[Whisper][NPU] Default optimum export of whisper model does not create KV cache decoder

[soumendukrg]

The latest optimum-cli command to download and convert whisper model to OV as given in the README does not create the decoder model with KV cache.

optimum-cli export openvino --trust-remote-code --model openai/whisper-base whisper-base

Since NPU inference pipeline needs the KV cache decoder model, running the sample throws out this error.

The following change to the exporting command can solve this issue as it generates the KV cache decoder also

optimum-cli export openvino --model openai/whisper-base whisper-base --task automatic-speech-recognition-with-past --disable-stateful

NOTE
Tried without the --disable-stateful flag, but did not produce the KV cache decoder.

As mentioned in this #1726 , --trust-remote-code is also not needed.

Environment:

• transformers 4.46.2
• openvino 2025.0.0
• openvino-genai 2025.0.0.0
• openvino-telemetry 2024.1.0
• openvino-tokenizers 2025.0.0.0
• optimum 1.24.0
• optimum-intel 1.23.0.dev0+dd4fe68

[eaidova]

@dmatveev @TolyaTalamanov probably need to provide command line for NPU model conversion in docs until stateful whisper is not supported. Could you please handle this?

[TolyaTalamanov]

@soumendukrg This information is already documented in the provided documentation: https://docs.openvino.ai/nightly/openvino-workflow-generative/inference-with-genai/inference-with-genai-on-npu.html

--disable-stateful should be provided during conversion:

optimum-cli export openvino --trust-remote-code --model openai/whisper-tiny whisper-tiny --disable-stateful

Once this PR gets into nightly package Whisper pipeline on NPU will be able to work with both 2 and 3 model configurations.