Add Qwen3 Embedding recipes (0.6B and 8B)

[natke]

Summary

Add Olive recipes for Qwen3-Embedding-0.6B and Qwen3-Embedding-8B embedding models, targeting CPU, CUDA, and WebGPU execution providers.

Models

• Qwen/Qwen3-Embedding-0.6B — 0.6B params
• Qwen/Qwen3-Embedding-8B — 8B params

Recipes (6 total)

• Qwen-Qwen3-Embedding-0.6B/{cpu,cuda,webgpu}
• Qwen-Qwen3-Embedding-8B/{cpu,cuda,webgpu}

Each recipe includes:

• INT4 precision with include_hidden_states=1 for embedding extraction
• ModelBuilder pass for ONNX/GenAI export
• MTEBEvaluator configured with STS17 benchmark
• README with setup and evaluation instructions
• info.yaml, requirements.txt, LICENSE

Dependencies

• Requires the MTEBEvaluator from Add MTEBEvaluator for embedding model evaluation Olive#2409 (merged)
• Requires device mapping and pooling fixes from Fix MTEBEvaluator: device mapping, padding-free inference, last-token pooling, L2 normalization Olive#2415 (merged)
• Manually works around tie_word_embeddings` not restored in model config after RTN/GPTQ quantization with retied embeddings Olive#2424

[Copilot]

Pull request overview

Adds new Olive recipe folders for the Qwen3 embedding model variants (0.6B, 8B) across CPU/CUDA/WebGPU, including configs, environment requirements, and documentation to support FP32 export with hidden states enabled and MTEB (STS17) evaluation wiring.

Changes:

• Added CPU/CUDA/WebGPU Olive configs using ModelBuilder (FP32) with include_hidden_states=1 for Qwen3 embedding models.
• Added per-recipe requirements.txt, info.yaml, and auto-generated README.md files.
• Added per-model Apache 2.0 LICENSE files.

Reviewed changes

Copilot reviewed 26 out of 26 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
Qwen-Qwen3-Embedding-8B/LICENSE	Adds license file for the 8B embedding model recipe set.
Qwen-Qwen3-Embedding-8B/cpu/requirements.txt	Python deps for running the CPU recipe and MTEB evaluation.
Qwen-Qwen3-Embedding-8B/cpu/README.md	CPU recipe usage + leaderboard references.
Qwen-Qwen3-Embedding-8B/cpu/Qwen-Qwen3-Embedding-8B_cpu_fp32.json	Olive config: HF model → ModelBuilder FP32 + MTEBEvaluator(STS17).
Qwen-Qwen3-Embedding-8B/cpu/info.yaml	Registers the CPU recipe metadata (device/EP/file).
Qwen-Qwen3-Embedding-8B/cuda/requirements.txt	Python deps for running the CUDA recipe and MTEB evaluation.
Qwen-Qwen3-Embedding-8B/cuda/README.md	CUDA recipe usage + leaderboard references.
Qwen-Qwen3-Embedding-8B/cuda/Qwen-Qwen3-Embedding-8B_cuda_fp32.json	Olive config: HF model → ModelBuilder FP32 + MTEBEvaluator(STS17) on CUDA EP.
Qwen-Qwen3-Embedding-8B/cuda/info.yaml	Registers the CUDA recipe metadata (device/EP/file).
Qwen-Qwen3-Embedding-8B/webgpu/requirements.txt	Python deps for running the WebGPU recipe and MTEB evaluation.
Qwen-Qwen3-Embedding-8B/webgpu/README.md	WebGPU recipe usage + leaderboard references.
Qwen-Qwen3-Embedding-8B/webgpu/Qwen-Qwen3-Embedding-8B_webgpu_fp32.json	Olive config: HF model → ModelBuilder FP32 + MTEBEvaluator(STS17) on WebGPU EP.
Qwen-Qwen3-Embedding-8B/webgpu/info.yaml	Registers the WebGPU recipe metadata (device/EP/file).
Qwen-Qwen3-Embedding-0.6B/LICENSE	Adds license file for the 0.6B embedding model recipe set.
Qwen-Qwen3-Embedding-0.6B/cpu/requirements.txt	Python deps for running the CPU recipe and MTEB evaluation.
Qwen-Qwen3-Embedding-0.6B/cpu/README.md	CPU recipe usage + leaderboard references + STS17 eval results section.
Qwen-Qwen3-Embedding-0.6B/cpu/Qwen-Qwen3-Embedding-0.6B_cpu_fp32.json	Olive config: HF model → ModelBuilder FP32 + MTEBEvaluator(STS17).
Qwen-Qwen3-Embedding-0.6B/cpu/info.yaml	Registers the CPU recipe metadata (device/EP/file).
Qwen-Qwen3-Embedding-0.6B/cuda/requirements.txt	Python deps for running the CUDA recipe and MTEB evaluation.
Qwen-Qwen3-Embedding-0.6B/cuda/README.md	CUDA recipe usage + leaderboard references.
Qwen-Qwen3-Embedding-0.6B/cuda/Qwen-Qwen3-Embedding-0.6B_cuda_fp32.json	Olive config: HF model → ModelBuilder FP32 + MTEBEvaluator(STS17) on CUDA EP.
Qwen-Qwen3-Embedding-0.6B/cuda/info.yaml	Registers the CUDA recipe metadata (device/EP/file).
Qwen-Qwen3-Embedding-0.6B/webgpu/requirements.txt	Python deps for running the WebGPU recipe and MTEB evaluation.
Qwen-Qwen3-Embedding-0.6B/webgpu/README.md	WebGPU recipe usage + leaderboard references.
Qwen-Qwen3-Embedding-0.6B/webgpu/Qwen-Qwen3-Embedding-0.6B_webgpu_fp32.json	Olive config: HF model → ModelBuilder FP32 + MTEBEvaluator(STS17) on WebGPU EP.
Qwen-Qwen3-Embedding-0.6B/webgpu/info.yaml	Registers the WebGPU recipe metadata (device/EP/file).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[shaahji]

/azp run

[hanbitmyths]

It would be good to remove fp32 cuda and webgpu recipes and add fp16 cuda and webgpu recipes.

[hanbitmyths]

Did you add int4 quantization at only 8B model intentionally?

[hanbitmyths]

Qwen3-Embedding-0.6B model supports tie_word_embeddings. Did you check if an ONNX model correctly ties embedding?