decode_attn_tiled.cuh File Reference

#include <cuda_runtime.h>
#include <math.h>

Go to the source code of this file.

Namespaces
namespace	ember_tiled

Macros
#define	EMBER_DECODE_ATTN_TILED_CUH

Functions
__host__ __device__ int	ember_tiled::tiled_chunk (int L, int min_chunk, int max_chunks)
__host__ __device__ int	ember_tiled::tiled_nchunks (int L, int chunk)
__host__ __device__ int	ember_tiled::tiled_grid_chunks (int max_len, int min_chunk, int max_chunks)
__host__ bool	ember_tiled::tiled_shape_ok (int heads, int head_dim, int max_len, int heads_max, int min_chunk, int max_chunks)
__host__ size_t	ember_tiled::tiled_shm_bytes (int head_dim, int max_len, int min_chunk, int max_chunks)
__device__ __forceinline__ void	ember_tiled::tiled_split_block (const float *__restrict__ Qh, const float *__restrict__ K, const float *__restrict__ V, int L, int c, int kvh, int num_kv_heads, int head_dim, float scale, int min_chunk, int max_chunks, float *__restrict__ part)
__device__ __forceinline__ void	ember_tiled::tiled_merge_block (const float *__restrict__ base, float *__restrict__ out_h, int L, int head_dim, int min_chunk, int max_chunks)

Variables
constexpr int	ember_tiled::kTiledThreads = 128
constexpr int	ember_tiled::kTiledMinChunk = 32
constexpr int	ember_tiled::kTiledMaxChunks = 64
constexpr int	ember_tiled::kTiledChunkMax = 2048
constexpr int	ember_tiled::kTiledHeadDimMax = 256

Macro Definition Documentation

EMBER_DECODE_ATTN_TILED_CUH

#define EMBER_DECODE_ATTN_TILED_CUH

Definition at line 22 of file decode_attn_tiled.cuh.