[BUG] Unexpected SystemExit: nvidia_cutlass_dsl/python_packages/cutlass/base_dsl/dsl.py(911): diagnostic

Which component has the problem?

CuTe DSL

Bug Report

Describe the bug

Executing NVBench Python example script using CuteDSL runs into "error: argument -diagnostic: invalid choice: '0' (choose from all, fail, success, info, suggestion)"

$ python cute_dsl_sgemm.py -d 0 -a "R=64" -a "N=256" --stopping-criterion entropy
# Devices

## [0] `NVIDIA RTX A6000`
* SM Version: 860 (PTX Version: 750)
* Number of SMs: 84
* SM Default Clock Rate: 1800 MHz
* Global Memory: 43478 MiB Free / 48532 MiB Total
* Global Memory Bus Peak: 768 GB/sec (384-bit DDR @8001MHz)
* Max Shared Memory: 100 KiB/SM, 48 KiB/Block
* L2 Cache Size: 6144 KiB
* Maximum Active Blocks: 16/SM
* Maximum Active Threads: 1536/SM, 1024/Block
* Available Registers: 65536/SM, 65536/Block
* ECC Enabled: No

# Log

# Benchmark Results

## cutlass_gemm

### [0] NVIDIA RTX A6000

No data -- check log.
Segmentation fault

Steps/Code to reproduce bug

On a machine with CTK 13 installed, and GPU available:

mkdir -p /tmp/reproducer
pushd /tmp/reproducer
git clone https://github.com/oleksandr-pavlyk/nvbench -b update-python-examples
cd nvbench/python

conda create -n cute-dsl-bug --yes python=3.12 pip 
conda activate cute-dsl-bug

# install dependencies to build NVBench python package
pip install cython scikit-build-core setuptools-scm typing-extensions cmake cuda-bindings

export CUDACXX=/usr/local/cuda/bin/nvcc
python -m pip install --no-build-isolation --no-deps --verbose .

cd examples

# install example dependencies
pip install cuda-core cupy nvidia-cutlass-dsl[cu13] 

# run the example that reproduces the behavior
python cute_dsl_sgemm.py -d 0 -a R=64 -a N=1024

I am aware of a work-around, which is to modify $CONDA_PREFIX/lib/python3.12/site-packages/nvidia_cutlass_dsl/python_packages/cutlass/base_dsl/dsl.py using the following diff:

$ diff $CONDA_PREFIX/lib/python3.12/site-packages/nvidia_cutlass_dsl/python_packages/cutlass/base_dsl/dsl_orig.py $CONDA_PREFIX/lib/python3.12/site-packages/nvidia_cutlass_dsl/python_packages/cutlass/base_dsl/dsl.py
911c911
<         args, _ = parser.parse_known_args()
---
>         # args, _ = parser.parse_known_args()
920c920
<         if args.diagnostic is None:
---
>         if True or args.diagnostic is None:

Expected behavior

The example should run without errors, like after applying the work-around

(report-bug) opavlyk@ee09c48-lcedt:/tmp/opavlyk/nvbench/python/examples$ python cute_dsl_sgemm.py -d 0 -a R=64 -a N=1024 --profile
# Devices

## [0] `NVIDIA RTX A6000`
* SM Version: 860 (PTX Version: 750)
* Number of SMs: 84
* SM Default Clock Rate: 1800 MHz
* Global Memory: 43627 MiB Free / 48532 MiB Total
* Global Memory Bus Peak: 768 GB/sec (384-bit DDR @8001MHz)
* Max Shared Memory: 100 KiB/SM, 48 KiB/Block
* L2 Cache Size: 6144 KiB
* Maximum Active Blocks: 16/SM
* Maximum Active Threads: 1536/SM, 1024/Block
* Available Registers: 65536/SM, 65536/Block
* ECC Enabled: No

# Log

# Benchmark Results

## cutlass_gemm

### [0] NVIDIA RTX A6000

| R  |  N   | Samples | CPU Time  | Noise | GPU Time  | Noise |
|----|------|---------|-----------|-------|-----------|-------|
| 64 | 1024 |      1x | 89.342 ms |  inf% | 89.331 ms |  inf% |

Environment details (please complete the following information):

Bare metal, Ubuntu 24.04, CTK 13.1,

Wed Apr  1 09:15:36 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 590.44.01              Driver Version: 590.44.01      CUDA Version: 13.1     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA RTX A6000               On  |   00000000:01:00.0 Off |                  Off |
| 39%   63C    P2             87W /  300W |    4712MiB /  49140MiB |      6%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
|   1  NVIDIA RTX A400                On  |   00000000:C1:00.0  On |                  N/A |
| 30%   42C    P8            N/A  /   50W |     131MiB /   4094MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

Additional context

In the same environment:

cd ../../../
git clone https://github.com/NVIDIA/cutlass
cd cutlass/examples/python/CuTeDSL/ampere

# install missing dependency
pip install torch[cu13]

# run the original example (runs fine)
python sgemm.py

Start with nvidia_cutlass_dsl/python_packages/cutlass/base_dsl/dsl.py around the diagnostic function and compare its argument parsing with nvbench/python/examples/cute_dsl_sgemm.py. Reproduce using the provided CTK 13 environment and command, then verify that the example completes with benchmark data instead of SystemExit and a segmentation fault.