Use single mma warp group for short q_len in FA to optimize decoding performance (#18985)

This commit is contained in:
Qiaolin Yu
2026-02-23 01:01:28 -08:00
committed by GitHub
parent 6448de7e96
commit d80c884a27

View File

@@ -92,7 +92,7 @@ FetchContent_Populate(repo-flashinfer)
FetchContent_Declare(
repo-flash-attention
GIT_REPOSITORY https://github.com/sgl-project/sgl-attn
GIT_TAG f866ec34002250e74c8bbcbcffa0e1ae71300b2d
GIT_TAG cc75c5c5979a607ad20a6828635646f9841acf01
GIT_SHALLOW OFF
)
FetchContent_Populate(repo-flash-attention)
@@ -464,21 +464,21 @@ if (SGL_KERNEL_ENABLE_FA3)
endif()
file(GLOB FA3_BF16_GEN_SRCS
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimall_bf16*_sm90.cu")
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9]*_bf16*_sm90.cu")
file(GLOB FA3_BF16_GEN_SRCS_
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimdiff_bf16*_sm90.cu")
list(APPEND FA3_BF16_GEN_SRCS ${FA3_BF16_GEN_SRCS_})
# FP16 source files
# FP16 source files - use individual hdim files instead of hdimall to avoid ptxas crash
file(GLOB FA3_FP16_GEN_SRCS
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimall_fp16*_sm90.cu")
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9]*_fp16*_sm90.cu")
file(GLOB FA3_FP16_GEN_SRCS_
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimdiff_fp16*_sm90.cu")
list(APPEND FA3_FP16_GEN_SRCS ${FA3_FP16_GEN_SRCS_})
# FP8 source files
file(GLOB FA3_FP8_GEN_SRCS
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimall_e4m3*_sm90.cu")
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdim[0-9]*_e4m3*_sm90.cu")
file(GLOB FA3_FP8_GEN_SRCS_
"${repo-flash-attention_SOURCE_DIR}/hopper/instantiations/flash_fwd_hdimdiff_e4m3*_sm90.cu")
list(APPEND FA3_FP8_GEN_SRCS ${FA3_FP8_GEN_SRCS_})