Silent kernel death

nlp
1

Environment:
OS : Linux
Python : 3.13.6
PyTorch : 2.6.0+cu124
torchvision : 0.21.0+cu124
transformers : 5.19.0
CUDA : 12.4 (cuDNN 9.1)
GPU : 8× Tesla V100-PCIE-32GB (compute capability 7.0)
What I’m doing:
I load a Qwen3.5-style model (contains GatedDeltaNet linear attention layers) with device_map=“balanced” across 8 V100 GPUs (dtype float16), then run model.generate(…).
The code worked, but printed a warning recommending to install flash-linear-attention. So I installed it via uv pip install flash-linear-attention. During generate(), the whole Jupyter kernel process died with no Python traceback (log: “kernel process died, ExitCode undefined”). At that point I had fla + Triton 3.2.0 installed.

what I did:

  1. I upgraded Triton to 3.3.0 — didn’t help.
  2. I uninstalled flash-linear-attention alone — no change.
  3. I then uninstalled both flash-linear-attention and triton. No more crashes, but now model.generate() is extremely slow: ~8 minutes for a simple generation.

What is the probable cause of the original silent kernel death? Is it Triton 3.3.0 incompatible with torch 2.6.0 (which expects Triton 3.2.0), or an sm70 issue — i.e. do the fla/GatedDeltaNet Triton kernels support V100 (sm70) at all?
Thanks!

2

Jupyter notebooks are unfortunately crashing sometimes without properly report error messages. Try to run the code in a terminal (you can export the notebook as a script) and check if proper errors are displayed there.

3

I run the code .py file but at that moment I got this error

LLVM ERROR: Failed to compute parent layout for slice layout.

4

Thanks you! Indeed I think your initial guess was right:

Triton does not support GPU architectures older than Ampere (sm_80). From the repo:

Supported Hardware:

  • NVIDIA GPUs (Compute Capability 8.0+)