[FA2] Release flash-attn-mma split-kv/q🎉 (#161)

* Update README.md

* Update README.md

Author: DefTruth

README.md

@@ -46,6 +46,7 @@ I have also implemented **FlashAttention-2** using pure MMA PTX instructions, wh
 
 ![flash-attn-mma](https://github.com/user-attachments/assets/6f66796d-44d5-4ec1-b224-af997bd152b2)
 
+
 |CUDA Cores|Sliced K (Loop over N/D)|Tile Block (Br, Bc, Bd)|MMA (m16n8k16)|
 |:---:|:---:|:---:|:---:|
 |✔️|✔️|✔️|✔️|
@@ -56,6 +57,8 @@ I have also implemented **FlashAttention-2** using pure MMA PTX instructions, wh
 
 The `Split KV` and `Split Q` implementations have been carried out in [flash-attention-mma⚡️⚡️](./kernels/flash-attn) for performance comparison. The `Split KV` method, which involves splitting all QKV across MMA (Warps), is slower than `Split Q` policy, which splitting Q across MMA(Warps) and keep access KV for all MMA(Warps).
 
+![flash-attn](https://github.com/user-attachments/assets/11490fbc-2a4a-4630-abe8-91a9d1251cba)
+
 - 📚 Split KV (Basic, FlashAttention-1)
 
 ```C++

kernels/flash-attn/README.md

@@ -26,6 +26,8 @@ This repository's implementation of FlashAttention is intended solely for learni
 
 The `Split KV` and `Split Q` implementations have been carried out in [flash-attention-mma⚡️⚡️](.) for performance comparison. The `Split KV` method, which involves splitting all QKV across MMA (Warps) using a naive matmul (MMA) and Warp tiling policy, is slower compared to the `Split Q` policy, which splitting Q across MMA(Warps) and keep access KV for all MMA(Warps).
 
+![flash-attn](https://github.com/user-attachments/assets/11490fbc-2a4a-4630-abe8-91a9d1251cba)
+
 ## 📖 Split KV (Basic, FlashAttention-1)
 <div id="mma-split-kv"></div>