Running With Buck

Running GPU Targets with Buck

This skill covers the mechanics of building and running GPU targets under Buck in fbcode. It is target-agnostic: substitute your own <buck target> and program arguments.

General requirements

1. Run from fbsource/fbcode. cd to fbsource/fbcode before invoking Buck. The @mode/opt flags (and other @mode/... files) only resolve when Buck is run from there.
2. Use @mode/opt. It provides the core GPU build configuration. @mode/opt generally sets up the GPU build, but some tritonbench targets still pass -c fbcode.enable_gpu_sections=true explicitly — add it if a target's GPU sections fail to build.
3. Build against the beta Triton with -m ovr_config//triton:beta. This directory is the beta Triton compiler. Pass this modifier so the target builds and runs against the beta Triton in this tree rather than the default/stable Triton — without it, changes made here are not exercised.
4. Select the GPU architecture with -c fbcode.nvcc_arch=<arch> and, where required, the CUDA version with -m ovr_config//third-party/cuda/constraints:<ver> (see Hardware requirements).
5. Environment variables prefix the buck2 run and are forwarded to the launched process:

   <ENV VARS> buck2 run @mode/opt -m ovr_config//triton:beta -c fbcode.nvcc_arch=<arch> [-m ovr_config//third-party/cuda/constraints:<ver>] \
     <buck target> -- <program args>
   

6. buck2 build vs buck2 run. Use buck2 build <target> to compile only (e.g. to surface a compile failure without executing); use buck2 run <target> -- <args> to build and run. Program arguments go after --.

Hardware requirements

Pick the arch (and CUDA version) for the single GPU you are targeting:

Notes:

• B200 / GB200 require CUDA >= 12.8. Existing tritonbench targets pin 12.8; use the version your build expects.
• GB300 requires arch b300a and CUDA >= 13.0.
• Set the CUDA version explicitly whenever a minimum applies, since the platform default may be older than the arch requires.

Examples

Blackwell GB300 (b300a, CUDA 13.0), from fbsource/fbcode:

buck2 run @mode/opt -m ovr_config//triton:beta \
  -c fbcode.nvcc_arch=b300a \
  -m ovr_config//third-party/cuda/constraints:13.0 \
  <buck target> -- <program args>

Blackwell B200 / GB200 (b200a, CUDA >= 12.8):

buck2 run @mode/opt -m ovr_config//triton:beta \
  -c fbcode.nvcc_arch=b200a \
  -m ovr_config//third-party/cuda/constraints:12.8 \
  <buck target> -- <program args>

Hopper (h100a):

buck2 run @mode/opt -m ovr_config//triton:beta \
  -c fbcode.nvcc_arch=h100a \
  <buck target> -- <program args>

With env vars forwarded (e.g. enabling a feature for the run):

SOME_ENV=1 buck2 run @mode/opt -m ovr_config//triton:beta -c fbcode.nvcc_arch=b300a \
  -m ovr_config//third-party/cuda/constraints:13.0 \
  <buck target> -- <program args>

Compile only (surface a build/compile failure without running):

buck2 build @mode/opt -m ovr_config//triton:beta -c fbcode.nvcc_arch=b300a \
  -m ovr_config//third-party/cuda/constraints:13.0 \
  <buck target>