cuda build_error ai_generated partial

RuntimeError: Triton compilation failed: unsupported instruction 'mma.sync.aligned.m16n8k16.row.col.f16.f16.f16'

ID: cuda/triton-asm-unsupported

Also available as: JSON · Markdown · 中文
80%Fix Rate
86%Confidence
1Evidence
2024-06-12First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
Triton 2.1.0 active — — —
PyTorch 2.2.0 active — — —
NVIDIA T4 (sm_75) active — — —

Root Cause

A Triton kernel uses a PTX instruction (e.g., mma.sync) that is not supported by the target GPU architecture, often due to an older GPU or incorrect compute capability.

generic

中文

Triton内核使用了目标GPU架构不支持的PTX指令（例如mma.sync），通常是由于GPU较旧或计算能力不正确。

Official Documentation

https://triton-lang.org/main/index.html

Workarounds

  1. 80% success Run the kernel on a GPU with compute capability >= 8.0 (Ampere or newer). Alternatively, disable Triton by setting environment variable `TORCHDYNAMO_USE_TRITON=0` to fall back to CUDA kernels. 
    Run the kernel on a GPU with compute capability >= 8.0 (Ampere or newer). Alternatively, disable Triton by setting environment variable `TORCHDYNAMO_USE_TRITON=0` to fall back to CUDA kernels.

中文步骤

  1. Run the kernel on a GPU with compute capability >= 8.0 (Ampere or newer). Alternatively, disable Triton by setting environment variable `TORCHDYNAMO_USE_TRITON=0` to fall back to CUDA kernels.

Dead Ends

Common approaches that don't work:

  1. 70% fail

    The error is hardware-limited, not software.

  2. 90% fail

    Triton uses its own JIT compiler, independent of TensorExpr.