Enum TensorInit 

pub enum TensorInit {
    Zero,
    Pattern {
        modulus: i64,
        shift: i64,
    },
    Values(Vec<i64>),
    FloatValues(Vec<u64>),
    ConstF(u64),
    RandomBf16 {
        seed: u32,
        shift_milli: i32,
    },
    External,
}

How a tensor’s contents are produced — recognized from the source’s array-build loop, so a host can reproduce the data and the numerics claim can bound the operand range.

Variants

Zero

Pattern

value[p] = (p % modulus) + shift.

Fields

modulus: i64

shift: i64

Values(Vec<i64>)

FloatValues(Vec<u64>)

Float values (as bits, so the init stays hashable).

ConstF(u64)

★ ONE CONSTANT, NOT N COPIES OF IT. A tensor every element of which is the same float is a fact about the tensor, not n facts — but the only float fill used to be FloatValues, so declaring one cost a Vec as long as the tensor. On Qwen3.8 that is not a rounding error: a single full-width layer materialized 386,754,784 elements — 3.1 GB — and the real 64-layer model with its 5120 x 248320 output projection exhausted 114 GB of a 125 GB box before it finished lowering. Held symbolically the same declaration is eight bytes.

RandomBf16

A vendor-style bf16 random fill (seed + shift in thousandths) — carried so a recorded PCC is reproducible.

Fields

seed: u32

shift_milli: i32

External

Supplied at run time.

Implementations

impl TensorInit

pub fn abs_bound(&self) -> Option<i64>

A bound on |value|, when the fill makes one derivable.

pub fn range(&self) -> Option<(i64, i64)>

The least and the greatest value of an integer fill.

pub fn element(&self, p: usize) -> Option<i64>

The element at flat position p, for the deterministic integer fills.

pub fn element_f64(&self, p: usize) -> Option<f64>

The element at flat position p as a float, for every deterministic fill.

pub fn is_float(&self) -> bool

Trait Implementations

impl Clone for TensorInit

fn clone(&self) -> TensorInit

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for TensorInit

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for TensorInit

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Hash for TensorInit

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for TensorInit

fn eq(&self, other: &TensorInit) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for TensorInit

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl Eq for TensorInit

impl StructuralPartialEq for TensorInit