pytorch - 💡(How to fix) Fix torch.compile introduces NaN in LayerNorm on valid float32 boundary inputs while eager remains finite [5 comments, 4 participants]

pytorch2026-03-22 06:37:01

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.

GitHub issue URL

Helpful · Quick feedback

GitHub stats

pytorch/pytorch#178084•Fetched 2026-04-08 01:12:14

View on GitHub

Comments

Participants

Timeline

Reactions

Author

Participants

Timeline (top)

commented ×5labeled ×3closed ×1cross-referenced ×1

Root Cause

input min=8.773236569945328e+31, max=3.4028234663852886e+38 mode cuda+float32+eager: has_nan=False, has_inf=False mode cuda+float32+compile: has_nan=True, has_inf=False Although the input contains extremely large values, these values are still valid float32 values. Therefore, this is not caused by invalid input.

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

from pathlib import Path

import torch
import torch.nn as nn
import torch._inductor.config as cfg

BASE = Path(__file__).resolve().parent
INPUT_PATH = BASE / "input.pt"
WEIGHT_PATH = BASE / "sd.pt"
HIDDEN_SIZE = 768


def make_input_and_ln():
    x = torch.load(str(INPUT_PATH), map_location="cpu").float()
    sd = torch.load(str(WEIGHT_PATH), map_location="cpu")

    ln = nn.LayerNorm(HIDDEN_SIZE, eps=1e-5).eval()
    ln.weight.data.copy_(sd["ln_weight"].float())
    ln.bias.data.copy_(sd["ln_bias"].float())
    return x, ln


def run(mode: int):
    x, ln = make_input_and_ln()
    x = x.to("cuda")
    ln = ln.to("cuda")
    if mode == 1:
        cfg.shape_padding = True
        ln = torch.compile(ln, dynamic=False)
    with torch.no_grad():
        y = ln(x).float()
    tag = "cuda+float32+eager" if mode == 0 else "cuda+float32+compile"
    print(f"input min={x.min().item()}, max={x.max().item()}")
    print(f"mode {tag}: has_nan={bool(torch.isnan(y).any())}, has_inf={bool(torch.isinf(y).any())}")


if __name__ == "__main__":
    run(0)
    run(1)

---

input min=8.773236569945328e+31, max=3.4028234663852886e+38
mode cuda+float32+eager: has_nan=False, has_inf=False
input min=8.773236569945328e+31, max=3.4028234663852886e+38
mode cuda+float32+compile: has_nan=True, has_inf=False

RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

I found a case where switching from eager execution to torch.compile causes a single LayerNorm layer to produce NaN, while eager execution on the same input and weights remains fully finite.

Observed output:

PyTorch documentation mentions that boundary values may lead to slight numerical inconsistencies across execution paths. However, this case is not a slight numerical mismatch: the compiled path produces NaN, while eager remains finite.

This is a serious issue because once NaN appears in an intermediate layer, it can propagate through downstream computation and eventually affect final model predictions.

from pathlib import Path

import torch
import torch.nn as nn
import torch._inductor.config as cfg

BASE = Path(__file__).resolve().parent
INPUT_PATH = BASE / "input.pt"
WEIGHT_PATH = BASE / "sd.pt"
HIDDEN_SIZE = 768


def make_input_and_ln():
    x = torch.load(str(INPUT_PATH), map_location="cpu").float()
    sd = torch.load(str(WEIGHT_PATH), map_location="cpu")

    ln = nn.LayerNorm(HIDDEN_SIZE, eps=1e-5).eval()
    ln.weight.data.copy_(sd["ln_weight"].float())
    ln.bias.data.copy_(sd["ln_bias"].float())
    return x, ln


def run(mode: int):
    x, ln = make_input_and_ln()
    x = x.to("cuda")
    ln = ln.to("cuda")
    if mode == 1:
        cfg.shape_padding = True
        ln = torch.compile(ln, dynamic=False)
    with torch.no_grad():
        y = ln(x).float()
    tag = "cuda+float32+eager" if mode == 0 else "cuda+float32+compile"
    print(f"input min={x.min().item()}, max={x.max().item()}")
    print(f"mode {tag}: has_nan={bool(torch.isnan(y).any())}, has_inf={bool(torch.isinf(y).any())}")


if __name__ == "__main__":
    run(0)
    run(1)

input min=8.773236569945328e+31, max=3.4028234663852886e+38
mode cuda+float32+eager: has_nan=False, has_inf=False
input min=8.773236569945328e+31, max=3.4028234663852886e+38
mode cuda+float32+compile: has_nan=True, has_inf=False

REMOVED

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

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-c [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 [pip3] torch==2.10.0+cu126 [pip3] torchaudio==2.11.0.de [pip3] torchvision==0.25.0+cu126 [pip3] triton==3.6.0+git9844da95 [conda] numpy [conda] nvidia-cublas-cu12 [conda] nvidia-cuda-cupti-cu12 [conda] nvidia-cuda-nvrtc-cu12 [conda] nvidia-cuda-runtime-cu12 [conda] nvidia-cudnn-cu12 [conda] nvidia-cufft-cu12 [conda] nvidia-curand-cu12 [conda] nvidia-cusolver-cu12 [conda] nvidia-cusparse-cu12 [conda] nvidia-cusparselt-cu12 [conda] nvidia-nccl-cu12 [conda] nvidia-nvjitlink-cu12 [conda] nvidia-nvtx-cu12 [conda] optree [conda] pytorch-triton [conda] torch [conda] torchaudio [conda] torchvision [conda] triton u12==12.6.77 +git4b3bb1f8 v20260127+cu126 1.26.4 pypi_0 pypi 12.6.4.1 pypi_0 pypi 12.6.80 pypi_0 pypi 12.6.77 pypi_0 pypi 12.6.77 pypi_0 pypi 9.10.2.21 pypi_0 pypi 11.3.0.4 pypi_0 pypi 10.3.7.77 pypi_0 pypi 11.7.1.2 pypi_0 pypi 12.5.4.2 pypi_0 pypi 0.7.1 pypi_0 pypi 2.27.5 pypi_0 pypi 12.6.85 pypi_0 pypi 12.6.77 pypi_0 pypi 0.18.0 pypi_0 pypi 3.2.0+git4b3bb1f8 pypi_0 pypi 2.10.0+cu126 pypi_0 pypi 2.11.0.dev20260127+cu126 pypi_0 pypi 0.25.0+cu126 pypi_0 pypi 3.6.0+git9844da95 pypi_0 pypi

extent analysis

Fix Plan

To address the issue of NaN values appearing in the LayerNorm layer when using torch.compile, we can try the following steps:

Clip input values: Clip the input values to prevent extremely large values from causing NaNs.
Increase epsilon value: Increase the epsilon value in the LayerNorm layer to reduce the likelihood of division by zero.
Use a different compilation mode: Try using a different compilation mode, such as dynamic=True, to see if it resolves the issue.

Here's an example of how you can modify your code to implement these changes:

import torch
import torch.nn as nn

def make_input_and_ln():
    x = torch.load(str(INPUT_PATH), map_location="cpu").float()
    sd = torch.load(str(WEIGHT_PATH), map_location="cpu")

    # Clip input values
    x = torch.clamp(x, min=-1e30, max=1e30)

    ln = nn.LayerNorm(HIDDEN_SIZE, eps=1e-3).eval()  # Increase epsilon value
    ln.weight.data.copy_(sd["ln_weight"].float())
    ln.bias.data.copy_(sd["ln_bias"].float())
    return x, ln

def run(mode: int):
    x, ln = make_input_and_ln()
    x = x.to("cuda")
    ln = ln.to("cuda")
    if mode == 1:
        cfg.shape_padding = True
        ln = torch.compile(ln, dynamic=True)  # Use dynamic compilation
    with torch.no_grad():
        y = ln(x).float()
    tag = "cuda+float32+eager" if mode == 0 else "cuda+float32+compile"
    print(f"input min={x.min().item()}, max={x.max().item()}")
    print(f"mode {tag}: has_nan={bool(torch.isnan(y).any())}, has_inf={bool(torch.isinf(y).any())}")

Verification

To verify that the fix worked, you can run the run function with both mode=0 and mode=1 and check the output for NaN values.

Extra Tips

Make sure to test your model with different input values to ensure that the fix works for all possible scenarios.
If you're still encountering issues, try reducing the precision of your model or using a different normalization layer.
Keep in mind that torch.compile is still an experimental feature, and you may need to update your PyTorch version or wait for future updates to resolve any issues.

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?

#api #API rate limit #retriever error #indexing error #inference speed

Still need to ship something?

×6

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

Data

Security

Network

Code

UI/UX

Text

System

Multimedia

Protocol

API

Engineering

pytorch - 💡(How to fix) Fix torch.compile introduces NaN in LayerNorm on valid float32 boundary inputs while eager remains finite [5 comments, 4 participants]

Recommended Tools

GitHub issue graph ai analysis

Root Cause

Fix Action

Fix / Workaround

Code Example

🐛 Describe the bug

Versions

extent analysis

Fix Plan

Verification

Extra Tips

Still need to ship something?

TRENDING

pytorch - 💡(How to fix) Fix torch.compile introduces NaN in LayerNorm on valid float32 boundary inputs while eager remains finite [5 comments, 4 participants]

Recommended Tools

GitHub issue graph ai analysis

Root Cause

Fix Action

Fix / Workaround

Code Example

🐛 Describe the bug

Versions

extent analysis

Fix Plan

Verification

Extra Tips

Still need to ship something?

RELATED_DISCOVERY

TRENDING