10000 torch.nn.functional.conv_transpose2d produces inconsistent output on CPU and CUDA · Issue #153276 · pytorch/pytorch · GitHub
[go: up one dir, main page]
Skip to content

torch.nn.functional.conv_transpose2d produces inconsistent output on CPU and CUDA #153276

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.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
jiren-the-gray opened this issue May 9, 2025 · 1 comment
Open

torch.nn.functional.conv_transpose2d produces inconsistent output on CPU and CUDA #153276

jiren-the-gray opened this issue May 9, 2025 · 1 comment
Assignees
@CaoE
Labels
module: half Related to float16 half-precision floats module: nn Related to torch.nn topic: bug fixes topic category triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module

Comments

@jiren-the-gray
Copy link
jiren-the-gray commented May 9, 2025 •
edited by pytorch-bot bot
Loading

🐛 Describe the bug

Passing float16 tensors as the input and weight tensor with stride=1 and padding=12 produces wildly different values on CPU vs CUDA in some cases even after setting torch.backends.cudnn.deterministic = True

Colab: link

Torch version: 2.6.0+cu124

Minimal reproduction code:

import torch
import numpy as np

torch.backends.cudnn.deterministic = True

rng = np.random.default_rng(1146)
input = torch.tensor(rng.uniform(-1, 13, (9, 9, 20, 52)), dtype=torch.float16)
weight = torch.tensor(rng.uniform(-2, -1, (9, 11, 12, 84)), dtype=torch.float16)
stride = 1
padding = 12

out_cpu = torch.nn.functional.conv_transpose2d(input, weight, stride=stride, padding=padding)
out_gpu = torch.nn.functional.conv_transpose2d(input.cuda(), weight.cuda(), stride=stride, padding=padding)

torch.testing.assert_close(out_cpu, out_gpu.cpu()) # Greatest absolute difference: 736.0 at index (4, 8, 2, 70)

Versions

Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) CPU @ 2.00GHz
CPU family: 6
Model: 85
Thread(s) per core: 2
Core(s) per socket: 1
Socket(s): 1
Stepping: 3
BogoMIPS: 4000.39
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves arat md_clear arch_capabilities
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 32 KiB (1 instance)
L1i cache: 32 KiB (1 instance)
L2 cache: 1 MiB (1 instance)
L3 cache: 38.5 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0,1
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Mitigation; PTE Inversion
Vulnerability Mds: Vulnerable; SMT Host state unknown
Vulnerability Meltdown: Vulnerable
Vulnerability Mmio stale data: Vulnerable
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Vulnerable: __user pointer sanitization and usercopy barriers only; no swapgs barriers
Vulnerability Spectre v2: Vulnerable; IBPB: disabled; STIBP: disabled; PBRSB-eIBRS: Not affected; BHI: Vulnerable
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Vulnerable

Versions of relevant libraries:
[pip3] numpy==2.0.2
[pip3] nvidia-cublas-cu12==12.5.3.2
[pip3] nvidia-cuda-cupti-cu12==12.5.82
[pip3] nvidia-cuda-nvrtc-cu12==12.5.82
[pip3] nvidia-cuda-runtime-cu12==12.5.82
[pip3] nvidia-cudnn-cu12==9.3.0.75
[pip3] nvidia-cufft-cu12==11.2.3.61
[pip3] nvidia-curand-cu12==10.3.6.82
[pip3] nvidia-cusolver-cu12==11.6.3.83
[pip3] nvidia-cusparse-cu12==12.5.1.3
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.5.82
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] nvtx==0.2.11
[pip3] optree==0.15.0
[pip3] pynvjitlink-cu12==0.5.2
[pip3] torch==2.6.0+cu124
[pip3] torchaudio==2.6.0+cu124
[pip3] torchsummary==1.5.1
[pip3] torchvision==0.21.0+cu124
[pip3] triton==3.2.0
[conda] Could not collect

cc @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki
@mingfeima mingfeima added the topic: bug fixes topic category label May 12, 2025
@mingfeima
Copy link
Collaborator

@CaoE could you please help debug this one? Try to disable oneDNN and see how the results turn out?

@colesbury colesbury added module: nn Related to torch.nn triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module labels May 13, 2025
@albanD albanD added the module: half Related to float16 half-precision floats label May 14, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@CaoE CaoE

Labels
module: half Related to float16 half-precision floats module: nn Related to torch.nn topic: bug fixes topic category triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
5 participants
@colesbury @albanD @mingfeima @CaoE @jiren-the-gray
0