Main branch compilation on nvcc 12.6

[roded2]

A recent commit ( 7a80279#diff-c5844b57c2c63ac3f0275f6f19f0f0fae920ef927cf4cba54ce3d52defae0aba ) introduced a change to hopper/setup.py which restricts the compilation to nvcc 12.3 (

flash-attention/hopper/setup.py

Line 392 in 6b1d059

if bare_metal_version != Version("12.3"): # nvcc 12.3 gives the best perf currently

). The reason for this change according to the comment in the code is because 12.3 gives better perf.
I tried to run hopper/setup.py with the nvcc which is installed on my machine (nvcc 12.6) by changing the if statement in the code attached above to evaluate to false. The compilation fails deterministically due to a segfault in nvcc:

[13/42] /home/user/bin/nvcc --generate-dependencies-with-compile --dependency-output /home/user/flash-attention/hopper/build/temp.linux-x86_64-cpython-312/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.o.d -I/home/user/flash-attention/hopper -I/home/user/flash-attention/csrc/cutlass/include -I/home/user/lib/python3.12/site-packages/torch/include -I/home/user/lib/python3.12/site-packages/torch/include/torch/csrc/api/include -I/home/user/lib/python3.12/site-packages/torch/include/TH -I/home/user/lib/python3.12/site-packages/torch/include/THC -I/home/user/include -I/home/user/include/python3.12 -c -c /home/user/flash-attention/hopper/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.cu -o /home/user/flash-attention/hopper/build/temp.linux-x86_64-cpython-312/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' --threads 4 -O3 -std=c++17 --ftemplate-backtrace-limit=0 --use_fast_math --resource-usage -lineinfo -DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED -DCUTLASS_DEBUG_TRACE_LEVEL=0 -DNDEBUG -gencode arch=compute_90a,code=sm_90a -DFLASHATTENTION_DISABLE_SM8x -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=flash_attn_3_cuda -D_GLIBCXX_USE_CXX11_ABI=1 -ccbin /home/user/bin/x86_64-conda-linux-gnu-cc FAILED: /home/user/flash-attention/hopper/build/temp.linux-x86_64-cpython-312/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.o /home/user/bin/nvcc --generate-dependencies-with-compile --dependency-output /home/user/flash-attention/hopper/build/temp.linux-x86_64-cpython-312/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.o.d -I/home/user/flash-attention/hopper -I/home/user/flash-attention/csrc/cutlass/include -I/home/user/lib/python3.12/site-packages/torch/include -I/home/user/lib/python3.12/site-packages/torch/include/torch/csrc/api/include -I/home/user/lib/python3.12/site-packages/torch/include/TH -I/home/user/lib/python3.12/site-packages/torch/include/THC -I/home/user/include -I/home/user/include/python3.12 -c -c /home/user/flash-attention/hopper/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.cu -o /home/user/flash-attention/hopper/build/temp.linux-x86_64-cpython-312/instantiations/flash_fwd_hdimall_bf16_paged_split_softcap_sm90.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options ''"'"'-fPIC'"'"'' --threads 4 -O3 -std=c++17 --ftemplate-backtrace-limit=0 --use_fast_math --resource-usage -lineinfo -DCUTE_SM90_EXTENDED_MMA_SHAPES_ENABLED -DCUTLASS_DEBUG_TRACE_LEVEL=0 -DNDEBUG -gencode arch=compute_90a,code=sm_90a -DFLASHATTENTION_DISABLE_SM8x -DTORCH_API_INCLUDE_EXTENSION_H '-DPYBIND11_COMPILER_TYPE="_gcc"' '-DPYBIND11_STDLIB="_libstdcpp"' '-DPYBIND11_BUILD_ABI="_cxxabi1016"' -DTORCH_EXTENSION_NAME=flash_attn_3_cuda -D_GLIBCXX_USE_CXX11_ABI=1 -ccbin /home/user/bin/x86_64-conda-linux-gnu-cc nvcc warning : incompatible redefinition for option 'compiler-bindir', the last value of this option was used Segmentation fault (core dumped)

Does the compilation has to happen on nvcc 12.3 from a correctness side as well and not just perf? Is it a known issue?
The command I used to compile is MAX_JOBS=64 FLASH_ATTENTION_FORCE_BUILD=TRUE FLASH_ATTENTION_DISABLE_SM80=TRUE python setup.py install but same issue happens with just running python setup.py install (ran on ubuntu 22).
Thanks

[ankutalev]

Same issue, can reproduce with just building CUDA sources with nvcc

[tridao]

I haven't tried nvcc 12.6 with all of the features.
You can try disabling more features. With these flags it should compile ok with 12.6 (with lower perf than compiling with 12.3).

FLASH_ATTENTION_DISABLE_BACKWARD=FALSE
FLASH_ATTENTION_DISABLE_SPLIT=TRUE
FLASH_ATTENTION_DISABLE_LOCAL=TRUE
FLASH_ATTENTION_DISABLE_PAGEDKV=TRUE
FLASH_ATTENTION_DISABLE_FP16=TRUE
FLASH_ATTENTION_DISABLE_FP8=TRUE
FLASH_ATTENTION_DISABLE_APPENDKV=TRUE
FLASH_ATTENTION_DISABLE_VARLEN=TRUE
FLASH_ATTENTION_DISABLE_CLUSTER=FALSE
FLASH_ATTENTION_DISABLE_PACKGQA=TRUE
FLASH_ATTENTION_DISABLE_SOFTCAP=TRUE
FLASH_ATTENTION_DISABLE_HDIM64=TRUE
FLASH_ATTENTION_DISABLE_HDIM96=TRUE
FLASH_ATTENTION_DISABLE_HDIM128=FALSE
FLASH_ATTENTION_DISABLE_HDIM192=TRUE
FLASH_ATTENTION_DISABLE_HDIM256=TRUE