Fix Qwen3.5-0.8B WebGPU optimization config

[xiaofeihan1]

No description provided.

[xiaofeihan1]

Hi @apsonawane PTAL, thanks!

[Copilot]

Pull request overview

This PR updates the Qwen3.5-0.8B “builtin” WebGPU Olive workflow configs to explicitly target a WebGPU-enabled local system, and adjusts the optimization script’s genai_config patching.

Changes:

• Add systems.local_system (WebGpuExecutionProvider) and set "target": "local_system" in the WebGPU embedding/text/vision Olive JSON configs.
• Update optimize.py to use vision_session_options when setting the decoder session_options.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
Qwen-Qwen3.5-0.8B/builtin/webgpu/vision.json	Adds WebGPU system definition and selects it via target.
Qwen-Qwen3.5-0.8B/builtin/webgpu/text.json	Adds WebGPU system definition and selects it via target.
Qwen-Qwen3.5-0.8B/builtin/webgpu/embedding.json	Adds WebGPU system definition and selects it via target.
Qwen-Qwen3.5-0.8B/builtin/optimize.py	Changes which session options are written into genai_config.json for the decoder.

---

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

[Copilot]

decoder.session_options is being set to vision_session_options. For device == "gpu", vision_provider_options explicitly disables CUDA graph capture, so this change unintentionally applies the vision/embedding workaround to the decoder too (and also hides the fact that session_options is otherwise undefined). Create a separate session_options based on provider_options and assign that to the decoder, keeping vision_session_options only for vision/embedding.

[apsonawane]

Why is this change required only for qwen3.5-0.8b model?

[xiaofeihan1]

Why is this change required only for qwen3.5-0.8b model?

Thanks @apsonawane . Applied for other Qwen3.5 models. I have a follow-up question:
Why Qwen3.5-2B use k_quant_mixed but Qwen3.5-4B use rtn?

[apsonawane]

@xiaofeihan1 We tried few combinations of quantization and selected the one that gives us best accuracy and performance. That's why there is a difference.
2B and 0.8B are very sensitive to int4 quant that's why we kept is mixed while 4B and 9B perform better on int4 and gives better performance