TensorFloat32

For conceptual model usage and type mapping, see:

• Model

class pyrogue.TensorFloat32(bitSize)

Model class for 32-bit TensorFloat32 (NVIDIA TF32) numbers.

Parameters:

• bitSize (int) – Number of bits being represented. Must be 32.

• args (Any)

• kwargs (Any)

Return type:

Any

Notes

Format: 1 sign bit, 8 exponent bits, 10 mantissa bits (1s/8e/10m). Bias = 127. Same exponent range as float32. Stored in a 4-byte word with the lower 13 mantissa bits zeroed. Supports infinity and NaN. Maximum representable finite value is approximately 3.40e38. Supported by NVIDIA Ampere (A100), Hopper (H100), and Blackwell GPUs.

pytype

alias of float

toBytes(value)

Convert float to 4-byte TensorFloat32 encoding.

TF32 zeros the lower 13 mantissa bits of the float32 bit pattern. All special values (NaN, infinity, zero, subnormals) are preserved.

Return type:

bytearray

Parameters:

value (float)

fromBytes(ba)

Decode 4-byte TensorFloat32 encoding to float.

TF32 bit pattern is a valid float32 with lower 13 mantissa bits zeroed. Direct reinterpretation as float32 is sufficient.

Return type:

float

Parameters:

ba (bytes)

fromString(string)

Parse a string into a float value.

Return type:

float

Parameters:

string (str)

minValue()

Return the minimum representable finite TF32 value (~-3.40e38).

Return type:

float

maxValue()

Return the maximum representable finite TF32 value (~3.40e38).

Return type:

float

class pyrogue.TensorFloat32BE(bitSize)

Model class for TensorFloat32 floats stored as big endian.

Parameters:

• args (Any)

• kwargs (Any)

Return type:

Any