# Smoke-test recipe for the `race` workload (`mode: fsdp`).
#
# `race` is a DISTRIBUTED workload (min_world_size=2), so it MUST be launched
# under torchrun -- a bare `aorta triage run` starts a single process
# (WORLD_SIZE=1) and is correctly refused by launch-mode validation. Use the
# `aorta` console script as torchrun's target (`-m aorta` is NOT a runnable
# module):
#
#   # validate only (no GPUs / no launcher needed):
#   aorta triage run --recipe recipes/example-fsdp-smoke.yaml --dry-run
#
#   # single node, 2 ranks (CX-7 / RCCL smoke; bump to your GPU count):
#   torchrun --standalone --nproc_per_node=2 $(which aorta) triage run -recipe recipes/example-fsdp-smoke.yaml
#
#   # multi-node (e.g. 1 rank/host x 40 hosts -- the AINIC topology):
#   torchrun --nnodes=40 --nproc_per_node=1 --rdzv-backend=c10d \
#     --rdzv-endpoint=$MIRAGE_HEAD_ADDR:$MIRAGE_HEAD_PORT $(which aorta) triage run \
#     --recipe recipes/example-fsdp-smoke.yaml
#
# Results are written by rank 0 only (dispatcher gates on RANK), so multi-rank
# launches don't clobber each other.
#
# Runs the simulated-collective FSDP harness in `aorta.race.modes.fsdp`. This
# is a PUBLIC, non-NDA smoke: it exercises the adapter -> create_reproducer ->
# real RCCL/NCCL collectives path end-to-end on any GPU box. It is NOT expected
# to reproduce the AINIC SDC bug on non-AINIC hardware (e.g. CX-7 / RCCL) -- a
# clean `passed: True` is the expected smoke result; the point is that the
# workload runs.
#
# Iterations are capped small so a smoke finishes in seconds -- the
# ReproducerConfig defaults (verify_iterations=10000, simulate_compute=True)
# would otherwise run for many minutes.
#
# Workload names resolve at cell-EXECUTION time, not load time: this recipe
# validates even before `pip install -e .` registers `race`; an unregistered
# name surfaces as a per-cell error at run time, not a load error.
schema_version: 1

ticket: EXAMPLE-151
workload: race

trials: 2
steps: 100

workload_config:
  mode: fsdp
  warmup_iterations: 2
  verify_iterations: 5
  simulate_compute: false   # skip GEMM work so the smoke is fast
  fsdp_shard_size: 100000

confound:
  threshold: 1.15
  # Left unset so the baseline auto-resolves to `baseline-local` via the
  # "first cell starting with baseline-" rule.

cells:
  - name: baseline-local
    mitigations: [none]
    environment: local

  - name: tf32_off-local
    mitigations: [tf32_off]
    environment: local

  - name: xnack-local
    mitigations: [xnack]
    environment: local