Nvidia 50 series might crash on Triton, add option to skip specific attention modes

[x5nder]

Nvidia GeForce RTX 5060 Ti 16GB

sageattention 2.2.0+cu130torch2.9.0andhigher.post4
torch 2.9.1+cu130
triton 3.5.1.post24

Attention Optimizer crashes on Triton:

[Benchmark] Running... (head_dim=128, seq_len=8192)
[Benchmark] Available: ['basic', 'sub_quad', 'split', 'pytorch', 'sage_auto', 'sage_cuda', 'sage_triton', 'sage_fp8_cuda', 'sage_fp8_cuda_fast', 'flash']
__triton_launcher.c
   Creating library C:\Users\x5nde\AppData\Local\Temp\tmprie7271v\__triton_launcher.cp312-win_amd64.lib and object C:\Users\x5nde\AppData\Local\Temp\tmprie7271v\__triton_launcher.cp312-win_amd64.exp
Exception Code: 0x80000003
 #0 0x00007ffb0fb2dc25 mlir::detail::FallbackTypeIDResolver::registerImplicitTypeID(class llvm::StringRef) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\triton\_C\libtriton.pyd+0x3d0dc25)
 #1 0x00007ffc5b6d1989 (C:\WINDOWS\System32\ucrtbase.dll+0xc1989)
 #2 0x00007ffc5b6b4ab1 (C:\WINDOWS\System32\ucrtbase.dll+0xa4ab1)
 #3 0x00007ffc5b6d19de (C:\WINDOWS\System32\ucrtbase.dll+0xc19de)
 #4 0x00007ffbd4351911 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\VCRUNTIME140_1.dll+0x1911)
 #5 0x00007ffbd435218f (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\VCRUNTIME140_1.dll+0x218f)
 #6 0x00007ffbd43521e9 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\VCRUNTIME140_1.dll+0x21e9)
 #7 0x00007ffbd4354019 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\VCRUNTIME140_1.dll+0x4019)
 #8 0x00007ffc5ddc62ff (C:\WINDOWS\SYSTEM32\ntdll.dll+0x1662ff)
 #9 0x00007ffc5dc72327 (C:\WINDOWS\SYSTEM32\ntdll.dll+0x12327)
#10 0x00007ffc5dc6a961 (C:\WINDOWS\SYSTEM32\ntdll.dll+0xa961)
#11 0x00007ffc5ae7a80a (C:\WINDOWS\System32\KERNELBASE.dll+0xca80a)
#12 0x00007ffbd50b5267 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\VCRUNTIME140.dll+0x5267)
#13 0x00007ffc2fbbbcfb c10::cuda::c10_cuda_check_implementation(int, char const *, char const *, int, bool) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\c10_cuda.dll+0x3bcfb)
#14 0x00007ffc2fbc01c5 c10::cuda::warn_or_error_on_sync(void) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\c10_cuda.dll+0x401c5)
#15 0x00007ffc2fc5909e c10::StorageImpl::~StorageImpl(void) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\c10.dll+0x2909e)
#16 0x00007ffb1ddd77e5 at::DynamicLibrary::sym(char const *) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\torch_cpu.dll+0x61477e5)
#17 0x00007ffc2fc44882 c10::ConstantSymNodeImpl<__int64>::~ConstantSymNodeImpl<__int64>(void) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\c10.dll+0x14882)
#18 0x00007ffc2fc7f67c c10::TensorImpl::~TensorImpl(void) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\c10.dll+0x4f67c)
#19 0x00007ffb1ddd7825 at::DynamicLibrary::sym(char const *) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\torch_cpu.dll+0x6147825)
#20 0x00007ffb1dd75f58 at::TensorBase::reset(void) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\torch_cpu.dll+0x60e5f58)
#21 0x00007ffbf247e395 initModule (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\torch_python.dll+0x4be395)
#22 0x00007ffbf252ba9f THPVariable_Wrap(class at::TensorBase const &) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\torch_python.dll+0x56ba9f)
#23 0x00007ffbf252bda3 THPVariable_Wrap(class at::TensorBase const &) (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\Lib\site-packages\torch\lib\torch_python.dll+0x56bda3)
#24 0x00007ffbd4a824ea (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x824ea)
#25 0x00007ffbd4a8304b (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x8304b)
#26 0x00007ffbd4a8103c (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x8103c)
#27 0x00007ffbd4a7be18 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x7be18)
#28 0x00007ffbd4ab286b (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0xb286b)
#29 0x00007ffbd4ab275f (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0xb275f)
#30 0x00007ffbd4a866ee (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x866ee)
#31 0x00007ffbd4a4b201 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x4b201)
#32 0x00007ffbd4c44f58 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x244f58)
#33 0x00007ffbd4c3b57d (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x23b57d)
#34 0x00007ffc4d5e7334 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\DLLs\_asyncio.pyd+0x7334)
#35 0x00007ffc4d5e6c7e (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\DLLs\_asyncio.pyd+0x6c7e)
#36 0x00007ffbd4a6560e (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x6560e)
#37 0x00007ffbd4c696e0 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x2696e0)
#38 0x00007ffbd4c699c0 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x2699c0)
#39 0x00007ffbd4a68376 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x68376)
#40 0x00007ffbd4ab2815 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0xb2815)
#41 0x00007ffbd4ab275f (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0xb275f)
#42 0x00007ffbd4a866ee (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x866ee)
#43 0x00007ffbd4a8103c (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x8103c)
#44 0x00007ffbd4a7be6c (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x7be6c)
#45 0x00007ffbd4ab2815 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0xb2815)
#46 0x00007ffbd4ab275f (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0xb275f)
#47 0x00007ffbd4a2304c (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x2304c)
#48 0x00007ffbd4a22ff2 (C:\Users\x5nde\AppData\Local\Programs\Python\Python312\python312.dll+0x22ff2)
#49 0x00007ffc5b6137b0 (C:\WINDOWS\System32\ucrtbase.dll+0x37b0)
#50 0x00007ffc5cf6e8d7 (C:\WINDOWS\System32\KERNEL32.DLL+0x2e8d7)
#51 0x00007ffc5dcec53c (C:\WINDOWS\SYSTEM32\ntdll.dll+0x8c53c)

Apparently this is a known issue with Nvidia 50-series card. It would be nice to be able to skip a specific attention mode (sage_triton, in this case) so ComfyUI won't crash while testing the other modes.