Add GPT-OSS mixed-width CUDA QMoE recipe

[David Fan (jiafatom)]

Summary

• add a GPT-OSS-20B CUDA recipe with INT4 dense weights
• requantize QMoE gate/up projections to INT2 and down projections to INT4
• use symmetric blockwise expert quantization with block size 64

Depends on:

• Add mixed-width INT2/INT4 QMoE builder support onnxruntime-genai#2624
• Add packed INT2 and mixed-width CUDA QMoE decode onnxruntime#32761

Validation

• shell syntax passes with bash -n
• required export options are present exactly once
• git diff --check passes

[Copilot]

Copilot review overview

🟡 Changes recommended

The export command has invalid arguments, and required recipe metadata and executable handling are incomplete.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 2 · 1

Open (3)

• Use --conversion_device instead of unsupported execution_provider · New
• Pass extra_mb_options as one comma-separated argument · New
• Add required info.yml metadata for recipe discovery · New

What changed in this PR

Adds a GPT-OSS-20B CUDA recipe with INT4 dense weights and mixed INT2/INT4 QMoE expert quantization.

Changes:

• Documents export prerequisites and inference usage.
• Adds the Olive export script with block size 64 and projection-specific quantization.

File	Description
gpt-oss-20b/​int4_cuda_int2_int4_qmoe/​README.md	Documents the recipe, export, and execution.
gpt-oss-20b/​int4_cuda_int2_int4_qmoe/​gpt-oss-20b.sh	Defines the mixed-width CUDA export command.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Xiaoyu Z (xiaoyu-work)]

Looks like this script is just a CLI, can we just move the CLI command to this README directly?

[Xiaoyu Z (xiaoyu-work)]

what is this used for?