# Proton capture of a Triton fused softmax, attributed to Python frames.
#
# Same attach mechanism as triton-vecadd, but with `context: python` so the
# hatchet tree is keyed by Python call stack rather than by launch site --
# the configuration you want once a script has more than one kernel.
#
#   aorta sweep run \
#       --recipe examples/profiling/proton/triton-softmax/recipe.yaml \
#       --output ./profiling_results \
#       -- python examples/profiling/proton/triton-softmax/softmax.py \
#            --rows 4096 --cols 1024 --iters 20
schema_version: 1
mode: probe

ticket: PROFILING-TRITON-SOFTMAX

trials: 1
timeout_per_trial: 1800

mitigation_axis: [none]
diagnostic_axis: [none]

env_passthrough_mode: inherit

# `context: python` is the only difference from triton-vecadd's block. See
# that recipe for why the backend is left on `auto`.
collect:
  proton:
    mode: "cli"
    backend: "auto"
    context: "python"
    data: "tree"