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 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151

use std::fs;

use candle_core::DType;
use candle_core::Device;

use super::SparseTextEmbedding;
use super::build;
use crate::device;
use crate::embed::builder::reject_unsupported_dtype;
use crate::embed::builder::resolve_dtype;
use crate::embed::fit;
use crate::error::CanddingError;
use crate::error::Result;
use crate::hub;
use crate::hub::ModelSource;
use crate::loader;
use crate::models;
use crate::registry;
use crate::registry::Catalog;
use crate::registry::sparse::SparseDescriptor;
use crate::tokenizer::TextTokenizer;
use crate::tokenizer::read_pad_token;
use crate::weights;

/// Eight keeps SPLADE's transient `(batch, 64, vocab)` logits chunk (59.61 MiB) near the dense pipeline's peak
/// hidden states at its default of 32 (48 MiB); it is a comparability choice, not a memory ceiling.
const DEFAULT_BATCH_SIZE: usize = 8;

pub struct SparseTextEmbeddingBuilder {
  spec: String,
  revision: String,
  device: Option<Device>,
  dtype: Option<DType>,
  max_length: Option<usize>,
  batch_size: usize,
  descriptor: Option<SparseDescriptor>,
}

impl SparseTextEmbeddingBuilder {
  pub(crate) fn new(spec: impl Into<String>) -> Self {
    Self {
      spec: spec.into(),
      revision: "main".to_string(),
      device: None,
      dtype: None,
      max_length: None,
      batch_size: DEFAULT_BATCH_SIZE,
      descriptor: None,
    }
  }

  pub fn device(mut self, device: Device) -> Self {
    self.device = Some(device);
    self
  }

  pub fn dtype(mut self, dtype: DType) -> Self {
    self.dtype = Some(dtype);
    self
  }

  /// Capped at the model's `max_position_embeddings` when building.
  pub fn max_length(mut self, max_length: usize) -> Self {
    self.max_length = Some(max_length);
    self
  }

  pub fn batch_size(mut self, batch_size: usize) -> Self {
    self.batch_size = batch_size.max(1);
    self
  }

  pub fn revision(mut self, revision: impl Into<String>) -> Self {
    self.revision = revision.into();
    self
  }

  /// Supplies the descriptor for a sparse head outside the registry, or replaces a registered one; without it an
  /// unregistered id is refused, since no file names a checkpoint's head or vocabulary size.
  pub fn descriptor(mut self, descriptor: SparseDescriptor) -> Self {
    self.descriptor = Some(descriptor);
    self
  }

  pub fn build(self) -> Result<SparseTextEmbedding> {
    let overridden = self.descriptor.is_some();
    let descriptor = match self.descriptor {
      Some(descriptor) => descriptor,
      None => {
        registry::sparse::lookup(&self.spec).ok_or_else(|| CanddingError::UnregisteredModel {
          id: self.spec.clone(),
          catalog: Catalog::Sparse,
        })?
      }
    };
    // A registry hit carries the entry's canonical spelling; an override keeps the caller's spec.
    let spec = if overridden {
      self.spec.as_str()
    } else {
      descriptor.id.as_str()
    };
    build::reject_computed_entry(&descriptor)?;
    let device = match self.device {
      Some(device) => device,
      None => device::auto()?,
    };
    let dtype = resolve_dtype(&self.spec, self.dtype, DType::F32, &device)?;
    reject_unsupported_dtype(&self.spec, dtype, descriptor.dtype_backends, &device)?;
    let source = ModelSource::parse(spec, &self.revision);
    let size_gb = descriptor.size_gb;
    fit::check_download(&self.spec, &device, &source, size_gb, DType::F32, dtype)?;
    let dir = hub::resolve(&source, descriptor.weights)?;
    let files = loader::discover(&dir, loader::Pipeline::Sparse)?;
    let config_json = fs::read_to_string(&files.config)?;
    let minicoil_side = build::resolve_minicoil_side(&descriptor)?;
    let weight_files = build::weight_files(&descriptor, &files, minicoil_side.as_ref(), dtype);
    fit::check_files(&self.spec, &device, &weight_files)?;
    tracing::info!(model = %self.spec, device = device::describe(&device), dtype = dtype.as_str(), "loading sparse model");
    let vb = unsafe { weights::var_builder(&files.weights, dtype, &device)? };
    let minicoil_files = minicoil_side.as_ref().map(|side| models::MiniCoilFiles {
      matrix: &side.matrix,
      vocab: &side.vocab,
      stopwords: &side.stopwords,
    });
    let model = models::load_sparse(
      descriptor.head,
      &config_json,
      vb,
      &files.tokenizer,
      files.sparse_linear.as_deref(),
      minicoil_files,
    )?;
    let max_length = self
      .max_length
      .unwrap_or(descriptor.default_max_length)
      .min(model.max_position_embeddings());
    let pad = read_pad_token(&config_json, files.tokenizer_config.as_deref())?;
    let tokenizer = TextTokenizer::from_file(&files.tokenizer, max_length, pad)?;
    Ok(SparseTextEmbedding::new(
      model,
      tokenizer,
      descriptor,
      device,
      dtype,
      self.batch_size,
    ))
  }
}

#[cfg(test)]
mod tests;