[Bug]: qwen3-30b-a3b on ovms, works on CPU, crashs with out of memory on iGPU

[ikirsh]

OpenVINO Version

2025.4.1, 2026.0

Operating System

Other (Please specify in description)

Device used for inference

GPU

Framework

None

Model used

qwen3-30b-a3b

Issue description

On Intel Ultra 7 265K with 32GB RAM + 128GB swap file, Kbuntu 25.10.

The model works well on the CPU with about 16GB RAM usage. When running on the iGPU, memory usage goes up until it fills the entire 32GB and then everything is killed.

The buffer length was reduced to 2048 but it didn't help.

The model is on the list of "AI Models verified for OpenVINO".

Step-by-step reproduction

1. Exported from optimum-cli

2. On CPU it worked:

LD_PRELOAD="/opt/openvino/ovms/lib/libopenvino_tokenizers.so \
/opt/openvino/ovms/lib/libopenvino_genai.so" \
/opt/openvino/ovms/bin/ovms \
  --model_repository_path /opt/openvino/models \
  --model_name Qwen3-30B-int4 \
  --task text_generation \
  --port 9001 \
  --rest_port 8000 \
  --target_device CPU

3. On GPU it crashed:

LD_PRELOAD="/opt/openvino/ovms/lib/libopenvino_tokenizers.so \
/opt/openvino/ovms/lib/libopenvino_genai.so" \
/opt/openvino/ovms/bin/ovms \
  --model_repository_path /opt/openvino/models \
  --model_name Qwen3-30B-int4 \
  --task text_generation \
  --port 9001 \
  --rest_port 8000 \
  --target_device GPU

Relevant log output

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.

[ikirsh]

Checked now, and the same happens with gpt_oss_20b_ov_int4, works on CPU, crashes on iGPU.

Note that other models, e.g. Qwen3-8B-int4-cw-ov, work well on both.

[Wan-Intel]

Hi @ikirsh
Could you please share the optimum-cli command that you used to download and convert qwen3-30b-a3b?

On another note, you may open a new issue regarding the bug on gpt_oss_20b_ov_int4 as current thread will be focus on qwen3-30b-a3b.

[ikirsh]

I have re-tested the issue using the following versions:

• OpenVINO Model Server 2026.0.0.4d3933c5c
• OpenVINO backend 2026.0.0.0rc3

I used the optimized model pulled directly via:

/opt/openvino/ovms/bin/ovms --pull \
  --source_model "OpenVINO/Qwen3-30B-A3B-int4-ov" \
  --model_repository_path /opt/openvino/models \
  --model_name Qwen3-30B-A3B-int4-ov \
  --task text_generation

The RAM was increased from 32GB to 96GB (+32 GB swap file).

Again, it worked with:

/opt/openvino/ovms/bin/ovms \
  --model_repository_path /opt/openvino/models \
  --model_name Qwen3-30B-A3B-int4-ov \
  --task text_generation \
  --port 9001 \
  --rest_port 8000 \
  --target_device CPU

and failed (killed by Linux) with

/opt/openvino/ovms/bin/ovms \
  --model_repository_path /opt/openvino/models \
  --model_name Qwen3-30B-A3B-int4-ov \
  --task text_generation \
  --port 9001 \
  --rest_port 8000 \
  --target_device GPU

Observed Behavior:
During the GPU load attempt, system memory usage climbs until it fills the entire 96GB of RAM and exhausts nearly all of the 32GB swap file. I have attempted to mitigate this by applying various parameters to limit the KV cache size, but the memory leak/exhaustion persists regardless of these settings.

[Wan-Intel]

Hi @ikirsh

Thank you for sharing the method you used to export the model. We will investigate the issue and provide an update shortly.

[Wan-Intel]

Hi @ikirsh

I've exported the model and run the model via the following command on Ubuntu 24.04. The application works fine.

./ovms/bin/ovms --model_repository_path <path_to_dir>/models_repo --model_name OpenVINO/Qwen3-30B-A3B-int4-ov --task text_generation --port 9001 --rest_port 8000 --target_device GPU --log_level DEBUG

With GPU:

For your information, Kbuntu 25.10 is not validated by OpenVINO. The supported operating system for using OpenVINO GPU plugin are as follows:

• Windows 11, 64-bit
• Windows 10, 64-bit
• Ubuntu 24.04 long-term support (LTS), 64-bit
• Ubuntu 22.04 long-term support (LTS), 64-bit
• Ubuntu 20.04 long-term support (LTS), 64-bit
• CentOS 7
• Red Hat Enterprise Linux (RHEL) 8 and 9, 64-bit

For more information, please refer to https://docs.openvino.ai/2026/about-openvino/release-notes-openvino/system-requirements.html

[ikirsh]

Hi,

Kubuntu is basically Ubuntu + KDE desktop, so probably this is not the issue.

Have you run it on Intel Core Ultra 7 265 iGPU? If not, could you please run it with Ubuntu 24.04 but on that specific GPU? Memory allocation is totally different when using iGPU, and we have a memory allocation issue here.

Thank you.

[Wan-Intel]

Let me check with relevant team, and we will provide an update as soon as possible.

[jgespino]

@ikirsh I was able to reproduce the issue on Ubuntu 24.04 using an Intel Core Ultra 2 265K processor. I’ll review this with the development team and update you once I have more information.

Ref. 182695