MAX_PROMPT_LEN seems to not be supported by VLMPipeline

[johncohen411]

Hi,

I followed these guides to export microsoft/Phi-3.5-vision-instruct to the openvino format so I can run it on my NPU:

• https://docs.openvino.ai/2025/model-server/ovms_demos_vlm_npu.html?utm_source=chatgpt.com#references
• https://docs.openvino.ai/2025/openvino-workflow-generative/inference-with-genai/inference-with-genai-on-npu.html

My goal is to use VLMPipeline on the NPU with MAX_PROMPT_LEN set to higher than 1024 (for example 2048).
However, when I pass the MAX_PROMPT_LEN parameter to VLMPipeline, same as @asteroidcow in issue #33385 , I am getting "NotFound: Unsupported property MAX_PROMPT_LEN by CPU plugin.". It seems like for some reason VLMPipeline does not support MAX_PROMPT_LEN even though LLMPipeline does.

I would appreciate some help.
(openvino version - 2025.4.0)

Thanks.

[krgkaushik]

Hi @johncohen411

This error happens because VLMPipeline attempts to apply MAX_PROMPT_LEN to every model in the pipeline, including the Visual Encoder (which usually runs on CPU). The CPU plugin rejects this property, causing the crash.

To fix this, you need to scope the configuration specifically to the NPU using DEVICE_PROPERTIES. This ensures the property is only sent to the NPU plugin and ignored by the CPU.

Try initializing your pipeline like this:
import openvino_genai as ov_genai
model_path = "path/to/model"
device = "NPU"

Scope the config so it only applies to the NPU device
config = {
"DEVICE_PROPERTIES": {
"NPU": {
"MAX_PROMPT_LEN": 2048
}
}
}

pipe = ov_genai.VLMPipeline(model_path, device, **config)
This should bypass the CPU plugin validation error. Let me know if that works for you\!

[johncohen411]

Thank you @krgkaushik .

I followed your advice and got this error:
Exception from C:\Jenkins\workspace\private-ci\ie\build-windows-vs2022\b\repos\openvino.genai\src\python\py_utils.cpp:255: Property "DEVICE_PROPERTIES" got unsupported type.