# Smoke test for the llm_determinism workload — single cell, tiny model.
#
# Goal: prove the deterministic-replay harness runs end-to-end and reports
# pass/fail, NOT to stress comm/compute. One cell, one trial, one step.
#
# Single-rank (no torchrun) — export the rendezvous vars first:
#   export RANK=0 WORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29500
#   aorta triage run --recipe recipes/example-llm-determinism-smoke.yaml
#
# Multi-GPU (real comm coverage):
#   torchrun --standalone --nproc_per_node=8 $(which aorta) triage run \
#     --recipe recipes/example-llm-determinism-smoke.yaml
#
# `steps` is a first-class recipe field — putting it in workload_config is rejected.
schema_version: 1

ticket: LLM-DET-SMOKE
workload: llm_determinism

trials: 1
steps: 1

confound:
  threshold: 1.15

# Tiny model so the smoke is fast and fits a single small GPU.
workload_config:
  num_layers: 4
  hidden_size: 1024
  ffn_size: 2816
  num_heads: 8
  seq_len: 256
  batch_size: 1
  dtype: bf16
  seed: 1234
  checksum_mode: per_rank

cells:
  - name: smoke-bf16-4L
    mitigations: [none]
    environment: local
    workload_config:
      capture_dir: ./llm_det_capture/smoke-bf16-4L