GPT-OSS model support (MXFP4 version)

[ravi9]

Add support for GPT-OSS MXFP4 quantized models including correct tensor handling and validation.

[arunthakur009]

I looked into the OpenVINO backend (ggml-openvino/ggml-quants.cpp), and it seems MXFP4 isn’t currently handled in the extraction pipeline.

While MXFP4 exists at the GGML level (quantize_row_mxfp4_ref / dequantize_row_mxfp4), the backend only supports types like Q4_0, Q4_1, Q4_K, Q5_K, Q6_K, and Q8_0.

There’s no extract_mxfp4_data implementation, and extract_quantized_weights() doesn’t include a case for MXFP4,so MXFP4 tensors would currently fail in the extract_quantized_weights() path during backend conversion.

It looks like the missing part is integrating MXFP4 into the OpenVINO extraction + weight construction pipeline rather than the quantization itself.

I’d like to work on this starting by tracing how Q4_* types are mapped and extending that path for MXFP4. Let me know if this direction makes sense.

[ravi9]

@zhaixuejun1993 @wine99 : Could you review the following:

ggml-org#21196

https://github.com/lucaspirola/llama.cpp/tree/fix/openvino-gpt-oss-inference

[5ghzx]

Wow, this is pretty cool, when will this be released?