electroglyph/snowflake-arctic-embed-m-v2.0-ONNX-uint8

snowflake-arctic-embed-m-v2.0-ONNX-uint8

This is a version of this model which outputs a qdrant compatible uint8 tensor.

It's very close to the accuracy of the full precision f32 ONNX model with f32 output.

This is an upgraded version of https://huggingface.co/electroglyph/snowflake2_m_uint8 with better accuracy.

Quantization method

For calibration data I used my own multilingual dataset of around 1.5m tokens: https://github.com/electroglyph/dataset_build

I ran all compatible tokens through the model and logged the highest/lowest values seen. I found a range of: -0.15288913249969482 to 0.1472320258617401

I hacked on the sentence_embedding output of the ONNX model and added QuantizeLinear node based on the range of -0.15288913249969482 to 0.15288913249969482 to keep it symmetric.

Benchmarks

I included results for one of my other models: https://huggingface.co/electroglyph/embeddinggemma-300m-ONNX-uint8

The results for this model are quite close to the full precision ONNX model with f32 output.

Example Benchmark Code

import mteb
from mteb.encoder_interface import PromptType
import numpy as np
import onnxruntime as rt
from transformers import AutoTokenizer

class CustomModel:
    def __init__(self) -> None:
        self.tokenizer = AutoTokenizer.from_pretrained("C:/LLM/snowflake-arctic-embed-m-v2.0")
        self.session = rt.InferenceSession("snowflake-arctic-embed-m-v2.0-uint8.onnx", providers=["CPUExecutionProvider"])
        self.scale = 0.15288913249969482 / 127.0
        
    def dequantize(self, quantized: list | np.ndarray, scale: float) -> np.ndarray:
        quantized = np.array(quantized)
        dequant = (quantized.astype(np.float32) - 128) * scale
        if dequant.ndim == 3 and dequant.shape[0] == 1:
            return np.squeeze(dequant, axis=0)
        return dequant

    def encode(
        self,
        sentences: list[str],
        task_name: str,
        prompt_type: PromptType | None = None,
        **kwargs,
    ) -> np.ndarray:
        inputs = self.tokenizer(sentences, padding=True, truncation=True, return_tensors="np", max_length=8192)
        q = self.session.run(["sentence_embedding"], dict(inputs))
        return self.dequantize(q, self.scale)


model = CustomModel()
benchmark = mteb.get_benchmark("NanoBEIR")
evaluation = mteb.MTEB(tasks=benchmark)
results = evaluation.run(model, corpus_chunk_size=4)
for r in results:
    print(r)

Example FastEmbed Usage

from fastembed import TextEmbedding
from fastembed.common.model_description import PoolingType, ModelSource

TextEmbedding.add_custom_model(
    model="snowflake-arctic-embed-m-v2.0-ONNX-uint8",
    pooling=PoolingType.DISABLED,
    normalization=False,
    sources=ModelSource(hf="electroglyph/snowflake-arctic-embed-m-v2.0-ONNX-uint8"),
    dim=768,
    model_file="onnx/model.onnx",
)

model = TextEmbedding(model_name="snowflake-arctic-embed-m-v2.0-ONNX-uint8")
embeddings = list(model.embed("test"))
print(embeddings)

License

Arctic is licensed under the Apache-2. The released models can be used for commercial purposes free of charge.