Back to browse
GitHub Repository

Deterministic numerical bisection debugger for Triton FP drift - find where allclose fails without device_print or TRITON_INTERPRET

1 starsPython

Triton-blackhole – debug Triton FP drift without device_print

by nesh23·Jul 23, 2026·1 point·0 comments

AI Analysis

●●●BangerWizardrySolve My Problem

Bisects Triton kernel failures to exact indices without flooding logs with device_print spam.

Strengths
  • Classifies drift types like reduction_order or dtype_cast instead of just reporting mismatch.
  • Works on compiled kernels without needing the slow and incompatible TRITON_INTERPRET mode.
Weaknesses
  • Live kernel execution requires NVIDIA hardware, limiting local testing for Mac users.
  • Niche audience limited to the relatively small subset of developers writing raw Triton code.
Target Audience

ML engineers and CUDA developers writing custom Triton kernels for GPU acceleration

Similar To

Nsight Compute · cuda-gdb · PyTorch Profiler

Similar Projects