char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Johannes GäßlerCUDA: fix Gemma E4B MTP FlashAttention (#25148)e495d1e74

char/aliceai-t5-moe
398 B11 linesraw
1// This file has been autogenerated by generate_cu_files.py, do not edit manually.
2
3#include "../fattn-mma-f16.cuh"
4
5DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2);
6DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2);
7DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2);
8DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2);
9DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2);
10DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2);
11DECL_FATTN_MMA_F16_CASE(512, 512, 4, 2);