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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
//! Shape/type introspection: `TensorTypeAndShapeInfo` over the owning
//! `GetTensorTypeAndShape` path (idx 65 — returns an OWNING handle, one release).
use crate::{Error, Result, api, check, sys};
use std::ffi::{CStr, c_char};
use std::ptr;
/// Owning wrapper over `OrtTensorTypeAndShapeInfo` obtained from a value via
/// `GetTensorTypeAndShape`. Released on drop (`ReleaseTensorTypeAndShapeInfo`, idx 99).
pub struct TensorTypeAndShapeInfo {
info: *mut sys::TensorTypeAndShapeInfoHandle,
}
impl TensorTypeAndShapeInfo {
/// Wrap an owning handle returned by `GetTensorTypeAndShape` (idx 65). The wrapper
/// assumes ownership and will release it on drop.
///
/// # Safety
/// `info` must be a freshly-allocated owning handle from `GetTensorTypeAndShape`.
pub(crate) unsafe fn from_owning(info: *mut sys::TensorTypeAndShapeInfoHandle) -> Self {
Self { info }
}
/// Build a fresh, empty type+shape info (`CreateTensorTypeAndShapeInfo`). Fill it with
/// [`Self::set_element_type`] + [`Self::set_dimensions`], then hand it to a shape-inference
/// context (`ShapeInferContext::set_output_type_shape` with the `custom-ops` feature) or inspect it. Owning —
/// released on drop.
pub fn new() -> Result<Self> {
let mut info: *mut sys::TensorTypeAndShapeInfoHandle = ptr::null_mut();
check(unsafe { api().create_tensor_type_and_shape_info()(&mut info) })?;
let info = crate::ensure_non_null(info, "tensor type and shape info")?;
// SAFETY: CreateTensorTypeAndShapeInfo allocates an owning handle.
Ok(unsafe { Self::from_owning(info) })
}
/// Set the element type (`SetTensorElementType`).
pub fn set_element_type(&mut self, ty: sys::ElementType) -> Result<()> {
check(unsafe { api().set_tensor_element_type()(self.info, ty) })
}
/// Set the concrete dimensions (`SetDimensions`).
pub fn set_dimensions(&mut self, dims: &[i64]) -> Result<()> {
check(unsafe { api().set_dimensions()(self.info, dims.as_ptr(), dims.len()) })
}
/// The raw owning handle (`pub(crate)` — shape-inference and model-editor wrappers pass it
/// to ORT APIs that borrow it).
#[cfg(any(feature = "custom-ops", feature = "model-editor"))]
pub(crate) fn as_ptr(&self) -> *const sys::TensorTypeAndShapeInfoHandle {
self.info as *const sys::TensorTypeAndShapeInfoHandle
}
/// Element type of the tensor.
pub fn element_type(&self) -> Result<sys::ElementType> {
let mut et = sys::ElementType::Undefined;
check(unsafe {
api().get_tensor_element_type()(
self.info as *const sys::TensorTypeAndShapeInfoHandle,
&mut et,
)
})?;
Ok(et)
}
/// Number of dimensions (rank).
pub fn rank(&self) -> Result<usize> {
let mut n: usize = 0;
check(unsafe {
api().get_dimensions_count()(
self.info as *const sys::TensorTypeAndShapeInfoHandle,
&mut n,
)
})?;
Ok(n)
}
/// Total element count (product of dimensions).
///
/// This computes from dimensions in Rust instead of calling ORT
/// `GetTensorShapeElementCount`, because ORT may report a SafeInt overflow for static
/// symbolic shapes such as `[-1, 1000]`. If any dimension is dynamic/unknown, this returns
/// a controlled ZRT error.
pub fn element_count(&self) -> Result<usize> {
checked_element_count(&self.dims()?)
}
/// Concrete dimensions, e.g. `[1, 1, 28, 28]`.
pub fn dims(&self) -> Result<Vec<i64>> {
let n = self.rank()?;
let mut out = vec![0i64; n];
check(unsafe {
api().get_dimensions()(
self.info as *const sys::TensorTypeAndShapeInfoHandle,
out.as_mut_ptr(),
n,
)
})?;
Ok(out)
}
/// Symbolic (named) dimensions: `Some("batch")` where the model declared a symbolic
/// dim, `None` where it is concrete. Length equals `rank()`. The strings are borrowed
/// from the engine-owned handle for the lifetime of `self`.
pub fn symbolic_dims(&self) -> Result<Vec<Option<&str>>> {
let n = self.rank()?;
let mut ptrs: Vec<*const c_char> = vec![ptr::null(); n];
check(unsafe {
api().get_symbolic_dimensions()(
self.info as *const sys::TensorTypeAndShapeInfoHandle,
ptrs.as_mut_ptr(),
n,
)
})?;
ptrs.iter()
.map(|&p| {
if p.is_null() {
Ok(None)
} else {
// SAFETY: the engine guarantees a NUL-terminated UTF-8-ish C string for the
// lifetime of the handle. We only borrow it; we do not free it.
unsafe { CStr::from_ptr(p) }
.to_str()
.map(Some)
.map_err(|_| {
Error::new(-1, "zrt: symbolic dimension name is not valid UTF-8")
})
}
})
.collect()
}
}
pub(crate) fn checked_element_count(dims: &[i64]) -> Result<usize> {
let mut count = 1usize;
for &dim in dims {
if dim < 0 {
return Err(Error::new(
-1,
format!("tensor shape contains a dynamic/unknown dimension ({dim})"),
));
}
let dim = usize::try_from(dim)
.map_err(|_| Error::new(-1, "tensor dimension does not fit usize"))?;
count = count
.checked_mul(dim)
.ok_or_else(|| Error::new(-1, "tensor shape element count overflows usize"))?;
}
Ok(count)
}
impl Drop for TensorTypeAndShapeInfo {
fn drop(&mut self) {
unsafe { api().release_tensor_type_and_shape_info()(self.info) }
}
}
/// Introspect a tensor value's full type+shape (owning path). The value MUST be a tensor;
/// for map/sequence values use `OwnedValue::value_type` instead.
pub(crate) fn tensor_type_and_shape(
value: *const sys::ValueHandle,
) -> Result<TensorTypeAndShapeInfo> {
let mut info: *mut sys::TensorTypeAndShapeInfoHandle = ptr::null_mut();
check(unsafe { api().get_tensor_type_and_shape()(value, &mut info) })?;
let info = crate::ensure_non_null(info, "tensor type and shape info")?;
Ok(unsafe { TensorTypeAndShapeInfo::from_owning(info) })
}
#[cfg(test)]
mod tests {
use super::*;
/// Engine-backed round trip of the builder: create → set element type + dims → read back.
/// No model needed; exercises CreateTensorTypeAndShapeInfo + SetTensorElementType +
/// SetDimensions (+ the read accessors), all released on drop.
#[test]
fn type_and_shape_info_builder_round_trip() {
let mut info = TensorTypeAndShapeInfo::new().expect("new");
info.set_element_type(sys::ElementType::Float)
.expect("set elem type");
info.set_dimensions(&[2, 3]).expect("set dims");
assert_eq!(info.element_type().unwrap(), sys::ElementType::Float);
assert_eq!(info.dims().unwrap(), vec![2, 3]);
assert_eq!(info.rank().unwrap(), 2);
assert_eq!(info.element_count().unwrap(), 6);
eprintln!("type_and_shape_info_builder_round_trip: create + set + read OK");
}
#[test]
fn type_info_accepts_newer_quantized_metadata_element_types() {
for ty in [
sys::ElementType::Float8E4M3FN,
sys::ElementType::Float8E5M2,
sys::ElementType::Uint4,
sys::ElementType::Int4,
sys::ElementType::Float4E2M1,
] {
let mut info = TensorTypeAndShapeInfo::new().expect("new");
info.set_element_type(ty).expect("set elem type");
assert_eq!(info.element_type().unwrap(), ty);
}
}
#[test]
fn checked_element_count_rejects_dynamic_and_overflow() {
assert_eq!(checked_element_count(&[1, 1000]).unwrap(), 1000);
assert!(checked_element_count(&[-1, 1000]).is_err());
assert!(checked_element_count(&[i64::MAX, 3]).is_err());
}
}