2025.4 GPT-OSS 20B failed to start

[roomer-s]

Followed instructions for Windows on
https://docs.openvino.ai/2025/model-server/ovms_demos_continuous_batching_agent.html
for GPT-OSS 20B.

After invoking:
ovms.exe --rest_port 8000 --source_model openai/gpt-oss-20b --model_repository_path models --tool_parser gptoss --reasoning_parser gptoss --target_device GPU --task text_generation --pipeline_type LM

it failed with
[2025-12-15 09:30:12.280][20660][serving][error][servable_initializer.cpp:104] Error during llm node initialization for models_path: C:\OpenVino\models\openai\gpt-oss-20b\./ exception: Exception from src\inference\src\cpp\core.cpp:117: Exception from src\inference\src\dev\plugin.cpp:53: Exception from src\core\src\pass\graph_rewrite.cpp:298: [ConvertFullyConnectedToFullyConnectedCompressed] END: node: gpu_opset::FullyConnected __module.model.layers.0.mlp.experts/aten::bmm/MatMul (opset1::Reshape __module.model.layers.0.mlp.experts/aten::view/Reshape[0]:f16[32,?,2880], opset1::Reshape Reshape_215611[0]:f16[32,5760,2880], gpu_opset::Placeholder Placeholder_113014[0]:dynamic[]) -> (f16[32,?,5760]) CALLBACK HAS THROWN: Check 'other_size == new_size' failed at src\core\src\op\constant.cpp:347: ov::Shape size 1061683200 is not equal to 16588800

[dkalinowski]

@roomer-s Running the command you provided alone, without prior export models causes OVMS to pull & export the model at runtime. And without missing --weight-format int4 parameter it exports the model with default int8 precision which is not supported yet.

To overcome, export the model first: https://docs.openvino.ai/2025/model-server/ovms_demos_continuous_batching_agent.html#export-llm-model

And then start using the command you provided. Or simply add --weight-format int4 to the ovms.exe command, it should work as well. I will update documentation to enforce int4 export.

• to get the latest accuracy fixes for GPT-OSS you need to either build OVMS from releases/2025/4 branch, or wait for 2025.4.1 release which is happening shortly.

[dkalinowski]

@roomer-s release 2025.4.1 is here: https://github.com/openvinotoolkit/model_server/releases/tag/v2025.4.1
let me know if you face any more issues