Mojo struct
MLAConfig
@register_passable(trivial)
struct MLAConfig
Fields
- fa4_config (
FA4Config): - MMA_M (
Int): - BM (
Int): - BN (
Int): - BK0 (
Int): - BK1 (
Int): - depth (
Int): - k_rope_depth (
Int): - kv_depth (
Int): - cache_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_qk_stages (
Int): - num_pv_stages (
Int): - smem_used (
Int): - dtype_size (
Int): - split_m (
Bool): - qkv_swizzle_mode (
TensorMapSwizzle): - k_rope_swizzle_mode (
TensorMapSwizzle): - output_swizzle_mode (
TensorMapSwizzle):
Implemented traits
AnyType,
Copyable,
ImplicitlyCopyable,
ImplicitlyDestructible,
Movable,
RegisterPassable,
TrivialRegisterPassable
comptime members
__copy_ctor_is_trivial
comptime __copy_ctor_is_trivial = True
__del__is_trivial
comptime __del__is_trivial = True
__move_ctor_is_trivial
comptime __move_ctor_is_trivial = True
mbar_size
comptime mbar_size = size_of[DType.int64]()
num_correction_cols
comptime num_correction_cols = 1
num_threads
comptime num_threads = 512
sm100_smem_carveout
comptime sm100_smem_carveout = (B200 - 1024)
sm100_tmem_cols
comptime sm100_tmem_cols = 512
TMEM_S0
comptime TMEM_S0 = 0
Methods
__init__
__init__(*, num_q_heads: Int, group: Int, depth: Int, qkv_dtype_size: Int, k_rope_dtype_size: Int, output_dtype_size: Int, page_size: Int) -> Self
num_qo
supported
correction_smem_elements
num_active_warps_per_group
num_active_threads_per_group
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!