Implements dual-chunk-flash-attn backend for dual chunk attention with sparse attention support.

Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>

Author: sighingnow

CMakeLists.txt

@@ -189,6 +189,7 @@ set(VLLM_EXT_SRC
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
+  "csrc/attention/vertical_slash_index.cu"
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
@@ -550,7 +551,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 96266b1111111f3d11aabefaf3bacbab6a89d03c
+          GIT_TAG 323b789ae92b0376c0adfcfe1b48571ac32dc411
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn