MLX Backend (Apple Silicon)¶

OpenMed v1.0.0 introduces native Apple Silicon acceleration via Apple MLX. On Macs with M1/M2/M3/M4 chips, NER and PII inference runs directly on the GPU — up to 10x faster than CPU-only PyTorch.

Installation¶

pip install "openmed[mlx]"

This installs mlx, huggingface-hub, tokenizers, and safetensors.

Quick Start¶

from openmed import analyze_text
from openmed.core.config import OpenMedConfig

# MLX is auto-detected on Apple Silicon — no config needed
result = analyze_text(
    "Patient John Doe, DOB 1990-05-15, SSN 123-45-6789",
    model_name="pii_detection",
)
print(result.entities)

To force a specific backend:

config = OpenMedConfig(backend="mlx")   # Force MLX
config = OpenMedConfig(backend="hf")    # Force HuggingFace/PyTorch
config = OpenMedConfig(backend=None)    # Auto-detect (default)

How It Works¶

1. Auto-detection: On Apple Silicon Macs with mlx installed, OpenMed automatically selects the MLX backend.
2. On-the-fly conversion: The first time you use a model with MLX, it's automatically converted from HuggingFace format and cached in ~/.cache/openmed/mlx/.
3. Identical output: MLX produces the same entity format as the HuggingFace backend — all downstream processing (entity merging, quality gates, PII detection) works identically.

Model Conversion¶

Automatic (recommended)¶

Models are converted automatically on first use. No manual step needed.

Manual conversion¶

For pre-converting models (e.g., for offline deployment):

python -m openmed.mlx.convert \
    --model OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1 \
    --output ./mlx-models/pii-small

With 8-bit quantization (reduces model size by ~4x):

python -m openmed.mlx.convert \
    --model OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1 \
    --output ./mlx-models/pii-small-q8 \
    --quantize 8

The output directory contains: - weights.npz — Model weights in MLX/NumPy format - config.json — Model architecture configuration - id2label.json — Label ID to entity name mapping

Supported Models¶

Currently, the MLX backend supports BERT-based token classification models:

DeBERTa, ModernBERT, and ELECTRA architectures will be added in future releases.

Fallback Behavior¶

If MLX is not available (non-Apple hardware, or mlx not installed), OpenMed automatically falls back to the HuggingFace/PyTorch backend. No code changes required.

Conversion Without MLX¶

You can convert models on any machine (even Linux) — the converter falls back to NumPy format:

# On Linux CI — produces NumPy .npz (no MLX needed for conversion)
python -m openmed.mlx.convert \
    --model OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1 \
    --output ./mlx-models/pii-small

The NumPy .npz files are fully compatible with the MLX backend.