furnace/
tensor.rs

1//! Tensors as the server sees them: named inputs and outputs of a model, with a datatype and a
2//! shape, held in host memory between the wire formats and the engine that runs the model.
3//!
4//! Every request reaches a model as [`HostTensor`]s, one per input, batch first. The flat
5//! `Vec<f32>` entry points of `Model` (`predict*`) wrap their buffer as one such tensor without
6//! copying it.
7
8use serde::{Deserialize, Serialize};
9
10use crate::error::{ModelError, Result};
11
12/// Element types a model's inputs and outputs can have
13#[derive(Debug, Clone, Copy, PartialEq, Eq)]
14pub enum DataType {
15    Fp32,
16    Int64,
17    Int32,
18}
19
20impl DataType {
21    /// The Open Inference Protocol name: `FP32`, `INT64`, `INT32`
22    pub fn v2_name(self) -> &'static str {
23        match self {
24            Self::Fp32 => "FP32",
25            Self::Int64 => "INT64",
26            Self::Int32 => "INT32",
27        }
28    }
29
30    /// The type an Open Inference Protocol name stands for; `None` for one the server does not
31    /// take
32    pub fn from_v2_name(name: &str) -> Option<Self> {
33        match name {
34            "FP32" => Some(Self::Fp32),
35            "INT64" => Some(Self::Int64),
36            "INT32" => Some(Self::Int32),
37            _ => None,
38        }
39    }
40
41    /// Bytes per element in the raw (binary) format
42    pub fn size(self) -> usize {
43        match self {
44            Self::Fp32 | Self::Int32 => 4,
45            Self::Int64 => 8,
46        }
47    }
48
49    /// Whether elements are integers (held as `i64` in [`Values::I64`])
50    pub fn is_int(self) -> bool {
51        matches!(self, Self::Int64 | Self::Int32)
52    }
53}
54
55/// A model input or output: its name, element type and per-item shape (without the batch
56/// dimension; `None` where the graph leaves a dimension open)
57#[derive(Debug, Clone, PartialEq)]
58pub struct IoSpec {
59    pub name: String,
60    pub dtype: DataType,
61    pub shape: Vec<Option<usize>>,
62}
63
64impl IoSpec {
65    /// A tensor of `dtype` with per-item `shape` (`None` for a dimension each request fixes)
66    pub fn new(name: &str, dtype: DataType, shape: &[Option<usize>]) -> Self {
67        Self {
68            name: name.to_string(),
69            dtype,
70            shape: shape.to_vec(),
71        }
72    }
73
74    /// An `f32` tensor with a fixed per-item shape
75    pub fn fp32(name: &str, shape: &[usize]) -> Self {
76        Self {
77            name: name.to_string(),
78            dtype: DataType::Fp32,
79            shape: shape.iter().map(|&d| Some(d)).collect(),
80        }
81    }
82
83    /// The per-item shape if every dimension is fixed
84    pub fn static_shape(&self) -> Option<Vec<usize>> {
85        self.shape.iter().copied().collect()
86    }
87
88    /// For metadata: the full shape with `-1` for the batch and for open dimensions
89    pub fn v2_shape(&self) -> Vec<i64> {
90        std::iter::once(-1)
91            .chain(self.shape.iter().map(|d| d.map_or(-1, |d| d as i64)))
92            .collect()
93    }
94}
95
96/// The legacy single-tensor description in `GET /model/info` (`input_spec`, `output_spec`):
97/// per-item shape, open dimensions as 0
98#[derive(Debug, Clone, Serialize, Deserialize)]
99pub struct TensorSpec {
100    pub shape: Vec<usize>,
101    /// `float32`, `int64` or `int32`
102    pub dtype: String,
103    pub min_value: Option<f32>,
104    pub max_value: Option<f32>,
105}
106
107/// Description of a model input or output in `GET /model/info`
108#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
109pub struct TensorInfo {
110    pub name: String,
111    /// `FP32`, `INT64` or `INT32`
112    pub datatype: String,
113    /// Full shape; `-1` for the batch dimension and for dimensions fixed per request
114    pub shape: Vec<i64>,
115}
116
117impl From<&IoSpec> for TensorInfo {
118    fn from(spec: &IoSpec) -> Self {
119        Self {
120            name: spec.name.clone(),
121            datatype: spec.dtype.v2_name().to_string(),
122            shape: spec.v2_shape(),
123        }
124    }
125}
126
127/// Tensor elements in host memory
128#[derive(Debug, Clone, PartialEq)]
129pub enum Values {
130    F32(Vec<f32>),
131    I64(Vec<i64>),
132}
133
134impl Values {
135    /// Number of elements
136    pub fn len(&self) -> usize {
137        match self {
138            Self::F32(v) => v.len(),
139            Self::I64(v) => v.len(),
140        }
141    }
142
143    /// Whether there are no elements
144    pub fn is_empty(&self) -> bool {
145        self.len() == 0
146    }
147}
148
149/// A tensor in host memory: full shape (batch first) and row-major elements.
150///
151/// What [`TensorModel::run`](crate::TensorModel::run) takes and returns. furnace-burn turns one
152/// into a Burn tensor and back (its `HostTensorExt`).
153#[derive(Debug, Clone, PartialEq)]
154pub struct HostTensor {
155    pub shape: Vec<usize>,
156    pub values: Values,
157}
158
159impl HostTensor {
160    /// Items along the batch dimension; 0 for a tensor without one (an empty shape), which no
161    /// model accepts
162    pub fn batch(&self) -> usize {
163        self.shape.first().copied().unwrap_or(0)
164    }
165
166    /// Items `start..end` along the batch dimension
167    pub fn slice_batch(&self, start: usize, end: usize) -> Self {
168        let per_item: usize = self.shape[1..].iter().product();
169        let range = start * per_item..end * per_item;
170        let mut shape = self.shape.clone();
171        shape[0] = end - start;
172        Self {
173            shape,
174            values: match &self.values {
175                Values::F32(v) => Values::F32(v[range].to_vec()),
176                Values::I64(v) => Values::I64(v[range].to_vec()),
177            },
178        }
179    }
180
181    /// Join tensors of the same type and per-item shape along the batch dimension
182    pub fn concat_batch(parts: Vec<Self>) -> Result<Self> {
183        let mut parts = parts.into_iter();
184        let mut joined = parts
185            .next()
186            .ok_or_else(|| ModelError::InferenceFailed("nothing to join".to_string()))?;
187        for part in parts {
188            if part.shape[1..] != joined.shape[1..] {
189                return Err(ModelError::InferenceFailed(format!(
190                    "chunks of one batch returned shapes {:?} and {:?}",
191                    joined.shape, part.shape
192                ))
193                .into());
194            }
195            joined.shape[0] += part.shape[0];
196            match (&mut joined.values, part.values) {
197                (Values::F32(a), Values::F32(b)) => a.extend(b),
198                (Values::I64(a), Values::I64(b)) => a.extend(b),
199                _ => {
200                    return Err(ModelError::InferenceFailed(
201                        "chunks of one batch returned different types".to_string(),
202                    )
203                    .into());
204                }
205            }
206        }
207        Ok(joined)
208    }
209}
210
211#[cfg(test)]
212mod tests {
213    use super::*;
214
215    #[test]
216    fn batches_split_and_join() {
217        let tensor = HostTensor {
218            shape: vec![3, 2],
219            values: Values::I64(vec![1, 2, 3, 4, 5, 6]),
220        };
221        let parts = vec![tensor.slice_batch(0, 1), tensor.slice_batch(1, 3)];
222        assert_eq!(parts[1].shape, [2, 2]);
223        assert_eq!(HostTensor::concat_batch(parts).unwrap(), tensor);
224    }
225
226    #[test]
227    fn metadata_shapes() {
228        let spec = IoSpec {
229            name: "ids".to_string(),
230            dtype: DataType::Int64,
231            shape: vec![Some(256), None],
232        };
233        assert_eq!(spec.v2_shape(), [-1, 256, -1]);
234        assert_eq!(spec.static_shape(), None);
235        assert_eq!(TensorInfo::from(&spec).datatype, "INT64");
236        assert_eq!(DataType::from_v2_name("INT32"), Some(DataType::Int32));
237        assert_eq!(DataType::from_v2_name("FP16"), None);
238    }
239}
furnace/tensor.rs

furnace/
tensor.rs
