Skip to content

numeric bug in CUDNN_ATTENTION, NaN #166211

New issue
New issue
Closed
Closed
numeric bug in CUDNN_ATTENTION, NaN#166211
Assignees
eqy
Labels
high prioritymodule: correctness (silent)issue that returns an incorrect result silentlymodule: cudnnRelated to torch.backends.cudnn, and CuDNN supportmodule: sdpaAll things related to torch.nn.functional.scaled_dot_product_attentiiontriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module
Milestone
2.9.1
@ad8e

Description

@ad8e
ad8e
opened on Oct 24, 2025

🐛 Describe the bug

The grad of q in F.scaled_dot_product_attention(q, k, v) NaNs occasionally. The activations are normal-sized, and the softmax logits are between -131 to 27, so this is a numeric bug.

I saved the gradients and activations to a file. q is of shape [20, 4, 4, 32], and only some terms in q.grad are NaNs; the other values are all normal. k and v's gradients are also normal. In every run, it's always q which explodes, although this may be caused by differences in how my codebase treats k and q.

test_sdpa_nan2.py
sdpa_nan_debug_Transformer_res-6.pt.txt - rename the extension to pt

Output:

================================================================================
SDPA NaN Gradient Test Case
================================================================================

Loading debug file: sdpa_nan_debug_Transformer_res-6.pt
Debug info from module: Transformer_res-6
NaN detected in: grad_q

=== Tensor Shapes ===
q: torch.Size([20, 4, 4, 32])
k: torch.Size([20, 4, 4, 32])
v: torch.Size([20, 4, 4, 32])
attn_output: torch.Size([20, 4, 4, 32])

=== Forward Pass Statistics ===
q - min: -9.187500e+00, max: 1.031250e+01, mean: 3.564453e-02, std: 3.281250e+00
q - has NaN: False, has Inf: False, NaN count: 0
k - min: -1.037500e+01, max: 9.625000e+00, mean: 3.247070e-02, std: 3.140625e+00
k - has NaN: False, has Inf: False, NaN count: 0
v - min: -3.417969e-01, max: 3.300781e-01, mean: 4.577637e-03, std: 1.064453e-01
v - has NaN: False, has Inf: False, NaN count: 0
attn_output - min: -2.158203e-01, max: 1.806641e-01, mean: 7.385254e-03, std: 7.275391e-02
attn_output - has NaN: False, has Inf: False, NaN count: 0

=== Gradient Statistics ===
grad_attn_output - min: -8.583069e-06, max: 8.165836e-06, mean: -7.566996e-09, std: 1.601875e-06
grad_attn_output - has NaN: False, has Inf: False, NaN count: 0
grad_q - min: -1.427907e-10, max: 1.518856e-10, mean: 1.803002e-13, std: 1.057288e-11
grad_q - has NaN: True, has Inf: False, NaN count: 640
grad_k - min: -1.182343e-10, max: 1.145963e-10, mean: -1.945111e-13, std: 8.810730e-12
grad_k - has NaN: False, has Inf: False, NaN count: 0
grad_v - min: -1.126528e-05, max: 1.096725e-05, mean: -7.566996e-09, std: 1.400709e-06
grad_v - has NaN: False, has Inf: False, NaN count: 0

=== Reconstructing Forward Pass ===

=== Manual Attention Score Calculation ===
Attention scores shape: torch.Size([20, 4, 4, 4])
attn_scores (Q@K^T / sqrt(d)) - min: -1.310000e+02, max: 2.737500e+01, mean: -2.475000e+01, std: 4.475000e+01
attn_scores (Q@K^T / sqrt(d)) - has NaN: False, has Inf: False, NaN count: 0
Attention weights shape: torch.Size([20, 4, 4, 4])
attn_weights (softmax) - min: 5.816114e-24, max: 1.000000e+00, mean: 2.500000e-01, std: 4.335938e-01
attn_weights (softmax) - has NaN: False, has Inf: False, NaN count: 0
Manual attention output shape: torch.Size([20, 4, 4, 32])
attn_output_manual - min: -2.158203e-01, max: 1.806641e-01, mean: 7.385254e-03, std: 7.275391e-02
attn_output_manual - has NaN: False, has Inf: False, NaN count: 0

=== SDPA Forward Pass ===
qkv shape torch.Size([20, 4, 4, 32]) torch.Size([20, 4, 4, 32]) torch.Size([20, 4, 4, 32])
qkv strides (512, 32, 128, 1) (512, 32, 128, 1) (512, 32, 128, 1)
Reconstructed attn_output - has NaN: False
Reconstructed attn_output - min: -2.158203e-01, max: 1.806641e-01, mean: 7.385254e-03, std: 7.275391e-02
Reconstructed attn_output - has NaN: False, has Inf: False, NaN count: 0
attn_output difference (ignoring NaN) - max: 0.000000e+00, mean: 0.000000e+00

=== Reconstructing Backward Pass ===
Reconstructed grad_q - has NaN: True
Reconstructed grad_k - has NaN: False
Reconstructed grad_v - has NaN: False

grad_q difference (ignoring NaN) - max: 0.000000e+00, mean: 0.000000e+00
grad_q difference - NaN count: 640
grad_k difference (ignoring NaN) - max: 0.000000e+00, mean: 0.000000e+00
grad_v difference (ignoring NaN) - max: 0.000000e+00, mean: 0.000000e+00

=== Test Complete ===
You can now inspect the tensors in debug_data for further analysis.

================================================================================

Running minimal reproduction test...
================================================================================
Loading debug file: sdpa_nan_debug_Transformer_res-6.pt
Forward output has NaN: False
grad_q has NaN: True
grad_k has NaN: False
grad_v has NaN: False

✓ Successfully reproduced NaN gradient!

You can switch attention variant by changing the attn_output_reconstructed calculation to this snippet, to see that the FLASH_ATTENTION backend has no such error:

    from torch.utils.flop_counter import FlopCounterMode
    from torch.nn.attention import SDPBackend, sdpa_kernel

    counter = FlopCounterMode()
    with sdpa_kernel(SDPBackend.FLASH_ATTENTION), counter:
        attn_output_reconstructed = F.scaled_dot_product_attention(q_test, k_test, v_test)

(thanks to mahouko for the clever flop counter trick to print the backend)

Versions

PyTorch version: 2.9.0
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0
Clang version: Could not collect
CMake version: version 4.1.0
Libc version: glibc-2.35

Python version: 3.10.12 (main, Aug 15 2025, 14:32:43) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.13-65-650-4141-22041-coreweave-amd64-85c45edc-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 13.0.88
CUDA_MODULE_LOADING set to: 
GPU models and configuration: GPU 0: NVIDIA H100 80GB HBM3
Nvidia driver version: 570.172.08
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.13.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.13.0
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, 57 bits virtual
Byte Order:                         Little Endian
CPU(s):                             128
On-line CPU(s) list:                0-127
Vendor ID:                          GenuineIntel
Model name:                         Intel(R) Xeon(R) Platinum 8462Y+
CPU family:                         6
Model:                              143
Thread(s) per core:                 2
Core(s) per socket:                 32
Socket(s):                          2
Stepping:                           8
CPU max MHz:                        4100.0000
CPU min MHz:                        800.0000
BogoMIPS:                           5600.00
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 tsc_known_freq 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 cat_l2 cdp_l3 invpcid_single cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hfi vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                     VT-x
L1d cache:                          3 MiB (64 instances)
L1i cache:                          2 MiB (64 instances)
L2 cache:                           128 MiB (64 instances)
L3 cache:                           120 MiB (2 instances)
NUMA node(s):                       2
NUMA node0 CPU(s):                  0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110,112,114,116,118,120,122,124,126
NUMA node1 CPU(s):                  1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79,81,83,85,87,89,91,93,95,97,99,101,103,105,107,109,111,113,115,117,119,121,123,125,127
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Not affected
Vulnerability Retbleed:             Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:           Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] clip-anytorch==2.6.0
[pip3] dctorch==0.1.2
[pip3] DISTS-pytorch==0.1
[pip3] gpytorch==1.14.2
[pip3] lovely-numpy==0.2.16
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.24.4
[pip3] onnx==1.19.1
[pip3] onnx-ir==0.1.11
[pip3] onnxscript==0.3.1
[pip3] torch==2.9.0
[pip3] torchaudio==2.9.0
[pip3] torchdata==0.11.0
[pip3] torchdiffeq==0.2.5
[pip3] torchsde==0.2.6
[pip3] torchtitan==0.1.0
[pip3] torchvision==0.24.0
[pip3] triton==3.5.0+gitbbb06c03
[pip3] welford-torch==0.2.5
[conda] Could not collect

cc @ezyang @gchanan @kadeng @msaroufim @csarofeen @ptrblck @xwang233 @eqy

Metadata

Metadata

Assignees

Labels

high prioritymodule: correctness (silent)issue that returns an incorrect result silentlymodule: cudnnRelated to torch.backends.cudnn, and CuDNN supportmodule: sdpaAll things related to torch.nn.functional.scaled_dot_product_attentiiontriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

Type

No type

Projects

PyTorch + CUDA

Status

Done

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions