Skip to content
  • Nicolas's avatar
    efe6e8bc
    Pin CUDA toolkit to 13.0 for the NVIDIA 580 driver (was 13.3) · efe6e8bc
    Nicolas authored
    
    
    The :gpu image shipped CUDA 13.3, but the target GPU nodes' NVIDIA 580 driver
    supports at most CUDA 13.0: a newer toolkit builds (nvcc needs no GPU) but every
    kernel launch then panics at runtime with "the provided PTX was compiled with an
    unsupported toolchain. (222)". Pin CUDA_VERSION=13-0 (Dockerfile default +
    build-images.sh fallback) and add a note in the README's "Choosing CUDA_VERSION"
    section on the driver-caps-toolkit constraint: CUDA is backward- not
    forward-compatible, so keep CUDA_VERSION <= the driver's max (nvidia-smi shows it);
    NVCC_APPEND_FLAGS=-arch=sm_86 builds SASS ahead-of-time as an escape hatch when a
    newer toolkit must run against an older driver.
    
    Co-Authored-By: default avatarClaude Opus 4.8 (1M context) <noreply@anthropic.com>
    efe6e8bc
    Pin CUDA toolkit to 13.0 for the NVIDIA 580 driver (was 13.3)
    Nicolas authored
    
    
    The :gpu image shipped CUDA 13.3, but the target GPU nodes' NVIDIA 580 driver
    supports at most CUDA 13.0: a newer toolkit builds (nvcc needs no GPU) but every
    kernel launch then panics at runtime with "the provided PTX was compiled with an
    unsupported toolchain. (222)". Pin CUDA_VERSION=13-0 (Dockerfile default +
    build-images.sh fallback) and add a note in the README's "Choosing CUDA_VERSION"
    section on the driver-caps-toolkit constraint: CUDA is backward- not
    forward-compatible, so keep CUDA_VERSION <= the driver's max (nvidia-smi shows it);
    NVCC_APPEND_FLAGS=-arch=sm_86 builds SASS ahead-of-time as an escape hatch when a
    newer toolkit must run against an older driver.
    
    Co-Authored-By: default avatarClaude Opus 4.8 (1M context) <noreply@anthropic.com>
Loading