VYPR
← All advisories
Moderate severityNVD Advisory· Published Jan 3, 2024· Updated Jun 6, 2025

FPE in paddle.nanmedian

CVE-2023-38674

Description

FPE in paddle.nanmedian in PaddlePaddle before 2.6.0. This flaw can cause a runtime crash and a denial of service.

AI Insight

LLM-synthesized narrative grounded in this CVE's description and references.

A floating-point exception in PaddlePaddle's `paddle.nanmedian` function before 2.6.0 causes a crash and denial of service when the input tensor has a zero-size dimension.

A floating-point exception (FPE) vulnerability exists in the paddle.nanmedian function of PaddlePaddle, an open-source deep learning platform. The root cause is a divide-by-zero operation that occurs when the function processes an input tensor whose last dimension size is zero [1][2]. The commit that fixes the issue adds a check to enforce that the stride (i.e., the size of the last dimension) is not zero, throwing an InvalidArgument error instead of performing the division [4]. This flaw affects versions before 2.6.0 [2].

To exploit this vulnerability, an attacker would need to supply a crafted tensor with a zero-length dimension to the paddle.nanmedian function. No special privileges or prior authentication are required beyond the ability to interact with the PaddlePaddle API [1][2]. The attack vector is local, but could be triggered indirectly if a user loads a malicious model or data that calls the vulnerable function [2].

If successfully triggered, the division by zero causes a runtime crash, leading to a denial of service (DoS) condition [2]. The impact is limited to availability; there is no evidence of data corruption or privilege escalation. The vulnerability is listed in the PyPA advisory database with a severity score of 7.5 (CVSS 3.1) [3].

A fix was merged into the PaddlePaddle GitHub repository on December 6, 2023, via commit 690ffe814dbfc5054d4e92df878687fd638fe3a5, and is included in the 2.6.0 release [4]. Users are strongly advised to upgrade to PaddlePaddle 2.6.0 or later. There is no known workaround for earlier versions beyond upgrading [2].

References
  1. GitHub - PaddlePaddle/Paddle: PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）
  2. NVD - CVE-2023-38674
  3. advisory-database/vulns/paddlepaddle/PYSEC-2024-129.yaml at main · pypa/advisory-database
  4. fix div 0 bug (#55644) · PaddlePaddle/Paddle@690ffe8

AI Insight generated on May 20, 2026. Synthesized from this CVE's description and the cited reference URLs; citations are validated against the source bundle.

Affected packages

Versions sourced from the GitHub Security Advisory.

PackageAffected versionsPatched versions
PaddlePaddlePyPI
< 2.6.02.6.0

Affected products

2

Patches

1
690ffe814dbf

fix div 0 bug (#55644)

https://github.com/PaddlePaddle/PaddlewanghuancoderJul 25, 2023via ghsa
commit
3 files changed · +28 −0
  • paddle/phi/kernels/cpu/nanmedian_kernel.cc+8 −0 modified
    @@ -116,6 +116,14 @@ void ProcessMedianKernel(const Context& dev_ctx,
   auto x_dim = x.dims();
   int64_t x_rank = x_dim.size();
   int64_t stride = x_dim[x_rank - 1];
+
+  PADDLE_ENFORCE_NE(
+      stride,
+      0,
+      phi::errors::InvalidArgument("The input Tensor x's shape[-1] should not "
+                                   "be 0, but shape is %s now.",
+                                   x_dim));
+
   int64_t pre_dim = numel / stride;
   int64_t i = 0;
  • paddle/phi/kernels/gpu/matrix_rank_tol_kernel.cu+12 −0 modified
    @@ -340,6 +340,18 @@ void MatrixRankTolKernel(const Context& dev_ctx,
   auto dim_out = out->dims();
   int rows = dim_x[dim_x.size() - 2];
   int cols = dim_x[dim_x.size() - 1];
+  PADDLE_ENFORCE_NE(
+      rows,
+      0,
+      phi::errors::InvalidArgument("The input Tensor x's shape[-2] should not "
+                                   "be 0, but shape is %s now.",
+                                   dim_x));
+  PADDLE_ENFORCE_NE(
+      cols,
+      0,
+      phi::errors::InvalidArgument("The input Tensor x's shape[-1] should not "
+                                   "be 0, but shape is %s now.",
+                                   dim_x));
   int k = std::min(rows, cols);
   auto numel = x.numel();
   int batches = numel / (rows * cols);
  • paddle/phi/kernels/gpu/nanmedian_kernel.cu+8 −0 modified
    @@ -149,6 +149,14 @@ void ProcessMedianKernel(const Context& dev_ctx,
   auto x_dim = x.dims();
   int64_t x_rank = x_dim.size();
   int64_t stride = x_dim[x_rank - 1];
+
+  PADDLE_ENFORCE_NE(
+      stride,
+      0,
+      phi::errors::InvalidArgument("The input Tensor x's shape[-1] should not "
+                                   "be 0, but shape is %s now.",
+                                   x_dim));
+
   int64_t pre_dim = numel / stride;
   int64_t i = 0;

Vulnerability mechanics

Generated on May 9, 2026. Inputs: CWE entries + fix-commit diffs from this CVE's patches. Citations validated against bundle.

References

5

News mentions

0

No linked articles in our index yet.