| license: other | |
| license_name: lfm1.0 | |
| license_link: https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M/blob/main/LICENSE | |
| base_model: LiquidAI/LFM2.5-Encoder-350M | |
| library_name: mlx | |
| tags: | |
| - mlx | |
| - encoder | |
| - lfm2 | |
| - liquidai | |
| - fp32 | |
| # LFM2.5 Encoder 350M — MLX FP32 | |
| Native MLX conversion of `LiquidAI/LFM2.5-Encoder-350M`, pinned to revision | |
| `b886781f7c6f10ca9b7096e21b83e30a073c2f39`. | |
| The checkpoint is kept entirely in float32. This implementation preserves the | |
| encoder's full bidirectional attention and centered, non-causal short | |
| convolutions; it is not a causal `mlx-lm` generation model. | |
| ```python | |
| import mlx.core as mx | |
| from model import load_model | |
| model = load_model(".") | |
| input_ids = mx.array([[1, 123, 456, 7]]) | |
| attention_mask = mx.ones_like(input_ids) | |
| logits = model(input_ids, attention_mask) | |
| mx.eval(logits) | |
| ``` | |
| Use the included tokenizer files with `transformers.AutoTokenizer` or the | |
| `tokenizers` package. See `validation.json` for the FP32 tensor audit and | |
| PyTorch-to-MLX numerical parity measurements. | |