Backpropagation to flex_attention `score_mod` biases fails based on presence of graph breaks

[stefan-baumann]

🐛 Describe the bug

The following minimal code snippet:

import torch
from torch.nn.attention.flex_attention import flex_attention

@torch.compile()
def test(x, y):
    # Materialize a bias matrix
    B, L, device = x.shape[0], x.shape[1], x.device
    b = torch.arange(B, device=device, dtype=torch.long).view(B, 1, 1)
    q_idx = torch.arange(L, device=device, dtype=torch.long).view(1, L, 1)
    kv_idx = torch.arange(L, device=device, dtype=torch.long).view(1, 1, L)
    bias_mat = y[b, q_idx] + y[b, kv_idx] # (B, L, L)

    # Dummy score_mod retrieving bias values
    def score_mod(score, b, h, q_idx, kv_idx):
        return score + bias_mat[b, q_idx, kv_idx]

    x_ = x[:, :, None].repeat(1, 1, 16, 1)
    # torch._dynamo.graph_break()
    return flex_attention(x_, x_, x_, score_mod=score_mod)


DEVICE = "cuda"
B, L, D = 2, 16, 64

x = torch.randn(B, L, D, device=DEVICE, requires_grad=True)
y = torch.randn(B, L, device=DEVICE, requires_grad=True)

out = test(x, y).mean().backward()

print(torch.__version__)
print(f"x: {(x.grad is not None) and (x.grad.norm() > 0)}, y: {(y.grad is not None) and (y.grad.norm() > 0)}")
assert x.grad.norm() > 0
assert y.grad.norm() > 0

fails to properly backpropagate gradients into y for me (on the current stable PyTorch 2.8.0-cu12.8 installed via pip, the current nightly, and 2.8.0a0+34c6371d24.nv25.08 via the nvidia/pytorch:25.08-py3 container, across two differently configured systems with H100s & H200s).

Curiously, un-commenting it enables the code to work as expected.

The code above is a minified version of a longer snippet I started out with that had the same behavior, and various other variations of this code snippet also (seemingly randomly to me) affect whether gradients are propagated to y. Gradients are always successfully propagated to x.

Some further things I've tried (in this case in the aforementioned NGC container):

• backend="eager": works
• backend="aot_eager": fails
• backend="aot_eager_decomp_partition": fails
• backend="inductor": fails

Error logs

Running the script above produces:

2.8.0a0+34c6371d24.nv25.08
x: True, y: False
Traceback (most recent call last):
  File "/home/hpc/v104dd/v104dd11/dev/diffusion-video-pixel-diffusion-fsdp2/weird_shit.py", line 33, in <module>
    assert y.grad.norm() > 0
           ^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'norm'

indicating that gradients are (wrongfully) not backpropagated to the bias term. Removing the # torch._dynamo.graph_break() comment fixes this error.

Versions

Version with stable PyTorch 2.8:

PyTorch version: 2.8.0+cu128
Is debug build: False
CUDA used to build PyTorch: 12.8
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 3.22.1
Libc version: glibc-2.35

Python version: 3.11.11 (main, Dec 11 2024, 16:28:39) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-130-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.8.93
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H200
Nvidia driver version: 550.144.03
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.7
Is XPU available: False
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:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               384
On-line CPU(s) list:                  0-383
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 9654 96-Core Processor
CPU family:                           25
Model:                                17
Thread(s) per core:                   2
Core(s) per socket:                   96
Socket(s):                            2
Stepping:                             1
Frequency boost:                      enabled
CPU max MHz:                          3707.8120
CPU min MHz:                          1500.0000
BogoMIPS:                             4800.14
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d
Virtualization:                       AMD-V
L1d cache:                            6 MiB (192 instances)
L1i cache:                            6 MiB (192 instances)
L2 cache:                             192 MiB (192 instances)
L3 cache:                             768 MiB (24 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-95,192-287
NUMA node1 CPU(s):                    96-191,288-383
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 Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Mitigation; safe RET
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 / Automatic IBRS; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] ema-pytorch==0.7.7
[pip3] msgpack-numpy==0.4.8
[pip3] mypy-extensions==1.0.0
[pip3] numpy==2.3.2
[pip3] nvidia-cublas-cu12==12.8.4.1
[pip3] nvidia-cuda-cupti-cu12==12.8.90
[pip3] nvidia-cuda-nvrtc-cu12==12.8.93
[pip3] nvidia-cuda-runtime-cu12==12.8.90
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cufft-cu12==11.3.3.83
[pip3] nvidia-curand-cu12==10.3.9.90
[pip3] nvidia-cusolver-cu12==11.7.3.90
[pip3] nvidia-cusparse-cu12==12.5.8.93
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-nccl-cu12==2.27.3
[pip3] nvidia-nvjitlink-cu12==12.8.93
[pip3] nvidia-nvtx-cu12==12.8.90
[pip3] nvtx==0.2.11
[pip3] open_clip_torch==3.1.0
[pip3] pynvjitlink-cu12==0.5.2
[pip3] pytorch-lightning==2.5.3
[pip3] pytorch-triton==3.4.0+gitf7888497
[pip3] torch==2.8.0
[pip3] torchcodec==0.6.0
[pip3] torchdata==0.11.0
[pip3] torchmetrics==1.8.1
[pip3] torchtitan==0.1.0
[pip3] torchvision==0.23.0
[pip3] triton==3.4.0
[conda] cuda-cudart               12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cudart-dev           12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cudart-dev_linux-64  12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cudart-static        12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cudart-static_linux-64 12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cudart_linux-64      12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cupti                12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-cupti-dev            12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-libraries            12.8.1                        0    nvidia/label/cuda-12.8.1
[conda] cuda-libraries-dev        12.8.1                        0    nvidia/label/cuda-12.8.1
[conda] cuda-nvrtc                12.8.93                       0    nvidia/label/cuda-12.8.1
[conda] cuda-nvrtc-dev            12.8.93                       0    nvidia/label/cuda-12.8.1
[conda] cuda-nvtx                 12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-opencl               12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] cuda-opencl-dev           12.8.90                       0    nvidia/label/cuda-12.8.1
[conda] ema-pytorch               0.7.7                    pypi_0    pypi
[conda] libcublas                 12.8.4.1                      0    nvidia/label/cuda-12.8.1
[conda] libcublas-dev             12.8.4.1                      0    nvidia/label/cuda-12.8.1
[conda] libcufft                  11.3.3.83                     0    nvidia/label/cuda-12.8.1
[conda] libcufft-dev              11.3.3.83                     0    nvidia/label/cuda-12.8.1
[conda] libcurand                 10.3.9.90                     0    nvidia/label/cuda-12.8.1
[conda] libcurand-dev             10.3.9.90                     0    nvidia/label/cuda-12.8.1
[conda] libcusolver               11.7.3.90                     0    nvidia/label/cuda-12.8.1
[conda] libcusolver-dev           11.7.3.90                     0    nvidia/label/cuda-12.8.1
[conda] libcusparse               12.5.8.93                     0    nvidia/label/cuda-12.8.1
[conda] libcusparse-dev           12.5.8.93                     0    nvidia/label/cuda-12.8.1
[conda] libnvjitlink              12.8.93                       1    nvidia/label/cuda-12.8.1
[conda] libnvjitlink-dev          12.8.93                       1    nvidia/label/cuda-12.8.1
[conda] msgpack-numpy             0.4.8                    pypi_0    pypi
[conda] numpy                     2.3.2                    pypi_0    pypi
[conda] nvidia-cublas-cu12        12.8.4.1                 pypi_0    pypi
[conda] nvidia-cuda-cupti-cu12    12.8.90                  pypi_0    pypi
[conda] nvidia-cuda-nvrtc-cu12    12.8.93                  pypi_0    pypi
[conda] nvidia-cuda-runtime-cu12  12.8.90                  pypi_0    pypi
[conda] nvidia-cudnn-cu12         9.10.2.21                pypi_0    pypi
[conda] nvidia-cufft-cu12         11.3.3.83                pypi_0    pypi
[conda] nvidia-curand-cu12        10.3.9.90                pypi_0    pypi
[conda] nvidia-cusolver-cu12      11.7.3.90                pypi_0    pypi
[conda] nvidia-cusparse-cu12      12.5.8.93                pypi_0    pypi
[conda] nvidia-cusparselt-cu12    0.7.1                    pypi_0    pypi
[conda] nvidia-nccl-cu12          2.27.3                   pypi_0    pypi
[conda] nvidia-nvjitlink-cu12     12.8.93                  pypi_0    pypi
[conda] nvidia-nvtx-cu12          12.8.90                  pypi_0    pypi
[conda] nvtx                      0.2.11                   pypi_0    pypi
[conda] open-clip-torch           3.1.0                    pypi_0    pypi
[conda] pynvjitlink-cu12          0.5.2                    pypi_0    pypi
[conda] pytorch-lightning         2.5.3                    pypi_0    pypi
[conda] pytorch-triton            3.4.0+gitf7888497          pypi_0    pypi
[conda] torch                     2.8.0                    pypi_0    pypi
[conda] torchcodec                0.6.0                    pypi_0    pypi
[conda] torchdata                 0.11.0                   pypi_0    pypi
[conda] torchmetrics              1.8.1                    pypi_0    pypi
[conda] torchtitan                0.1.0                    pypi_0    pypi
[conda] torchvision               0.23.0                   pypi_0    pypi
[conda] triton                    3.4.0                    pypi_0    pypi

Version with NGC Container:

PyTorch version: 2.8.0a0+34c6371d24.nv25.08
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.2 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.39

Python version: 3.12.3 (main, Jun 18 2025, 17:59:45) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-5.14.0-570.35.1.el9_6.x86_64-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 13.0.48
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H100
Nvidia driver version: 580.65.06
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.12.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.12.0
Is XPU available: False
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:                           52 bits physical, 57 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  128
On-line CPU(s) list:                     0-127
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 9554 64-Core Processor
CPU family:                              25
Model:                                   17
Thread(s) per core:                      1
Core(s) per socket:                      64
Socket(s):                               2
Stepping:                                1
Frequency boost:                         enabled
CPU(s) scaling MHz:                      93%
CPU max MHz:                             3762.9880
CPU min MHz:                             1500.0000
BogoMIPS:                                6200.54
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good amd_lbr_v2 nopl xtopology nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba perfmon_v2 ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d debug_swap
L1d cache:                               4 MiB (128 instances)
L1i cache:                               4 MiB (128 instances)
L2 cache:                                128 MiB (128 instances)
L3 cache:                                512 MiB (16 instances)
NUMA node(s):                            8
NUMA node0 CPU(s):                       0-15
NUMA node1 CPU(s):                       16-31
NUMA node2 CPU(s):                       32-47
NUMA node3 CPU(s):                       48-63
NUMA node4 CPU(s):                       64-79
NUMA node5 CPU(s):                       80-95
NUMA node6 CPU(s):                       96-111
NUMA node7 CPU(s):                       112-127
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: 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 Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Mitigation; Safe RET
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; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsx async abort:           Not affected

Versions of relevant libraries:
[pip3] ema-pytorch==0.7.7
[pip3] intel-openmp==2021.4.0
[pip3] mkl==2021.1.1
[pip3] mkl-devel==2021.1.1
[pip3] mkl-include==2021.1.1
[pip3] mypy_extensions==1.1.0
[pip3] numpy==1.26.4
[pip3] nvidia-cudnn-frontend==1.13.0
[pip3] nvtx==0.2.11
[pip3] onnx==1.18.0
[pip3] optree==0.17.0
[pip3] pynvjitlink==0.7.0
[pip3] pytorch-triton==3.3.1+gitc8757738
[pip3] tbb==2021.13.1
[pip3] torch==2.8.0a0+34c6371d24.nv25.8
[pip3] torch_tensorrt==2.8.0a0
[pip3] torchao==0.12.0+git
[pip3] torchdata==0.11.0
[pip3] torchdiffeq==0.2.5
[pip3] torchprofile==0.0.4
[pip3] torchtitan==0.1.0
[pip3] torchvision==0.23.0a0+428a54c9
[conda] Could not collect

cc @ezyang @gchanan @zou3519 @kadeng @msaroufim @chauhang @penguinwu @bdhirsh

[stefan-baumann]

I also just threw it into a default Google Colab notebook with a T4 GPU just to be sure, and the same problem is reproducible, although the behavior is slightly different, and seems to vary across different instances.

Attempt 1:

• # torch._dynamo.graph_break(): fails
• torch._dynamo.graph_break(): fails
• Deleting that line: works

Attempt 2:

• # torch._dynamo.graph_break(): fails
• torch._dynamo.graph_break(): works
• Deleting that line: fails

Attempt 2 is somewhat sane, but attempt 1 has a complete lack of sanity in its behavior.

[williamwen42]

cc @bdhirsh for dispatcher correctness - my initial guess is that were failing to handle the complicated indexing.

[stefan-baumann]

I actually found a second, likely related bug:

import random
import torch

@torch.compile()
def test(x):
    offset = random.uniform(-1, 1)
    # torch._dynamo.graph_break()
    return x + offset


for _ in range(5):
    x = torch.randn(1)
    print(test(x) - x)

Running this (either with fullgraph=True or not) on CPU tensors results in random offsets outside the [-1, 1] range:

tensor([7528.3672])
tensor([3.9518e+15])
tensor([0.])
tensor([4.6613e+26])
tensor([0.])

Adding the graph break makes this work again:

tensor([0.2417])
tensor([0.1916])
tensor([-0.0258])
tensor([-0.2303])
tensor([0.3440])

I would've expected, for the original code, one of two things to happen:

1. It works with a random offset in [-1, 1] per forward.
2. It works with a random offset in [-1, 1] that is drawn once during compilation and then fixed afterwards.

Depending on the presence of the graph break, it seems to change between case 1 and case "3" - adding either 0 or some completely random number, which I assume is caused by it reading from unallocated memory?

I think either 1 or 2 would be reasonable behavior. On GPU tensors, it seems to work as in case 1 consistently for me, whether the graph break is there or not.

I'll file a separate issue for this later once I find the time, but thought I'd drop this here already.

[williamwen42]

cc @drisspg