doc: add Windows ML CLI build flow section to model conversion#9874
Conversation
|
|
||
| The behavior on first entry depends on how the model was added to the project: | ||
|
|
||
| - **Built-in models** (the curated entries that already have a model card) ship with a prepared configuration. The Build Flow card opens directly in the **Configured** state — no auto-configuration runs. Select **Edit Config** to review the prepared recipe, then select **Build**. |
There was a problem hiding this comment.
Built-in models already include validated configurations for WinML CLI workflows.
These curated models ship with prepared Build Config files optimized for Windows ML EPs.
| The behavior on first entry depends on how the model was added to the project: | ||
|
|
||
| - **Built-in models** (the curated entries that already have a model card) ship with a prepared configuration. The Build Flow card opens directly in the **Configured** state — no auto-configuration runs. Select **Edit Config** to review the prepared recipe, then select **Build**. | ||
| - **Hugging Face models added by ID** are downloaded and analyzed automatically on first entry. The card transitions through these states: |
There was a problem hiding this comment.
Step 1: Generate the build config
Hugging Face models added by ID are automatically processed on first entry.
WinML CLI queries Hugging Face, auto-detects the task and model type, and generates Build Config JSON files automatically.
During onboarding, WinML CLI generates three configuration variants:
config-noquant.json
config-w8a16.json
config-w8a8.json
The primary difference between them is the quantization strategy:
No Quant — full precision model
W8A16 — 8-bit weights with 16-bit activations
W8A8 — 8-bit weights with 8-bit activations for more aggressive compression and performance optimization
Step 2: Customize the Config
You can customize the workflow before running the build pipeline.
Typical customization areas include: Task type, compile target, precision details.
By default, Compile is set to null.
You can customize the Compile with target EP, e.g. xxxx.
Step 3: Run the build
This build runs all four pipeline stages in sequence:
- Export
- Optimize
- Quantize
- Compile
The workflow reads the settings recorded in config*.json.
After the build step, WinML CLI automatically generates a declarative build_config.json file that defines how the workflow runs end-to-end. Developers can inspect and customize it through View Config.
This declarative configuration model makes it easy to integrate WinML CLI into CI/CD pipelines with reproducible and portable build workflows.
WinML CLI also generates an analyze report, which can be viewed through View Analyze.
The analyze results provide detailed model compatibility insights, including:
supported operators, partially supported operators, unsupported operators for Windows ML EPs.
During analysis, WinML CLI automatically inspects the ONNX graph, detects optimization patterns, and generates recommended Windows ML optimization workflows.
…-in config, failure causes)
ntrogh
left a comment
ntrogh
left a comment
There was a problem hiding this comment.
@ssss141414 Let me know when this can be merged
Sure. In discussion with our PM. Will inform you when ready. Thank you! |
…onfig, no download)
5 participants
No description provided.