[BUG] CuTe DSL 4.8 dev SIGABRTs in TVM-FFI launch for SM100 ragged SDPA (4.7 passes)

Which component has the problem?

CuTe DSL

Bug Report

Describe the bug

On SM100, the latest CuTe DSL 4.8 development wheel aborts the Python
process while launching a ragged/THD SDPA kernel through TVM-FFI.

The failure reproduces in isolation in about 8 seconds with:

• nvidia-cutlass-dsl==4.8.0a0+20260907123000.a76982f
• NVIDIA/cudnn-frontend at 4af0c9f645cf41a3547c638f9eec0c2861ff7aa3
• apache-tvm-ffi==0.1.13.post3

The process reports:

CUDA Dialect Assertion failed on CUDA error:
cudaErrorInvalidValue - invalid argument

Fatal Python error: Aborted

  File "<string>", line 3 in wrapper
  File ".../cutlass_dsl/tvm_ffi_provider.py", line 1084 in __call__
  File ".../cudnn/sdpa/fwd/api_dsl.py", line 1942 in _execute_thd
  ...
  File ".../test_mhas_v2.py", line 507 in test_sdpa_random_bwd_ragged_L0

This is a native abort rather than a Python exception.

Steps to reproduce

1. Check out cudnn-frontend:

git clone https://github.com/NVIDIA/cudnn-frontend.git
cd cudnn-frontend
git checkout 4af0c9f645cf41a3547c638f9eec0c2861ff7aa3

2. Install the latest 4.8 development wheel and TVM-FFI:

pip install \
  "nvidia-cutlass-dsl[cu13]==4.8.0a0+20260907123000.a76982f" \
  --extra-index-url \
  https://urm.nvidia.com/artifactory/api/pypi/nv-shared-pypi-local/simple/

pip install "apache-tvm-ffi==0.1.13.post3"

3. Build/install cudnn-frontend in the usual way, then run:

CUDNN_FRONTEND_ENABLE_FROST_ENGINES=1 \
pytest -vv -s \
  'test/python/test_mhas_v2.py::test_sdpa_random_bwd_ragged_L0[test370]'

The deterministic test configuration is:

dtype:        fp16
B:            16
H:            5
S_q:          7333
S_kv:         8048
D_qk:         192
D_v:          128
alignment:    bottom-right
right_bound:  0
layout:       ragged
stats layout: head_major

seq_q:
[5252, 4398, 3528, 2933, 0, 1446, 0, 836,
 5526, 1774, 3493, 3848, 6549, 3619, 2920, 5128]

seq_kv:
[6695, 2088, 556, 7866, 5601, 2414, 6438, 5467,
 3232, 381, 6855, 1976, 2164, 6981, 1580, 1277]

Expected behavior

The kernel should launch and produce a result, or at minimum report a
recoverable Python/DSL error. The process must not abort.

The same cudnn-frontend suites complete without this abort with CuTe DSL
4.7.0.

Regression evidence

At identical cudnn-frontend SHAs, the 4.8 lane repeatedly aborts inside the
same call path while the 4.7.0 lane completes:

• FE 32d955d9c1057496a3284dcc9584cd48136589a4
  • 4.8 build 20260905123000.6083892: three SIGABRTs, eventual timeout
  • 4.7.0: completed with no worker abort
• FE 58fc746a1cfdf528d50584412ce4d1966849629f
  • 4.8 build 20260906123000.6083892: three SIGABRTs, eventual timeout
  • 4.7.0: completed with no worker abort

The comparison also crossed two SM100 runners and driver versions: whichever
runner used 4.8 failed, while the 4.7 job completed.

The longer history shows this is a recurring 4.8-only regression:

• Known-good 4.8 builds through 20260823210556.ac70faa
• Identical aborts from 20260825210638.952ea40 through
  20260831123001.b066920
• Temporarily green with 20260901123001.7b963f4 and
  20260902123001.54ab983
• Regressed again with 6083892
• Still present in the latest RC build a76982f

The Python tvm_ffi_provider.py file is identical between the temporarily
good 54ab983 build and the failing 6083892/a76982f builds, so the abort
appears to originate below that Python wrapper or in generated launch code.

Environment

• GPU: B200, compute capability 10.0
• Driver: 595.71.05
• CUDA: 13.2
• Python: 3.12.3
• PyTorch: 2.12.0a0+5aff3928d8.nv26.05
• cuDNN backend: 9.30.0
• cuDNN Frontend: 1.29.0
• Apache TVM-FFI: 0.1.13.post3
• CuTe DSL: 4.8.0a0+20260907123000.a76982f

The nightly CI reproductions also occurred with CUDA 13.4, cuDNN 9.27,
PyTorch 2.14 nightly, and TVM-FFI 0.1.13.post3.

This is separate from the SM80 NVVM compilation regression tracked in #3594.
Both currently block validation of CuTe DSL 4.8 before GA.

Run the pinned pytest case on SM100 with CuTe DSL 4.8 and compare it with 4.7. Trace from test/python/test_mhas_v2.py::test_sdpa_random_bwd_ragged_L0[test370] through cutlass_dsl/tvm_ffi_provider.py:1084 and the generated launch path, using the known-good and failing builds to isolate the regression. Done means the ragged SDPA test launches without a native abort and reports a recoverable error if execution cannot proceed.