Mojo struct
MLA_SM100_Decode_Config
struct MLA_SM100_Decode_Config
Fields
- MMA_M (
Int): - BM (
Int): - BN (
Int): - BK0 (
Int): - BK1 (
Int): - q_depth (
Int): - depth (
Int): - padded_depth (
Int): - padded_q_depth (
Int): - rope_depth (
Int): - group (
Int): - num_q_heads (
Int): - num_kv_heads (
Int): - tmem_used (
Int): - num_kv_stages (
Int): - smem_used (
Int): - dtype_size (
Int): - swizzle_mode (
TensorMapSwizzle): - kv_swizzle_mode (
TensorMapSwizzle): - decoding_warp_split_k (
Bool): - out_rows (
Int):
Implemented traits
AnyType,
ImplicitlyDestructible
comptime members
__del__is_trivial
comptime __del__is_trivial = True
cta_group
comptime cta_group = 1
mbar_size
comptime mbar_size = size_of[DType.int64]()
MMA_K
comptime MMA_K = 16
num_threads
comptime num_threads = 384
sm100_smem_carveout
comptime sm100_smem_carveout = (B200 - 1024)
sm100_tmem_cols
comptime sm100_tmem_cols = 512
TMEM_CORR_LI
comptime TMEM_CORR_LI = 321
TMEM_CORR_SCALE
comptime TMEM_CORR_SCALE = 320
TMEM_O
comptime TMEM_O = 0
TMEM_S0
comptime TMEM_S0 = 256
TMEM_S1
comptime TMEM_S1 = 288
Methods
__init__
__init__(out self, *, num_q_heads: Int, group: Int, depth: Int, q_depth: Int, dtype_size: Int, swizzle_mode: TensorMapSwizzle, kv_swizzle_mode: TensorMapSwizzle, page_size: Int, decoding_warp_split_k: Bool)
supported
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!