Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix CUDA fp32: eager vs torch.compile mismatch on single Conv2d [6 comments, 4 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
pytorch/pytorch#178247•Fetched 2026-04-08 01:21:05
View on GitHub
Comments
6
Participants
4
Timeline
14
Reactions
0
Author
hiahu329hiahu329
Participants
albanDalbanDdesertfiredesertfiregithub-actions[bot]github-actions[bot]hiahu329hiahu329
Timeline (top)
commented ×6labeled ×3mentioned ×2subscribed ×2

Fix Action

Fix / Workaround

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 40 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 48 On-line CPU(s) list: 0-47 Vendor ID: GenuineIntel Model name: QEMU Virtual CPU version 2.5+ CPU family: 15 Model: 107 Thread(s) per core: 1 Core(s) per socket: 48 Socket(s): 1 Stepping: 1 BogoMIPS: 4190.15 Flags: fpu de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx lm constant_tsc nopl xtopology cpuid tsc_known_freq pni ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c hypervisor lahf_lm abm cpuid_fault pti bmi1 avx2 bmi2 avx512f avx512dq avx512cd avx512bw avx512vl Hypervisor vendor: KVM Virtualization type: full L1d cache: 1.5 MiB (48 instances) L1i cache: 1.5 MiB (48 instances) L2 cache: 192 MiB (48 instances) L3 cache: 16 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-47 Vulnerability Gather data sampling: Not affected Vulnerability Itlb multihit: KVM: Mitigation: VMX unsupported Vulnerability L1tf: Mitigation; PTE Inversion Vulnerability Mds: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Unknown: No mitigations Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Vulnerable Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Retpoline Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Not affected

Code Example

import os
import contextlib
import torch
import torch.nn as nn
import torch._inductor.config as inductor_config

EPS = 1e-12
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
INPUT_PATH = os.path.join(BASE_DIR, "input.pt")
SD_PATH = os.path.join(BASE_DIR, "sd.pt")

DEVICE = "cuda"
DTYPE = torch.float32

COMPILE_EPILOGUE_FUSION = True

@contextlib.contextmanager
def set_epilogue_fusion(enabled: bool):
    old_epilogue = inductor_config.epilogue_fusion
    try:
        inductor_config.epilogue_fusion = enabled
        yield
    finally:
        inductor_config.epilogue_fusion = old_epilogue


def build_conv(sd: dict, compile_flag: bool):
    conv = nn.Conv2d(3, 96, kernel_size=3, stride=2, padding=1, bias=True).to(device=DEVICE, dtype=DTYPE).eval()
    with torch.no_grad():
        conv.weight.copy_(sd["weight"].to(device=DEVICE, dtype=DTYPE))
        conv.bias.copy_(sd["bias"].to(device=DEVICE, dtype=DTYPE))
    if compile_flag:
        conv = torch.compile(conv)
    return conv


def run_conv(x_cpu: torch.Tensor, sd: dict, compile_flag: bool) -> torch.Tensor:
    with set_epilogue_fusion(COMPILE_EPILOGUE_FUSION if compile_flag else False):
        conv = build_conv(sd, compile_flag)
    with torch.no_grad(), set_epilogue_fusion(COMPILE_EPILOGUE_FUSION if compile_flag else False):
        y = conv(x_cpu.to(device=DEVICE, dtype=DTYPE))
    return y.detach().float().cpu()


def stats(ref: torch.Tensor, new: torch.Tensor):
    abs_max = (new - ref).abs().max().item()
    ref_f = ref.reshape(ref.shape[0], -1)
    new_f = new.reshape(new.shape[0], -1)
    valid = torch.isfinite(ref_f) & torch.isfinite(new_f)
    rel = torch.where(valid, (new_f - ref_f).abs() / (ref_f.abs() + EPS), torch.zeros_like(ref_f))
    rel_max = rel.max().item()
    return abs_max, rel_max
def has_nan_inf(t: torch.Tensor):
    return bool(torch.isnan(t).any().item()), bool(torch.isinf(t).any().item())


def main():
    x = torch.load(INPUT_PATH, map_location="cpu").float()
    sd = torch.load(SD_PATH, map_location="cpu")
    y_eager = run_conv(x, sd, compile_flag=False)
    y_compile = run_conv(x, sd, compile_flag=True)

    abs_max, rel_max = stats(y_eager, y_compile)
    on, oi = has_nan_inf(y_eager)
    sn, si = has_nan_inf(y_compile)

    print("compare=eager VS compile")
    print(f"input shape={tuple(x.shape)}")
    print(f"conv1 output shape={tuple(y_eager.shape)}")
    print(f"compile_epilogue_fusion={COMPILE_EPILOGUE_FUSION}")
    print(
        f"conv1 abs_max={abs_max:.6e} rel_max={rel_max:.6e} "
        f"eager(nan={on},inf={oi}) compile(nan={sn},inf={si})"
    )


if __name__ == "__main__":
    main()

---

compare=eager VS compile
input shape=(64, 3, 32, 32)
conv1 output shape=(64, 96, 16, 16)
compile_epilogue_fusion=True
conv1 abs_max=7.152557e-07 rel_max=1.042344e-01 eager(nan=False,inf=False) compile(nan=False,inf=False)
RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

I can consistently reproduce a numerical mismatch between CUDA FP32 eager and torch.compile on a minimal Conv2d-only repro.

Both modes produce finite outputs (no NaN/Inf), but there is still a non-zero output drift between eager and compile.

Is this behavior expected for CUDA FP32, or should this be considered an accuracy issue?

import os
import contextlib
import torch
import torch.nn as nn
import torch._inductor.config as inductor_config

EPS = 1e-12
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
INPUT_PATH = os.path.join(BASE_DIR, "input.pt")
SD_PATH = os.path.join(BASE_DIR, "sd.pt")

DEVICE = "cuda"
DTYPE = torch.float32

COMPILE_EPILOGUE_FUSION = True

@contextlib.contextmanager
def set_epilogue_fusion(enabled: bool):
    old_epilogue = inductor_config.epilogue_fusion
    try:
        inductor_config.epilogue_fusion = enabled
        yield
    finally:
        inductor_config.epilogue_fusion = old_epilogue


def build_conv(sd: dict, compile_flag: bool):
    conv = nn.Conv2d(3, 96, kernel_size=3, stride=2, padding=1, bias=True).to(device=DEVICE, dtype=DTYPE).eval()
    with torch.no_grad():
        conv.weight.copy_(sd["weight"].to(device=DEVICE, dtype=DTYPE))
        conv.bias.copy_(sd["bias"].to(device=DEVICE, dtype=DTYPE))
    if compile_flag:
        conv = torch.compile(conv)
    return conv


def run_conv(x_cpu: torch.Tensor, sd: dict, compile_flag: bool) -> torch.Tensor:
    with set_epilogue_fusion(COMPILE_EPILOGUE_FUSION if compile_flag else False):
        conv = build_conv(sd, compile_flag)
    with torch.no_grad(), set_epilogue_fusion(COMPILE_EPILOGUE_FUSION if compile_flag else False):
        y = conv(x_cpu.to(device=DEVICE, dtype=DTYPE))
    return y.detach().float().cpu()


def stats(ref: torch.Tensor, new: torch.Tensor):
    abs_max = (new - ref).abs().max().item()
    ref_f = ref.reshape(ref.shape[0], -1)
    new_f = new.reshape(new.shape[0], -1)
    valid = torch.isfinite(ref_f) & torch.isfinite(new_f)
    rel = torch.where(valid, (new_f - ref_f).abs() / (ref_f.abs() + EPS), torch.zeros_like(ref_f))
    rel_max = rel.max().item()
    return abs_max, rel_max
def has_nan_inf(t: torch.Tensor):
    return bool(torch.isnan(t).any().item()), bool(torch.isinf(t).any().item())


def main():
    x = torch.load(INPUT_PATH, map_location="cpu").float()
    sd = torch.load(SD_PATH, map_location="cpu")
    y_eager = run_conv(x, sd, compile_flag=False)
    y_compile = run_conv(x, sd, compile_flag=True)

    abs_max, rel_max = stats(y_eager, y_compile)
    on, oi = has_nan_inf(y_eager)
    sn, si = has_nan_inf(y_compile)

    print("compare=eager VS compile")
    print(f"input shape={tuple(x.shape)}")
    print(f"conv1 output shape={tuple(y_eager.shape)}")
    print(f"compile_epilogue_fusion={COMPILE_EPILOGUE_FUSION}")
    print(
        f"conv1 abs_max={abs_max:.6e} rel_max={rel_max:.6e} "
        f"eager(nan={on},inf={oi}) compile(nan={sn},inf={si})"
    )


if __name__ == "__main__":
    main()
compare=eager VS compile
input shape=(64, 3, 32, 32)
conv1 output shape=(64, 96, 16, 16)
compile_epilogue_fusion=True
conv1 abs_max=7.152557e-07 rel_max=1.042344e-01 eager(nan=False,inf=False) compile(nan=False,inf=False)

Versions

PyTorch version: 2.10.0+cu126 Is debug build: False CUDA used to build PyTorch: 12.6 ROCm used to build PyTorch: N/A

OS: Ubuntu 24.04.3 LTS (x86_64) GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 Clang version: Could not collect CMake version: Could not collect Libc version: glibc-2.39

Python version: 3.10.19 | packaged by conda-forge | (main, Jan 26 2026, 23:45:08) [GCC 14.3.0] (64-bit runtime) Python platform: Linux-6.8.0-90-generic-x86_64-with-glibc2.39 Is CUDA available: True CUDA runtime version: 12.6.20 CUDA_MODULE_LOADING set to: GPU models and configuration: GPU 0: NVIDIA GeForce RTX 3090 GPU 1: NVIDIA GeForce RTX 3090

Nvidia driver version: 560.35.03 cuDNN version: Could not collect Is XPU available: False HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True Caching allocator config: N/A

CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 40 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 48 On-line CPU(s) list: 0-47 Vendor ID: GenuineIntel Model name: QEMU Virtual CPU version 2.5+ CPU family: 15 Model: 107 Thread(s) per core: 1 Core(s) per socket: 48 Socket(s): 1 Stepping: 1 BogoMIPS: 4190.15 Flags: fpu de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx lm constant_tsc nopl xtopology cpuid tsc_known_freq pni ssse3 fma cx16 sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c hypervisor lahf_lm abm cpuid_fault pti bmi1 avx2 bmi2 avx512f avx512dq avx512cd avx512bw avx512vl Hypervisor vendor: KVM Virtualization type: full L1d cache: 1.5 MiB (48 instances) L1i cache: 1.5 MiB (48 instances) L2 cache: 192 MiB (48 instances) L3 cache: 16 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-47 Vulnerability Gather data sampling: Not affected Vulnerability Itlb multihit: KVM: Mitigation: VMX unsupported Vulnerability L1tf: Mitigation; PTE Inversion Vulnerability Mds: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown Vulnerability Meltdown: Mitigation; PTI Vulnerability Mmio stale data: Unknown: No mitigations Vulnerability Reg file data sampling: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Vulnerable Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Retpoline Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Not affected

Versions of relevant libraries: [pip3] numpy==1.26.4 [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==9.10.2.21 [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==0.7.1 [pip3] nvidia-nccl-cu12==2.27.5 [pip3] nvidia-nvjitlink-cu12==12.6.85 [pip3] nvidia-nvtx-cu12==12.6.77 [pip3] onnxruntime-gpu==1.23.2 [pip3] optree==0.18.0 [pip3] pytorch-triton==3.2.0+git4b3bb1f8 [pip3] torch==2.10.0+cu126 [pip3] torchaudio==2.11.0.dev20260127+cu126 [pip3] torchvision==0.25.0+cu126 [pip3] triton==3.6.0+git9844da95 [conda] numpy 1.26.4 pypi_0 pypi [conda] nvidia-cublas-cu12 12.6.4.1 pypi_0 pypi [conda] nvidia-cuda-cupti-cu12 12.6.80 pypi_0 pypi [conda] nvidia-cuda-nvrtc-cu12 12.6.77 pypi_0 pypi [conda] nvidia-cuda-runtime-cu12 12.6.77 pypi_0 pypi [conda] nvidia-cudnn-cu12 9.10.2.21 pypi_0 pypi [conda] nvidia-cufft-cu12 11.3.0.4 pypi_0 pypi [conda] nvidia-curand-cu12 10.3.7.77 pypi_0 pypi [conda] nvidia-cusolver-cu12 11.7.1.2 pypi_0 pypi [conda] nvidia-cusparse-cu12 12.5.4.2 pypi_0 pypi [conda] nvidia-cusparselt-cu12 0.7.1 pypi_0 pypi [conda] nvidia-nccl-cu12 2.27.5 pypi_0 pypi [conda] nvidia-nvjitlink-cu12 12.6.85 pypi_0 pypi [conda] nvidia-nvtx-cu12 12.6.77 pypi_0 pypi [conda] optree 0.18.0 pypi_0 pypi [conda] pytorch-triton 3.2.0+git4b3bb1f8 pypi_0 pypi [conda] torch 2.10.0+cu126 pypi_0 pypi [conda] torchaudio 2.11.0.dev20260127+cu126 pypi_0 pypi [conda] torchvision 0.25.0+cu126 pypi_0 pypi [conda] triton 3.6.0+git9844da95 pypi_0 pypi

cc @chauhang @penguinwu

extent analysis

Fix Plan

To address the numerical mismatch between CUDA FP32 eager and torch.compile, we can try the following steps:

  • Disable epilogue fusion: Set COMPILE_EPILOGUE_FUSION to False to see if it resolves the issue.
  • Use a different convolution implementation: Try using torch.nn.functional.conv2d instead of torch.nn.Conv2d to see if the issue persists.
  • Verify input and weight data: Ensure that the input and weight data are identical for both eager and compiled modes.

Here's an example code snippet that demonstrates these steps:

# Disable epilogue fusion
COMPILE_EPILOGUE_FUSION = False

# Use a different convolution implementation
def build_conv(sd: dict, compile_flag: bool):
    if compile_flag:
        # Use torch.nn.functional.conv2d for compiled mode
        def conv(x):
            return torch.nn.functional.conv2d(x, sd["weight"], bias=sd["bias"], stride=2, padding=1)
        conv = torch.compile(conv)
    else:
        # Use torch.nn.Conv2d for eager mode
        conv = nn.Conv2d(3, 96, kernel_size=3, stride=2, padding=1, bias=True).to(device=DEVICE, dtype=DTYPE).eval()
        with torch.no_grad():
            conv.weight.copy_(sd["weight"].to(device=DEVICE, dtype=DTYPE))
            conv.bias.copy_(sd["bias"].to(device=DEVICE, dtype=DTYPE))
    return conv

# Verify input and weight data
def run_conv(x_cpu: torch.Tensor, sd: dict, compile_flag: bool) -> torch.Tensor:
    with set_epilogue_fusion(COMPILE_EPILOGUE_FUSION if compile_flag else False):
        conv = build_conv(sd, compile_flag)
    with torch.no_grad(), set_epilogue_fusion(COMPILE_EPILOGUE_FUSION if compile_flag else False):
        if compile_flag:
            y = conv(x_cpu.to(device=DEVICE, dtype=DTYPE))
        else:
            y = conv(x_cpu.to(device=DEVICE, dtype=DTYPE))
    return y.detach().float().cpu()

Verification

To verify that the fix worked, you can compare the output of the eager and compiled modes using the stats function:

y_eager = run_conv(x, sd, compile_flag=False)
y_compile = run_conv(x, sd, compile_flag=True)

abs_max, rel_max = stats(y_eager, y_compile)
print(f"conv1 abs_max={abs_max:.6e} rel_max={rel_max:.6e}")

If the abs_max and rel_max values are close to zero, it indicates that the numerical mismatch has been resolved.

Extra Tips

  • Ensure that the input and weight data

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#ISR setup#authentication setup#request error#file not found

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING