Nx.Floating (Nx v0.11.0)

Copy Markdown View Source

Functions for loading and dumping floating-point formats not supported by the Erlang VM.

Summary

Functions

dump_bf16(x)

Dumps an Elixir value to BF16 (Brain Float 16) binary representation.

dump_f8(x)

Dumps an Elixir value to F8 (E5M2) binary representation.

dump_f8_e4m3fn(x)

Dumps an Elixir value to F8 E4M3FN binary representation.

load_bf16(bf16)

Loads a BF16 (Brain Float 16) value from its binary representation.

load_f8(arg)

Loads an F8 (E5M2) value from its binary representation.

load_f8_e4m3fn(arg)

Loads an F8 E4M3FN value from its binary representation.

Functions

dump_bf16(x)

Dumps an Elixir value to BF16 (Brain Float 16) binary representation.

Examples

iex> Nx.Floating.load_bf16(Nx.Floating.dump_bf16(1.0))
1.0

iex> Nx.Floating.load_bf16(Nx.Floating.dump_bf16(-2.5))
-2.5

iex> Nx.Floating.load_bf16(Nx.Floating.dump_bf16(:nan))
:nan

dump_f8(x)

Dumps an Elixir value to F8 (E5M2) binary representation.

Examples

iex> Nx.Floating.load_f8(Nx.Floating.dump_f8(1.0))
1.0

iex> Nx.Floating.load_f8(Nx.Floating.dump_f8(-2.0))
-2.0

iex> Nx.Floating.load_f8(Nx.Floating.dump_f8(:nan))
:nan

dump_f8_e4m3fn(x)

Dumps an Elixir value to F8 E4M3FN binary representation.

Finite values are clamped to the E4M3FN range [-448.0, 448.0]. Since E4M3FN has no infinity, :infinity and :neg_infinity saturate to max/min finite values. Only :nan maps to NaN.

Examples

iex> Nx.Floating.dump_f8_e4m3fn(1.0)
<<0x38>>

iex> Nx.Floating.dump_f8_e4m3fn(0.0)
<<0x00>>

iex> Nx.Floating.dump_f8_e4m3fn(448.0)
<<0x7E>>

iex> Nx.Floating.dump_f8_e4m3fn(-448.0)
<<0xFE>>

iex> Nx.Floating.dump_f8_e4m3fn(:infinity)
<<0x7E>>

iex> Nx.Floating.dump_f8_e4m3fn(:neg_infinity)
<<0xFE>>

iex> Nx.Floating.dump_f8_e4m3fn(:nan)
<<0x7F>>

load_bf16(bf16)

Loads a BF16 (Brain Float 16) value from its binary representation.

Returns a float, :nan, :infinity, or :neg_infinity.

Examples

iex> Nx.Floating.load_bf16(Nx.Floating.dump_bf16(1.0))
1.0

iex> Nx.Floating.load_bf16(Nx.Floating.dump_bf16(:infinity))
:infinity

iex> Nx.Floating.load_bf16(Nx.Floating.dump_bf16(:neg_infinity))
:neg_infinity

load_f8(arg)

Loads an F8 (E5M2) value from its binary representation.

Returns a float, :nan, :infinity, or :neg_infinity.

Examples

iex> Nx.Floating.load_f8(Nx.Floating.dump_f8(1.0))
1.0

iex> Nx.Floating.load_f8(Nx.Floating.dump_f8(:infinity))
:infinity

iex> Nx.Floating.load_f8(Nx.Floating.dump_f8(:neg_infinity))
:neg_infinity

load_f8_e4m3fn(arg)

Loads an F8 E4M3FN value from its binary representation.

Returns a float or :nan (E4M3FN has no infinity).

Examples

iex> Nx.Floating.load_f8_e4m3fn(Nx.Floating.dump_f8_e4m3fn(1.0))
1.0

iex> Nx.Floating.load_f8_e4m3fn(Nx.Floating.dump_f8_e4m3fn(:nan))
:nan

iex> Nx.Floating.load_f8_e4m3fn(Nx.Floating.dump_f8_e4m3fn(0.0))
0.0

iex> Nx.Floating.load_f8_e4m3fn(Nx.Floating.dump_f8_e4m3fn(448.0))
448.0