candding 0.1.0

Pure candle embeddings for Rust with no ONNX: dense, sparse, multi-vector and rerank models today, image models on the roadmap
Documentation
use crate::models::MultiVectorHead;
use crate::multi_vector::MaskedRowFill;
use crate::multi_vector::ScoreReduction;
use crate::registry::DtypeSupport;
use crate::registry::Family;
use crate::registry::Support;
use crate::registry::WeightsFormat;
use crate::registry::multi_vector::entry::MultiVectorEntry;
use crate::registry::t4::T4_BF16_UNTESTED;

/// `colbert-ir/colbertv2.0`: a bias-free root `Linear(768, 128)` over `bert-base-uncased`, with the reference's
/// fallback markers `[unused0] ` and `[unused1] `; documents cut at 180 tokens, queries padded to 32 rows.
pub(super) const COLBERTV2_MULTI_VECTOR: MultiVectorEntry = MultiVectorEntry {
  id: "colbert-ir/colbertv2.0",
  family: Family::Bert,
  head: MultiVectorHead::StanfordColbert,
  default_max_length: 180,
  keeps_cls_row: true,
  projection_in: 768,
  projection_out: 128,
  projection_bias: false,
  normalize: true,
  score_reduction: ScoreReduction::Sum,
  query_marker: Some("[unused0] "),
  document_marker: Some("[unused1] "),
  query_length: Some(32),
  attends_to_mask_tokens: false,
  masked_document_row_fill: MaskedRowFill::DtypeMinimum,
  position_scheme: None,
  cpu: Support::Verified,
  metal: Support::Verified,
  cuda: Support::Verified,
  dtypes: DtypeSupport {
    f32: Support::Verified,
    f16: Support::Verified,
    bf16: Support::Verified,
  },
  license: "MIT",
  size_gb: 0.44,
  phase: 4,
  weights: WeightsFormat::Safetensors,
  gated: false,
  accel_only_dtypes: &[],
  cell_exceptions: T4_BF16_UNTESTED,
};