Flex Attention Extremely Slow

[why-in-Shanghaitech]

🐛 Describe the bug

I try to use flex attention in huggingface transformer, only to find it very slow. Compared to the sdpa implementation, flex attention is about 4-5 times slower, but it does save the CUDA memory.

Tested on RTX3090, A6000 and A100.

Here is the example code: https://gist.github.com/why-in-Shanghaitech/8b8205f98568c6741a2e38dfcdb9d362

I have no idea what is happening. Is this normal? Can anyone reproduce this? Or this problem is related to huggingface transformers?

Versions

PyTorch version: 2.6.0.dev20240914
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.2 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35

Python version: 3.9.19 | packaged by conda-forge | (main, Mar 20 2024, 12:50:21)  [GCC 12.3.0] (64-bit runtime)
Python platform: Linux-5.15.0-86-generic-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: 12.2.140
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                       x86_64
CPU op-mode(s):                     32-bit, 64-bit
Address sizes:                      46 bits physical, 48 bits virtual
Byte Order:                         Little Endian
CPU(s):                             64
On-line CPU(s) list:                0-63
Vendor ID:                          GenuineIntel
Model name:                         Intel(R) Xeon(R) Gold 6226R CPU @ 2.90GHz
CPU family:                         6
Model:                              85
Thread(s) per core:                 2
Core(s) per socket:                 16
Socket(s):                          2
Stepping:                           7
CPU max MHz:                        3900.0000
CPU min MHz:                        1200.0000
BogoMIPS:                           5800.00
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts pku ospke avx512_vnni md_clear flush_l1d arch_capabilities
Virtualization:                     VT-x
L1d cache:                          1 MiB (32 instances)
L1i cache:                          1 MiB (32 instances)
L2 cache:                           32 MiB (32 instances)
L3 cache:                           44 MiB (2 instances)
NUMA node(s):                       2
NUMA node0 CPU(s):                  0-15,32-47
NUMA node1 CPU(s):                  16-31,48-63
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit:        KVM: Mitigation: VMX disabled
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Retbleed:             Mitigation; Enhanced IBRS
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:           Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Mitigation; TSX disabled

Versions of relevant libraries:
[pip3] numpy==2.0.2
[pip3] torch==2.6.0.dev20240914
[pip3] torchaudio==2.5.0.dev20240914
[pip3] torchvision==0.20.0.dev20240914
[pip3] triton==3.1.0
[conda] blas                      1.0                         mkl  
[conda] libblas                   3.9.0            16_linux64_mkl    conda-forge
[conda] libcblas                  3.9.0            16_linux64_mkl    conda-forge
[conda] libjpeg-turbo             2.0.0                h9bf148f_0    pytorch-nightly
[conda] liblapack                 3.9.0            16_linux64_mkl    conda-forge
[conda] mkl                       2022.1.0           hc2b9512_224  
[conda] numpy                     2.0.2            py39h9cb892a_0    conda-forge
[conda] pytorch                   2.6.0.dev20240914 py3.9_cuda12.1_cudnn9.1.0_0    pytorch-nightly
[conda] pytorch-cuda              12.1                 ha16c6d3_6    pytorch-nightly
[conda] pytorch-mutex             1.0                        cuda    pytorch
[conda] torchaudio                2.5.0.dev20240914      py39_cu121    pytorch-nightly
[conda] torchtriton               3.1.0+5fe38ffd73            py39    pytorch-nightly
[conda] torchvision               0.20.0.dev20240914      py39_cu121    pytorch-nightly

cc @ezyang @chauhang @penguinwu @zou3519 @ydwu4 @bdhirsh @Chillee @drisspg @yanboliang @BoyuanFeng

[BoyuanFeng]

It looks like you are using flex_attention in eager mode. Please use compilation mode instead. i.e., flex_attention = torch.compile(flex_attention)

[why-in-Shanghaitech]

It looks like you are using flex_attention in eager mode. Please use compilation mode instead. i.e., flex_attention = torch.compile(flex_attention)

Thank you so much! This exactly solves the problem.