github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-07-16 21:25:46 -05:00
Code Issues 16.9k Packages Projects Releases 122 Wiki Activity

[GH-ISSUE #15993] [Bug] 500 Error with OLLAMA_FLASH_ATTENTION=true on Intel iGPU (Vulkan) when processing high-res images #87870

New Issue
Open
opened 2026-05-10 06:29:14 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-05-10 06:29:14 -05:00
Owner
Copy Link

Originally created by @crackerfly on GitHub (May 6, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15993

What is the issue?

Describe the bug

When using the Vulkan backend on an Intel iGPU, setting OLLAMA_FLASH_ATTENTION=true causes a 500 Error when attempting to process high-resolution images with multimodal models.

If I set OLLAMA_FLASH_ATTENTION=false, the models can process the images without crashing, but the generation quality and output accuracy drop significantly.

Steps to reproduce

  1. Set the environment variable: OLLAMA_FLASH_ATTENTION=true
  2. Run Ollama using the Vulkan backend on an Intel Core Ultra hardware setup.
  3. Load a multimodal model (e.g., qwen3.5:9b or qwen3.6:35b).
  4. Input a high-resolution image in the prompt.
  5. The server immediately returns a 500 Internal Server Error.

Expected behavior

The model should successfully process the high-resolution image with Flash Attention enabled, similar to how it behaves on other GPU architectures.

Environment

  • OS: Windows 11 25H2
  • Hardware: Intel Core Ultra 7 358H + iGPU (B390)
  • Ollama Version: 0.19.0 ~ 0.23.1
  • Backend: Vulkan
  • Models Tested: qwen3.5:9b, qwen3.6:35b

Additional context & Troubleshooting

I have done some isolation testing to narrow down the issue:

  • Upstream Testing: I tested the exact same workflow using the latest version of llama.cpp (Vulkan build) directly. It works perfectly fine with Flash Attention enabled on this Intel hardware.
  • Alternative Hardware: I tested Ollama with OLLAMA_FLASH_ATTENTION=true on an AMD iGPU environment, and it works perfectly without any 500 errors.

Proposed Solution

Given that the latest upstream llama.cpp handles this correctly, this appears to be an Intel-specific Vulkan bug that has already been resolved upstream. Syncing/updating the ggml-vulkan backend in Ollama to the latest version should fix this issue.

Relevant log output

OS

Windows

GPU

Intel

CPU

Intel

Ollama version

0.19.0 ~ 0.23.1

Originally created by @crackerfly on GitHub (May 6, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15993 ### What is the issue? ### Describe the bug When using the Vulkan backend on an Intel iGPU, setting `OLLAMA_FLASH_ATTENTION=true` causes a `500 Error` when attempting to process high-resolution images with multimodal models. If I set `OLLAMA_FLASH_ATTENTION=false`, the models can process the images without crashing, but the generation quality and output accuracy drop significantly. ### Steps to reproduce 1. Set the environment variable: `OLLAMA_FLASH_ATTENTION=true` 2. Run Ollama using the Vulkan backend on an Intel Core Ultra hardware setup. 3. Load a multimodal model (e.g., `qwen3.5:9b` or `qwen3.6:35b`). 4. Input a high-resolution image in the prompt. 5. The server immediately returns a `500 Internal Server Error`. ### Expected behavior The model should successfully process the high-resolution image with Flash Attention enabled, similar to how it behaves on other GPU architectures. ### Environment * **OS:** Windows 11 25H2 * **Hardware:** Intel Core Ultra 7 358H + iGPU (B390) * **Ollama Version:** 0.19.0 ~ 0.23.1 * **Backend:** Vulkan * **Models Tested:** `qwen3.5:9b`, `qwen3.6:35b` ### Additional context & Troubleshooting I have done some isolation testing to narrow down the issue: * **Upstream Testing:** I tested the exact same workflow using the latest version of `llama.cpp` (Vulkan build) directly. It works perfectly fine with Flash Attention enabled on this Intel hardware. * **Alternative Hardware:** I tested Ollama with `OLLAMA_FLASH_ATTENTION=true` on an AMD iGPU environment, and it works perfectly without any 500 errors. ### Proposed Solution Given that the latest upstream `llama.cpp` handles this correctly, this appears to be an Intel-specific Vulkan bug that has already been resolved upstream. Syncing/updating the `ggml-vulkan` backend in Ollama to the latest version should fix this issue. ### Relevant log output ```shell ``` ### OS Windows ### GPU Intel ### CPU Intel ### Ollama version 0.19.0 ~ 0.23.1
GiteaMirror added the bug label 2026-05-10 06:29:14 -05:00
GiteaMirror commented 2026-05-10 06:29:16 -05:00
Author
Owner
Copy Link

@crackerfly commented on GitHub (May 9, 2026):

Is anyone here?

<!-- gh-comment-id:4412265677 --> @crackerfly commented on GitHub (May 9, 2026): Is anyone here?
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#87870
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?