Matrix multiply with blis+numpy from conda-forge crashes

[cbu-hw]

In a Python codebase, I encountered mysterious crashes deep within matplotlib when I switch my BLAS to blis, and traced them back to a suspiciously simple numpy.dot multiplication of two identity matrices crashing the interpreter. MWE using Python via miniforge on

• Windows 11 with Powershell (edited - I mistakenly wrote "10" before)
• Python 3.13
• blis 2.0
• numpy 2.5.1

> conda create -n blis-bug python=3.13 "libblas=*=*_blis" numpy
Channels:
 - conda-forge
Platform: win-64
Collecting package metadata (repodata.json): done
Solving environment: done

## Package Plan ##

  environment location: C:\Users\myuser\AppData\Local\miniforge3\envs\blis-bug

  added / updated specs:
    - libblas[build=*_blis]
    - numpy
    - python=3.13


The following NEW packages will be INSTALLED:

  blis               conda-forge/win-64::blis-2.0-pthreads_h02eceb6_4
  bzip2              conda-forge/win-64::bzip2-1.0.8-h0ad9c76_9
  ca-certificates    conda-forge/noarch::ca-certificates-2026.6.17-h4c7d964_0
  libblas            conda-forge/win-64::libblas-3.11.0-8_h281ba03_blis
  libcblas           conda-forge/win-64::libcblas-3.11.0-8_h8e9ec6c_blis
  libexpat           conda-forge/win-64::libexpat-2.8.1-hac47afa_1
  libffi             conda-forge/win-64::libffi-3.5.2-h3d046cb_0
  liblapack          conda-forge/win-64::liblapack-3.11.0-8_h018ca30_netlib
  liblzma            conda-forge/win-64::liblzma-5.8.3-hfd05255_0
  libmpdec           conda-forge/win-64::libmpdec-4.0.0-hfd05255_1
  libsqlite          conda-forge/win-64::libsqlite-3.53.3-hf5d6505_0
  libzlib            conda-forge/win-64::libzlib-1.3.2-hfd05255_2
  numpy              conda-forge/win-64::numpy-2.5.1-py313ha8dc839_0
  openssl            conda-forge/win-64::openssl-3.6.3-hf411b9b_0
  pip                conda-forge/noarch::pip-26.1.2-pyh145f28c_0
  python             conda-forge/win-64::python-3.13.14-h09917c8_100_cp313
  python_abi         conda-forge/noarch::python_abi-3.13-8_cp313
  tk                 conda-forge/win-64::tk-8.6.13-h6ed50ae_3
  tzdata             conda-forge/noarch::tzdata-2025c-hc9c84f9_1
  ucrt               conda-forge/win-64::ucrt-10.0.26100.0-h57928b3_0
  vc                 conda-forge/win-64::vc-14.5-h1b7c187_39
  vc14_runtime       conda-forge/win-64::vc14_runtime-14.51.36231-h1b9f54f_39
  vcomp14            conda-forge/win-64::vcomp14-14.51.36231-h1b9f54f_39


Proceed ([y]/n)?


Downloading and Extracting Packages:

Preparing transaction: done
Verifying transaction: done
Executing transaction: done
#
# To activate this environment, use
#
#     $ conda activate blis-bug
#
# To deactivate an active environment, use
#
#     $ conda deactivate

WARNING conda.conda_pypi.main:notify_externally_managed_future(209):
  This environment has pip installed. A future conda release will
  protect conda environments from accidental 'pip install' usage.
  Try the beta to install PyPI packages natively with conda:
    conda config --set solver rattler
    conda config --append channels conda-pypi
    conda install <package>
  More info: https://docs.conda.io/projects/conda/en/stable/new-features.html
  
> conda activate blis-bug
> python -c "import numpy as np;A = np.array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);B = np.array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);print(np.dot(A, B))"
> $LastExitCode
-1073741819
> python -c "import numpy as np; np._pytesttester._show_numpy_info()"
NumPy version 2.5.1
NumPy CPU features:  X86_V2 X86_V3*

Expected result observed with openblas:

> conda install "libblas=*=*_openblas"
...
    package                    |            build
    ---------------------------|-----------------
    libblas-3.11.0             |8_h0adab6e_openblas          66 KB  conda-forge
    libcblas-3.11.0            |8_h2a8eebe_openblas          67 KB  conda-forge
    liblapack-3.11.0           |8_hd232482_openblas          79 KB  conda-forge
    libopenblas-0.3.33         |pthreads_h877e47f_0         3.9 MB  conda-forge
    ------------------------------------------------------------
                                           Total:         4.1 MB

The following NEW packages will be INSTALLED:

  libopenblas        conda-forge/win-64::libopenblas-0.3.33-pthreads_h877e47f_0

The following packages will be REMOVED:

  blis-2.0-pthreads_h02eceb6_4

The following packages will be REVISED:

  libblas                            3.11.0-8_h281ba03_blis --> 3.11.0-8_h0adab6e_openblas
  libcblas                           3.11.0-8_h8e9ec6c_blis --> 3.11.0-8_h2a8eebe_openblas
  liblapack                        3.11.0-8_h018ca30_netlib --> 3.11.0-8_hd232482_openblas
...
> python -c "import numpy as np;A = np.array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);B = np.array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);print(np.dot(A, B))"
[[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]]

This is on a laptop with an Intel(R) Core(TM) Ultra 7 265U (2.10 GHz). I already did a memcheck and processor check to confirm that the hardware is ok.

Do you need any more information to analyze?
Can anybody reproduce? I feel kinda silly, as I would not expect such a simple repro to have slipped through testing cracks...hence the hardware check 😅

As this touches conda-forge, I hope it's ok to ping @h-vetinari as an FYI, too.

I'm seeing the same crash on Python 3.14 (3.13 was just closest to my original setup).

[devinamatthews]

Unfortunately I can't reproduce on Win11 with Core Ultra 9 285H.

(base) PS C:\Users\46774035> conda create -n blis-bug -c conda-forge python=3.13 "libblas=*=*_blis" numpy
3 channel Terms of Service accepted
Retrieving notices: done
Channels:
 - conda-forge
 - defaults
Platform: win-64
Collecting package metadata (repodata.json): done
Solving environment: done

## Package Plan ##

  environment location: C:\Users\46774035\AppData\Local\miniconda3\envs\blis-bug

  added / updated specs:
    - libblas[build=*_blis]
    - numpy
    - python=3.13


The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    blis-2.0                   |pthreads_h02eceb6_4         5.1 MB  conda-forge
    bzip2-1.0.8                |       h0ad9c76_9          55 KB  conda-forge
    ca-certificates-2026.7.22  |       h4c7d964_0         129 KB  conda-forge
    libblas-3.11.0             |  8_h281ba03_blis          66 KB  conda-forge
    libcblas-3.11.0            |  8_h8e9ec6c_blis          67 KB  conda-forge
    libexpat-2.8.1             |       hac47afa_1          70 KB  conda-forge
    libffi-3.5.2               |       h3d046cb_0          45 KB  conda-forge
    liblapack-3.11.0           |8_h018ca30_netlib         2.0 MB  conda-forge
    liblzma-5.8.3              |       hfd05255_0         104 KB  conda-forge
    libmpdec-4.0.0             |       hfd05255_1          87 KB  conda-forge
    libsqlite-3.53.3           |       hf5d6505_0         1.3 MB  conda-forge
    libzlib-1.3.2              |       hfd05255_2          57 KB  conda-forge
    numpy-2.5.1                |  py313ha8dc839_0         7.0 MB  conda-forge
    openssl-3.6.3              |       hf411b9b_0         9.0 MB  conda-forge
    pip-26.1.2                 |     pyh145f28c_0         1.1 MB  conda-forge
    python-3.13.14             |h09917c8_100_cp313        16.0 MB  conda-forge
    python_abi-3.13            |          8_cp313           7 KB  conda-forge
    tk-8.6.13                  |       h967ab96_3         3.6 MB  conda-forge
    tzdata-2026c               |       h151e31d_0         116 KB  conda-forge
    ucrt-10.0.26100.0          |       h57928b3_0         678 KB  conda-forge
    vc-14.5                    |      h1b7c187_39          20 KB  conda-forge
    vc14_runtime-14.51.36231   |      h1b9f54f_39         720 KB  conda-forge
    vcomp14-14.51.36231        |      h1b9f54f_39         118 KB  conda-forge
    ------------------------------------------------------------
                                           Total:        47.5 MB

The following NEW packages will be INSTALLED:

  blis               conda-forge/win-64::blis-2.0-pthreads_h02eceb6_4
  bzip2              conda-forge/win-64::bzip2-1.0.8-h0ad9c76_9
  ca-certificates    conda-forge/noarch::ca-certificates-2026.7.22-h4c7d964_0
  libblas            conda-forge/win-64::libblas-3.11.0-8_h281ba03_blis
  libcblas           conda-forge/win-64::libcblas-3.11.0-8_h8e9ec6c_blis
  libexpat           conda-forge/win-64::libexpat-2.8.1-hac47afa_1
  libffi             conda-forge/win-64::libffi-3.5.2-h3d046cb_0
  liblapack          conda-forge/win-64::liblapack-3.11.0-8_h018ca30_netlib
  liblzma            conda-forge/win-64::liblzma-5.8.3-hfd05255_0
  libmpdec           conda-forge/win-64::libmpdec-4.0.0-hfd05255_1
  libsqlite          conda-forge/win-64::libsqlite-3.53.3-hf5d6505_0
  libzlib            conda-forge/win-64::libzlib-1.3.2-hfd05255_2
  numpy              conda-forge/win-64::numpy-2.5.1-py313ha8dc839_0
  openssl            conda-forge/win-64::openssl-3.6.3-hf411b9b_0
  pip                conda-forge/noarch::pip-26.1.2-pyh145f28c_0
  python             conda-forge/win-64::python-3.13.14-h09917c8_100_cp313
  python_abi         conda-forge/noarch::python_abi-3.13-8_cp313
  tk                 conda-forge/win-64::tk-8.6.13-h967ab96_3
  tzdata             conda-forge/noarch::tzdata-2026c-h151e31d_0
  ucrt               conda-forge/win-64::ucrt-10.0.26100.0-h57928b3_0
  vc                 conda-forge/win-64::vc-14.5-h1b7c187_39
  vc14_runtime       conda-forge/win-64::vc14_runtime-14.51.36231-h1b9f54f_39
  vcomp14            conda-forge/win-64::vcomp14-14.51.36231-h1b9f54f_39


Proceed ([y]/n)?


Downloading and Extracting Packages:

Preparing transaction: done
Verifying transaction: done
Executing transaction: done
#
# To activate this environment, use
#
#     $ conda activate blis-bug
#
# To deactivate an active environment, use
#
#     $ conda deactivate


Channel "defaults" has the following notices:
  [info] -- Tue Jun  9 00:00:00 2026
  PyTorch 2.12 with CUDA support is now available to install with your current channel (Anaconda Main). Learn more: https://anacondas

WARNING conda.conda_pypi.main:notify_externally_managed_future(156):
  Did you know? You can install many PyPI packages with conda
  using the conda-pypi beta. Get started:
    https://docs.conda.io/projects/conda/en/stable/new-features.html

(base) PS C:\Users\46774035> conda activate blis-bug
(blis-bug) PS C:\Users\46774035> python -c "import numpy as np;A = np.array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);B = np.array([[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]);print(np.dot(A, B))"
[[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]]
(blis-bug) PS C:\Users\46774035> $LastExitCode
0

[cbu-hw]

Thank you for trying to repro! I did some more digging. I tried on other machines, and could not reproduce there, either.
The closest I could get to w.r.t. CPU is a Intel(R) Core(TM) Ultra 7 265HX (2.60 GHz) (Intel64 Family 6 Model 198 Stepping 2, GenuineIntel), which is on a more powerful version of my laptop (265HX instead of 265U). This also does not show the problem.

Interestingly, I found out that forcing single-threading makes the issue disappear!

$env:BLIS_NUM_THREADS=1;python -c "import numpy as np; A=np.eye(3); print(np.dot(A,A))"
[[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]]​‌

works​‌! Any thread count >1 consistently crashes.

Going to dgemm directly crashes with BLIS_NUM_THREADS>1

$env:BLIS_NUM_THREADS=2;python -c "import scipy.linalg.blas as blas;import numpy as np;A = np.eye(3);print(blas.dgemm(alpha=1.0, a=A, b=A))"

Matrix size does not seem to play a role:

$env:BLIS_NUM_THREADS=1;python -c "import numpy as np; A=np.random.rand(500,500); print((A@A)[0,0])"
123.17867102205722

works

$env:BLIS_NUM_THREADS=2;python -c "import numpy as np; A=np.random.rand(500,500); print((A@A)[0,0])"

crashes.

Could it be that the problem is in BLIS multithreading/runtime management rather than the DGEMM kernel itself?

This is a CPU with a maybe unexpected core setup (2 performance, 8 efficient, 2 low-power efficient cores)?

Get-CimInstance Win32_Processor | Format-List *


Availability                            : 3
CpuStatus                               : 1
CurrentVoltage                          : 12
DeviceID                                : CPU0
ErrorCleared                            :
ErrorDescription                        :
LastErrorCode                           :
LoadPercentage                          : 39
Status                                  : OK
StatusInfo                              : 3
AddressWidth                            : 64
DataWidth                               : 64
ExtClock                                : 100
L2CacheSize                             : 10240
L2CacheSpeed                            :
MaxClockSpeed                           : 2100
PowerManagementSupported                : False
ProcessorType                           : 3
Revision                                :
SocketDesignation                       : U3E1
Version                                 :
VoltageCaps                             :
Caption                                 : Intel64 Family 6 Model 181 Stepping 0
Description                             : Intel64 Family 6 Model 181 Stepping 0
InstallDate                             :
Name                                    : Intel(R) Core(TM) Ultra 7 265U
ConfigManagerErrorCode                  :
ConfigManagerUserConfig                 :
CreationClassName                       : Win32_Processor
PNPDeviceID                             :
PowerManagementCapabilities             :
SystemCreationClassName                 : Win32_ComputerSystem
SystemName                              : XXXXXXX
CurrentClockSpeed                       : 2100
Family                                  : 774
OtherFamilyDescription                  :
Role                                    : CPU
Stepping                                :
UniqueId                                :
UpgradeMethod                           : 83
Architecture                            : 9
AssetTag                                :
Characteristics                         : 252
L3CacheSize                             : 12288
L3CacheSpeed                            : 0
Level                                   : 6
Manufacturer                            : GenuineIntel
NumberOfCores                           : 12
NumberOfEnabledCore                     : 12
NumberOfLogicalProcessors               : 14
PartNumber                              :
ProcessorId                             : BFEBFBFF000B0650
SecondLevelAddressTranslationExtensions : False
SerialNumber                            :
ThreadCount                             : 14
VirtualizationFirmwareEnabled           : False
VMMonitorModeExtensions                 : False
PSComputerName                          :
CimClass                                : root/cimv2:Win32_Processor
CimInstanceProperties                   : {Caption, Description, InstallDate, Name...}
CimSystemProperties                     : Microsoft.Management.Infrastructure.CimSystemProperties

What can I do to help analysis?