# Per-block deterministic-replay recipe for a single repeated transformer block.
#
# Pass workload knobs through `workload_config` at recipe scope (every cell
# inherits) and/or `workload_config` at cell scope (cell wins on key
# collision). The dispatcher forwards the merged dict to
# `LlmDeterminismWorkload(config)` verbatim — same keys the standalone
# launcher in docs/llm-determinism.md uses.
#
# Run on a multi-GPU node (the `aorta` console script is the entry point;
# `python -m aorta.triage.cli` is NOT a runnable module):
#   torchrun --standalone --nproc_per_node=8 $(which aorta) triage run \
#     --recipe recipes/example-llm-determinism.yaml
#
# Each cell writes its own capture_dir, so results don't collide.
#
# Fresh-process isolation keeps TF32/process-global state independent between
# cells. Each trial pays its own FSDP2/RCCL/hipBLAS startup cost, so compare
# checksums rather than startup-dominated wall time.
# `steps` is a first-class recipe field — the loader rejects it inside
# workload_config (the dispatcher would silently clobber it otherwise).
schema_version: 1

ticket: LLM-DET-001
workload: llm_determinism
trial_isolation: process

trials: 1
steps: 1

confound:
  threshold: 1.15

# Recipe-scope defaults — applied to every cell, overridden per-cell below.
workload_config:
  hidden_size: 2048
  ffn_size: 5632
  num_heads: 16
  seq_len: 512
  batch_size: 1
  dtype: bf16
  seed: 1234
  checksum_mode: per_rank

cells:
  # Baseline: 24-layer bf16, default everything else.
  - name: baseline-bf16-24L
    mitigations: [none]
    environment: local
    workload_config:
      num_layers: 24
      capture_dir: ./llm_det_capture/baseline-bf16-24L

  # Half the comm/compute stages — useful for scaling down with rank count.
  - name: bf16-12L
    mitigations: [none]
    environment: local
    workload_config:
      num_layers: 12
      capture_dir: ./llm_det_capture/bf16-12L

  # Same shape as baseline but with tf32_off mitigation. Catches matmul
  # paths whose determinism depends on TF32 reduction ordering.
  - name: tf32_off-bf16-24L
    mitigations: [tf32_off]
    environment: local
    workload_config:
      num_layers: 24
      capture_dir: ./llm_det_capture/tf32_off-bf16-24L

  # MoE variant — top-1 router over 4 GLU experts. Different kernel mix.
  - name: moe4-bf16-12L
    mitigations: [none]
    environment: local
    workload_config:
      num_layers: 12
      num_experts: 4
      capture_dir: ./llm_det_capture/moe4-bf16-12L