Mojo struct

FA4Config

@register_passable(trivial) struct FA4Config

Fields

MMA_M (Int):
BM (Int):
BN (Int):
BK0 (Int):
BK1 (Int):
depth (Int):
padded_depth (Int):
group (Int):
num_q_heads (Int):
num_kv_heads (Int):
TMEM_S1 (Int):
TMEM_O0 (Int):
TMEM_O1 (Int):
TMEM_P0 (Int):
TMEM_P1 (Int):
TMEM_C0 (Int):
TMEM_C1 (Int):
tmem_used (Int):
num_kv_stages (Int):
num_mma_stages (Int):
smem_used (Int):
dtype_size (Int):
split_m (Bool):
swizzle_mode (TensorMapSwizzle):

Implemented traits

AnyType, Copyable, ImplicitlyCopyable, Movable, UnknownDestructibility

Aliases

`copyinitis_trivial`

alias __copyinit__is_trivial = True

`delis_trivial`

alias __del__is_trivial = True

`moveinitis_trivial`

alias __moveinit__is_trivial = True

`mbar_size`

alias mbar_size = DType.int64.size_of()

`MMA_K`

alias MMA_K = 16

`num_correction_cols`

alias num_correction_cols = 1

`num_threads`

alias num_threads = 512

`sm100_smem_carveout`

alias sm100_smem_carveout = (GPUInfo.from_family(AcceleratorArchitectureFamily(32, 2048, 233472, 65536, 1024), "B200", Vendor(2), "cuda", "blackwell", 10, "sm_100a", 148) - 1024)

`sm100_tmem_cols`

alias sm100_tmem_cols = 512

`TMEM_S0`

alias TMEM_S0 = 0

Methods

`init`

__init__(*, num_q_heads: Int, group: Int, depth: Int, dtype_size: Int, swizzle_mode: TensorMapSwizzle) -> Self

`num_qo`

num_qo(self) -> Int

Returns:

Int

`supported`

supported(self) -> Bool

Returns:

Bool

`use_tmem_for_correction`

use_tmem_for_correction(self) -> Bool

Returns:

Bool

`correction_smem_elements`

correction_smem_elements(self) -> Int

Returns:

Int

`num_active_warps_per_group`

num_active_warps_per_group(self) -> Int

Returns:

Int

`num_active_threads_per_group`

num_active_threads_per_group(self) -> Int

Returns:

Int

Fields​

Implemented traits​

Aliases​

__copyinit__is_trivial​

__del__is_trivial​

__moveinit__is_trivial​

mbar_size​

MMA_K​

num_correction_cols​

num_threads​

sm100_smem_carveout​

sm100_tmem_cols​

TMEM_S0​

Methods​

__init__​

num_qo​

supported​

use_tmem_for_correction​

correction_smem_elements​

num_active_warps_per_group​

num_active_threads_per_group​

Fields

Implemented traits

Aliases

`copyinitis_trivial`

`delis_trivial`

`moveinitis_trivial`

`mbar_size`

`MMA_K`

`num_correction_cols`

`num_threads`

`sm100_smem_carveout`

`sm100_tmem_cols`

`TMEM_S0`

Methods

`init`

`num_qo`

`supported`

`use_tmem_for_correction`

`correction_smem_elements`

`num_active_warps_per_group`

`num_active_threads_per_group`