8000 IInconsistent Error Handling in `torch.fused_moving_avg_obs_fake_quant` Between CPU and GPU Implementations · Issue #153310 · pytorch/pytorch · GitHub
[go: up one dir, main page]
Skip to content

IInconsistent Error Handling in torch.fused_moving_avg_obs_fake_quant Between CPU and GPU Implementations #153310

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
SilentTester73 opened this issue May 10, 2025 · 4 comments · May be fixed by #153699
Open

IInconsistent Error Handling in torch.fused_moving_avg_obs_fake_quant Between CPU and GPU Implementations #153310

SilentTester73 opened this issue May 10, 2025 · 4 comments · May be fixed by #153699
Assignees
@Xia-Weiwen
Labels
module: edge cases Adversarial inputs unlikely to occur in practice oncall: quantization Quantization support in PyTorch

Comments

@SilentTester73
Copy link
SilentTester73 commented May 10, 2025 •
edited by pytorch-bot bot
Loading

🐛 Describe the bug

The torch.fused_moving_avg_obs_fake_quant function exhibits inconsistent behavior between CPU and GPU implementations. When provided with boolean tensors instead of the expected integer tensors, the CPU version silently accepts them while the GPU version raises a type error.

import torch

print(torch.__version__)

# self_cpu = list(torch.jit.load("self_cpu.pt").parameters())[0]
self_cpu = torch.tensor([ 0.0000e+00,  0.0000e+00,  0.0000e+00, 0.0000e+00,
            0.0000e+00,  0.0000e+00], dtype=torch.float32)
observer_on_cpu = torch.tensor(False)
fake_quant_on_cpu = torch.tensor(False)
running_min_cpu = torch.tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.])
running_max_cpu = torch.tensor([0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010,
        0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010])
scale_cpu = torch.tensor([0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010,
        0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010, 0.0010])
zero_point_cpu = torch.tensor([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], dtype=torch.int32)


try:
    torch.fused_moving_avg_obs_fake_quant(
        self_cpu,
        observer_on_cpu,
        fake_quant_on_cpu,
        running_min_cpu,
        running_max_cpu,
        scale_cpu,
        zero_point_cpu,
        0,
        0,
        1,
        0,
        False,
        False
    )
    print("CPU Success")
except RuntimeError as e:
    print("CPU Error:", e)

# run on gpu
self_cpu_cuda = self_cpu.cuda()
observer_on_cpu_cuda = observer_on_cpu.cuda()
fake_quant_on_cpu_cuda = fake_quant_on_cpu.cuda()
running_min_cpu_cuda = running_min_cpu.cuda()
running_max_cpu_cuda = running_max_cpu.cuda()
scale_cpu_cuda = scale_cpu.cuda()
zero_point_cpu_cuda = zero_point_cpu.cuda()

try:
    torch.fused_moving_avg_obs_fake_quant(
        self_cpu_cuda,
        observer_on_cpu_cuda,
        fake_quant_on_cpu_cuda,
        running_min_cpu_cuda,
        running_max_cpu_cuda,
        scale_cpu_cuda,
        zero_point_cpu_cuda,
        0,
        0,
        1,
        0,
        False,
        False
    )
    print("GPU Success")
except RuntimeError as e:
    print("GPU Error:", e)

Output:

2.7.0+cu126
CPU Success
GPU Error: expected scalar type Long but found Bool

Maybe a validity checker is needed for CPU implementation

Colab link: https://colab.research.google.com/drive/1SG-7Zo6UHIJ9vXwbZXAs5wSpapMClpnn?usp=sharing

Versions

Collecting environment information...
PyTorch version: 2.7.0+cu126
Is debug build: False
CUDA used to build PyTorch: 12.6
ROCM used to build PyTorch: N/A

OS: Ubuntu 24.04.1 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu224.04) 13.3.0
Clang version: 18.1.8 (++20240731025043+3b5b5c1ec4a3-1exp1~20240731145144.92)
CMake version: version 4.0.0
Libc version: glibc-2.39

Python version: 3.12.3 (main, Feb 4 2025, 14:48:35) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-6.8.0-59-generic-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 12.8.93
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090
Nvidia driver version: 570.133.20
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: 39 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: GenuineIntel
Model name: 13th Gen Intel(R) Core(TM) i9-13900F
CPU family: 6
Model: 183
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 1
Stepping: 1
CPU(s) scaling MHz: 23%
CPU max MHz: 5600.0000
CPU min MHz: 800.0000
BogoMIPS: 3993.60
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 est tm2 ssse3 sdbg fma cx16 xtpr pdcm sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect user_shstk avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq rdpid movdiri movdir64b fsrm md_clear serialize arch_lbr ibt flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 896 KiB (24 instances)
L1i cache: 1.3 MiB (24 instances)
L2 cache: 32 MiB (12 instances)
L3 cache: 36 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
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: Mitigation; Clear Register File
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] numpy==2.2.5
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] optree==0.15.0
[pip3] torch==2.7.0
[pip3] triton==3.3.0
[conda] Could not collect

cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel @msaroufim
@ngimel ngimel added the module: edge cases Adversarial inputs unlikely to occur in practice label May 11, 2025
@mingfeima
Copy link
Collaborator

@Xia-Weiwen could you please add argument checks here?

@colesbury colesbury added the oncall: quantization Quantization support in PyTorch label May 13, 2025
@Xia-Weiwen
Copy link
Collaborator

Looks like the problem is with the CUDA implementation, not the CPU implementation.
The argument observer_on is a flag that indicates whether to observe the input tensor or not. And there is no document saying it should be a Long tensor rather than Boolean. So, it totally makes sense that observer_on is a Boolean tensor.
In the CPU implementation, the value in observer_on is taken and convert to an integer (0 means false, otherwise true):

pytorch/aten/src/ATen/native/quantized/cpu/fused_obs_fake_quant.cpp

Line 159 in edc2d53

auto observe = observer_on.item().toInt();

But in the CUDA implementation, the code gets data pointer of observer_on by assuming it's data type is int64_t without any checks or conversion:

pytorch/aten/src/ATen/native/quantized/cuda/FusedObsFakeQuant.cu

Line 144 in edc2d53

int64_t* observer_on_data = observer_on.data_ptr<int64_t>();

So, I think maybe it's the CUDA implementation that needs to convert observer_on to int64_t before getting its data pointer.

Hi @mingfeima does it make sense to you? If so, I can add a conversion in the CUDA implementation. Thanks.

@Xia-Weiwen
AA38
Copy link
Collaborator

Hi @mingfeima Do you have any suggestions?

@mingfeima
Copy link
Collaborator

if so, you can just open up a pr to fix it:)

@Xia-Weiwen Xia-Weiwen linked a pull request May 16, 2025 that will close this issue
Open
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@Xia-Weiwen Xia-Weiwen

Labels
module: edge cases Adversarial inputs unlikely to occur in practice oncall: quantization Quantization support in PyTorch
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

Support boolean tensor for torch.fused_moving_avg_obs_fake_quant on CUDA Xia-Weiwen/pytorch
5 participants
@colesbury @Xia-Weiwen @ngimel @mingfeima @SilentTester73
0