Add option to enable hardware acceleration for Adapter models

[kziovas]

🧩 Feature Request: Add option to enable hardware acceleration for Adapter models

Summary:
Enable users to specify the hardware device (cuda, cpu, mps, npu, etc.) used for computation when initializing adapter-based models such as TextEmbedding.

---

Current Behavior

Currently, the TextEmbedding adapter in vecs.adapters.text_embedding initializes a SentenceTransformer model without any control over which device is used for inference:

self.model = ST(model, use_auth_token=use_auth_token)

This defaults to automatic device detection inside sentence-transformers, but users cannot explicitly override it or ensure that hardware acceleration (e.g., GPU or Apple MPS) is utilized when available.

---

Proposed Improvement

Add a new optional argument to the TextEmbedding class:

def __init__(
    *,
    model: TextEmbeddingModel,
    batch_size: int = 8,
    use_auth_token: str = None,
    device: str | None = None,
):

and use it in model initialization:

self.model = ST(model, use_auth_token=use_auth_token, device=device)

This aligns with the SentenceTransformer constructor signature:

SentenceTransformer(model_name_or_path, device: Optional[str] = None, ...)

---

Benefits

• Improved performance: Enables GPU, MPS, or other accelerator usage for large-scale text embedding generation.
• Better control: Allows deployment in heterogeneous environments (CPU-only, multi-GPU, Apple Silicon, etc.).
• Parity with sentence-transformers API: Matches user expectations and improves developer ergonomics.

---

Example Usage

from vecs.adapters import TextEmbedding

embedding_adapter = TextEmbedding(
    model="sentence-transformers/all-MiniLM-L6-v2",
    device="cuda",  # or "mps", "cpu"
)

---

Backward Compatibility

• Default behavior remains unchanged (device=None → automatic device detection).
• No breaking API changes.

---

[Hallidayo]

Hi @kziovas - Thanks for opening this one. I'm going to move across as a enhancement/feature request. Thanks again for opening!