[FlexAttention] The flex attention test takes too long time to run in CI. #4265
Description
The flex attention test cases takes about ~5 hours which is not acceptable in CI.
The major problem for now are:
- Some configurations are suboptimal but in the autotune lists. It causes 4 times at most in IGC codegen to reduce the register spilling size.
-- Maybe we can try to use the auto GRF mode on Triton side
-- Need to update the Torch to use Intel turned configuration instead of using the common configuration from CUDA. - Some test cases uses high precision on fp32 matmul. It fall back to the FMA GEMM version which take a long time to generate the kernel with ~30000 instructions in one function.
-- Maybe we can try to enhance the FMA GEMM with optimized layout to reduce the kernel size and register size.