-
Notifications
You must be signed in to change notification settings - Fork 21.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.
Already on GitHub? Sign in to your account
SDPA does not switch off dropout during evaluation #124464
Comments
Hey @Giventicket, in your first code block, you're setting the model to evaluation mode, but still passing Is the purpose of your issue that this is also happening in the doc example? We'd accept a PR updating the doc to fix this. |
yes, still passing F.dropout(..., training=True) is the problem. Thanks for commenting and PR. |
@jbschlosser checkout plz |
@Giventicket Not sure I'm clear on exactly what the malfunction is, but I'll hazard a guess that you're trying to use SDPA from your module and it isn't respecting the module's evaluation status. Can you workaround this by manually providing class Model1(nn.Module):
def __init__(self, p=0.5):
super().__init__()
self.p = p
def forward(self, ...):
return F.scaled_dot_product_attention(..., dropout_p=(self.p if self.training else 0.0)) If I'm wrong about the problem, is it possible to get more detail on the behavior you're looking for? |
@jbschlosser whoop that makes sense, thank a lot! |
I stumbled upon this behavior as well. Yes, it appears that SDPA doesn't respect .eval() and uses non-zero value of dropout_p during inference. I think what @Giventicket was trying to say is that in the pseudocode for SDPA there is a line:
which makes you think as if train is set to permanently as True, regardless of whether the model is in train or eval mode. If that's indeed how SDPA works in C, it is, at minimum, peculiar. Doing this |
@anmorgunov Note that SDPA (AKA We could have a An alternative would be to bleed the module-level training concept down into the function: ... = F.scaled_dot_product_attention(..., dropout_p=0.4, train=False) # if train is False, ignore passed-in dropout_p and use 0.0 |
once you spell it out, it makes perfect sense to behave it like this and in no other way |
if we detect Yes, SDPA is a function, not a module, but having userland avoid footguns is our imperative and responsibility, so we gotta do something about it. |
I don't think it's possible to detect eval mode of the module that made the function call from within SDPA without a signature change. We can detect |
Reopening to resolve this. At a minimum, we should add a doc warning about this behavior. |
Fixes #124464 TBD: validate formatting [ghstack-poisoned]
Fixes pytorch#124464 Pull Request resolved: pytorch#126294 Approved by: https://github.com/mikaylagawarecki, https://github.com/drisspg
馃悰 Describe the bug
I find out the detailed pseudo-like code from https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html while using flash attention.
However the metric measured from flash attention model is changed (especially when I use dropout_p = 0.1) which is supposed to be unchanged. Flash attention does not affect the output as I know. Then I figure out the expected bug from the code above and marked with annotation "Bug Here!!".
I execute the two code blocks following:
and
The result for them is following:
As you look at the model 1, the same process as dropout in scaled_dot_product_attention, it does not act correctly even though I turn on the evaluation option. So please fix this one as soon as possible. Thank you.
Versions
Collecting environment information...
PyTorch version: 2.3.0.dev20240121+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35
Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.4.0-159-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.0.140
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA GeForce RTX 3090
GPU 1: NVIDIA GeForce RTX 3090
GPU 2: NVIDIA GeForce RTX 3090
GPU 3: NVIDIA GeForce RTX 3090
Nvidia driver version: 535.154.05
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): 80
On-line CPU(s) list: 0-79
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Gold 5218R CPU @ 2.10GHz
CPU family: 6
Model: 85
Thread(s) per core: 2
Core(s) per socket: 20
Socket(s): 2
Stepping: 7
CPU max MHz: 4000.0000
CPU min MHz: 800.0000
BogoMIPS: 4200.00
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: 1.3 MiB (40 instances)
L1i cache: 1.3 MiB (40 instances)
L2 cache: 40 MiB (40 instances)
L3 cache: 55 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78
NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit: KVM: Mitigation: Split huge pages
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 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] mypy-extensions==1.0.0
[pip3] numpy==1.24.1
[pip3] pytorch-lightning==2.1.2
[pip3] pytorch-triton==2.2.0+e28a256d71
[pip3] torch==2.3.0.dev20240121+cu121
[pip3] torch-tb-profiler==0.4.3
[pip3] torchaudio==2.2.0.dev20240121+cu121
[pip3] torchmetrics==1.3.0.post0
[pip3] torchvision==0.18.0.dev20240121+cu121
[conda] Could not collect
cc @svekars @brycebortree @jbschlosser @bhosmer @cpuhrsch @erichan1 @drisspg @mikaylagawarecki
The text was updated successfully, but these errors were encountered: