faster-clip 0.1.0

Dependency-light pure-numpy CPU inference for CLIP embeddings, inference-free SPLADE, and BERT fill-mask — a drop-in replacement for the transformers / sentence-transformers models in a RAG stack (no PyTorch, no ONNX).

faster-clip

Dependency-light, pure-numpy CPU inference for the models a multilingual RAG stack typically needs — with no PyTorch, no transformers, no sentence-transformers, no ONNX. Just numpy + tokenizers (+ safetensors / huggingface_hub for weights, pillow for images).

pip install faster-clip

It provides three drop-in replacements, each numerically matching the original HuggingFace pipeline:

Class	Replaces	Output
ClipModel	SentenceTransformer('clip-ViT-B-32') + clip-ViT-B-32-multilingual-v1	512-d CLIP image / text embeddings (cosine ≈ 1.0)
SparseEncoder	SparseEncoder('…inference-free-splade…')	{token_id: weight} sparse dicts (bit-identical)
FillMask	pipeline('fill-mask', 'bert-base-uncased' / 'bert-base-portuguese-cased')	top-k {token_str, sequence, score} (identical strings)

Model weights are downloaded from HuggingFace and cached on first use (the SPLADE weights, ~120 KB, ship inside the wheel).

CLIP embeddings

from faster_clip import ClipModel
from PIL import Image

clip = ClipModel()  # lazily downloads image / text weights on first use

text_emb = clip.encode_text(["um carro vermelho", "a red car"])   # (2, 512)
img_emb  = clip.encode_image(Image.open("photo.jpg"))             # (1, 512)

Image and multilingual text land in the same 512-d CLIP space, so you can do cross-modal search. (Normalize yourself if you need unit vectors.)

SPLADE sparse encoding (no sparse tensor materialized)

The model is an inference-free SPLADE: encoding is a pure token→weight lookup, so the result is just a dict of the active dimensions — no dense/sparse vector is ever built.

from faster_clip import SparseEncoder

sparse = SparseEncoder()
sparse.encode("banco de dados vetorial")
# -> {'15060': 1.09, '2951': 1.11, '2078': 1.04, ...}   # {token_id: weight}

encode, encode_query, and encode_document are equivalent (the model's query route is the default) and accept a string or a list of strings.

Fill-mask autocomplete

from faster_clip import FillMask

fm = FillMask()
fm.predict("the weather is very [MASK]", lang="en")
# -> [{'token_str': '.', 'sequence': 'the weather is very.', 'score': 0.30}, ...]
fm.predict("o brasil é um [MASK]", lang="pt")
# -> [{'token_str': 'lixo', 'sequence': 'o brasil é um lixo', ...}, ...]

lang is "en" (bert-base-uncased) or "pt" (neuralmind/bert-base-portuguese-cased).

Model cache

Weights are cached by huggingface_hub under ~/.cache/huggingface. To run offline, pre-download the models once (or set HF_HOME). CLIP vision/text and the BERT MLMs are fetched on first use of each capability.

License

MIT. Model weights belong to their respective authors (OpenAI / Microsoft / sentence-transformers / neuralmind), used under their original licenses.