --> @powerpiggy commented on GitHub (Mar 13, 2025): I have the same problem.
GiteaMirror commented 2026-04-12 17:50:15 -05:00
Author
Owner
Copy Link

@noobmaster29 commented on GitHub (Mar 13, 2025):

I'm running on windows and I am having the same issue (just installed everything fresh today). I thought it was perhaps a memory issue but I tried it with 27b and 4b and I get the same issue. Does not happen with only 1 image attached.

With 27b, if I attached 1 photo, and then a 2nd photo in another request, it also crashes with the same error. However, does not seem to be the case with 4b.

Edit: Ran the same request on openrouter and it runs fine.

Image

Image

<!-- gh-comment-id:2720383685 --> @noobmaster29 commented on GitHub (Mar 13, 2025): I'm running on windows and I am having the same issue (just installed everything fresh today). I thought it was perhaps a memory issue but I tried it with 27b and 4b and I get the same issue. Does not happen with only 1 image attached. With 27b, if I attached 1 photo, and then a 2nd photo in another request, it also crashes with the same error. However, does not seem to be the case with 4b. Edit: Ran the same request on openrouter and it runs fine. ![Image](https://github.com/user-attachments/assets/200006af-ce98-4275-ac59-793f8952c596) ![Image](https://github.com/user-attachments/assets/442e8cb9-6b59-43f8-ab49-fd0852485394)
GiteaMirror commented 2026-04-12 17:50:17 -05:00
Author
Owner
Copy Link

@rdzotz commented on GitHub (Mar 13, 2025):

I get the following on ubuntu

ollama run gemma3:27b
>>> Please describe the following images: ..._after.jpg _before.jpg 
Added image '_after.jpg'
Added image '_before.jpg'
Error: vision model only supports a single image per message
<!-- gh-comment-id:2720595111 --> @rdzotz commented on GitHub (Mar 13, 2025): I get the following on ubuntu ``` ollama run gemma3:27b >>> Please describe the following images: ..._after.jpg _before.jpg Added image '_after.jpg' Added image '_before.jpg' Error: vision model only supports a single image per message ```
GiteaMirror commented 2026-04-12 17:50:17 -05:00
Author
Owner
Copy Link

@fighter3005 commented on GitHub (Mar 13, 2025):

Same thing on ubuntu (docker).

Image

Image

It behaves like Llama3.2-vision. However, both should support multiple images, if I am not mistaken.

I read here, that this is how it should be for Llama3.2-vision.
I still believe this is wrong. vLLM also lists both models with multiple image support. This has been and still is a major inconvenience with Ollama.

Disclaimer: I have not looked into either codebase and might be terribly wrong here. :)

<!-- gh-comment-id:2721244663 --> @fighter3005 commented on GitHub (Mar 13, 2025): Same thing on ubuntu (docker). ![Image](https://github.com/user-attachments/assets/c72e231a-e1af-4506-b5b3-f4656357b633) ![Image](https://github.com/user-attachments/assets/fed29f71-9af9-4934-ad51-63377010f9ad) It behaves like Llama3.2-vision. However, both should support multiple images, if I am not mistaken. I read [here](https://github.com/ollama/ollama/issues/7477), that this is how it should be for Llama3.2-vision. I still believe this is wrong. vLLM also lists both models with multiple image support. This has been and still is a major inconvenience with Ollama. Disclaimer: I have not looked into either codebase and might be terribly wrong here. :)
GiteaMirror commented 2026-04-12 17:50:17 -05:00
Author
Owner
Copy Link

@rdzotz commented on GitHub (Mar 13, 2025):

Same thing on ubuntu (docker).

Image

Image

It behaves like Llama3.2-vision. However, both should support multiple images, if I am not mistaken.

I read here, that this is how it should be for Llama3.2-vision. I still believe this is wrong. vLLM also lists both models with multiple image support. This has been and still is a major inconvenience with Ollama.

Disclaimer: I have not looked into either codebase and might be terribly wrong here. :)

llama3.2-vision works for me with multiple image input.

<!-- gh-comment-id:2721667857 --> @rdzotz commented on GitHub (Mar 13, 2025): > Same thing on ubuntu (docker). > > ![Image](https://github.com/user-attachments/assets/c72e231a-e1af-4506-b5b3-f4656357b633) > > ![Image](https://github.com/user-attachments/assets/fed29f71-9af9-4934-ad51-63377010f9ad) > > It behaves like Llama3.2-vision. However, both should support multiple images, if I am not mistaken. > > I read [here](https://github.com/ollama/ollama/issues/7477), that this is how it should be for Llama3.2-vision. I still believe this is wrong. vLLM also lists both models with multiple image support. This has been and still is a major inconvenience with Ollama. > > Disclaimer: I have not looked into either codebase and might be terribly wrong here. :) llama3.2-vision works for me with multiple image input.
GiteaMirror commented 2026-04-12 17:50:19 -05:00
Author
Owner
Copy Link

@fighter3005 commented on GitHub (Mar 13, 2025):

Same thing on ubuntu (docker).
Image
Image
It behaves like Llama3.2-vision. However, both should support multiple images, if I am not mistaken.
I read here, that this is how it should be for Llama3.2-vision. I still believe this is wrong. vLLM also lists both models with multiple image support. This has been and still is a major inconvenience with Ollama.
Disclaimer: I have not looked into either codebase and might be terribly wrong here. :)

llama3.2-vision works for me with multiple image input.

Multiple images in one prompt? What quantization do you use? And what settings do you use with Ollama?
MiniCPM-V-2.6 works for me with multiple inputs through open webui, llama3.2-vision not. I will test through ollama directly...

Update: Llama3.2-vision does not support multiple images in one prompt. So I cannot say: "compare the images ./image1.png ./image2.png"

<!-- gh-comment-id:2722089980 --> @fighter3005 commented on GitHub (Mar 13, 2025): > > Same thing on ubuntu (docker). > > ![Image](https://github.com/user-attachments/assets/c72e231a-e1af-4506-b5b3-f4656357b633) > > ![Image](https://github.com/user-attachments/assets/fed29f71-9af9-4934-ad51-63377010f9ad) > > It behaves like Llama3.2-vision. However, both should support multiple images, if I am not mistaken. > > I read [here](https://github.com/ollama/ollama/issues/7477), that this is how it should be for Llama3.2-vision. I still believe this is wrong. vLLM also lists both models with multiple image support. This has been and still is a major inconvenience with Ollama. > > Disclaimer: I have not looked into either codebase and might be terribly wrong here. :) > > llama3.2-vision works for me with multiple image input. Multiple images in one prompt? What quantization do you use? And what settings do you use with Ollama? MiniCPM-V-2.6 works for me with multiple inputs through open webui, llama3.2-vision not. I will test through ollama directly... Update: Llama3.2-vision does not support multiple images in one prompt. So I cannot say: "compare the images ./image1.png ./image2.png"
GiteaMirror commented 2026-04-12 17:50:20 -05:00
Author
Owner
Copy Link

@noobmaster29 commented on GitHub (Mar 14, 2025):

Very cool, do we know when the fix will be commited? I'm happy to test and report back.

<!-- gh-comment-id:2723097756 --> @noobmaster29 commented on GitHub (Mar 14, 2025): Very cool, do we know when the fix will be commited? I'm happy to test and report back.
GiteaMirror commented 2026-04-12 17:50:21 -05:00
Author
Owner
Copy Link

@BruceMacD commented on GitHub (Mar 14, 2025):

Hey everyone, Jesse has a fix in review for this, we have a pre-release built right now, so it won't make it in for that, but it will be in the next one!

<!-- gh-comment-id:2725236407 --> @BruceMacD commented on GitHub (Mar 14, 2025): Hey everyone, Jesse has a fix in review for this, we have a pre-release built right now, so it won't make it in for that, but it will be in the next one!
GiteaMirror commented 2026-04-12 17:50:22 -05:00
Author
Owner
Copy Link

@noobmaster29 commented on GitHub (Mar 19, 2025):

Just tested this after the latest update and it works perfectly. Thank you very much!

<!-- gh-comment-id:2735033488 --> @noobmaster29 commented on GitHub (Mar 19, 2025): Just tested this after the latest update and it works perfectly. Thank you very much!
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#6331
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?