FIX #245 - Set required DirectML session options (memory_pattern + sequential execution)

[markgandolfo]

proposed fix for #245

DirectML requires two ONNX Runtime session options that fastembed wasn't setting:

• enable_mem_pattern must be false
• execution_mode must be sequential (not parallel)

Without these, DirectML can OOM on operations that would otherwise fit in VRAM. This became visible after candle 0.10 changed GPU memory allocation patterns.

on the directml page

If creating the onnxruntime InferenceSession object directly, you must set the appropriate fields on the onnxruntime::SessionOptions struct. Specifically, execution_mode must be set to ExecutionMode::ORT_SEQUENTIAL, and enable_mem_pattern must be false.

Changes:

• Added directml = ["ort/directml"] feature flag in Cargo.toml
• Patched all 4 session builder sites (try_new and try_new_from_user_defined in both text_embedding/impl.rs and image_embedding/impl.rs) to detect DirectML at runtime and apply the required settings
• Uses #[cfg(feature = "directml")] guards so it compiles on all platforms
• No change for CPU, CUDA, or any other execution provider.

[Anush008]

LGTM!

Thanks for taking the time to contribute @markgandolfo.

[github-actions]

🎉 This PR is included in version 5.13.2 🎉

The release is available on:

• GitHub release
• v5.13.2

Your semantic-release bot 📦🚀