# Phase 4: hyperparameter sweep on the 9.4h stage. A100 40GB.
# This is the run the stop/go gate is decided on (see PROJECT_PLAN.md).
# Sweep variants: copy this file, change one thing, change output_dir.
output_dir: runs/lora_10h_r32
seed: 42

data:
  train_manifest: manifests/syspin_10h/train.csv
  val_manifest: manifests/syspin_10h/val.csv
  max_frames_per_batch: 38400   # ~410 s of audio per batch; halve for a 24GB card
  max_samples_per_batch: 48
  num_workers: 8

model:
  vocab_file: checkpoints/IndicF5/checkpoints/vocab.txt
  base_checkpoint: checkpoints/IndicF5/model.safetensors
  arch: {dim: 1024, depth: 22, heads: 16, ff_mult: 2, text_dim: 512, conv_layers: 4}

lora:
  rank: 32
  alpha: 32
  dropout: 0.0
  target_modules:
    - 'transformer\.transformer_blocks\.\d+\.attn\.(to_q|to_k|to_v|to_out\.0)'
    - 'transformer\.transformer_blocks\.\d+\.ff\.ff\.(0\.0|2)'
  # Variant B of the sweep: the text embedding is where character-to-sound mapping lives,
  # so this is the first thing to try if pronunciation barely moves.
  extra_trainable: []
  # extra_trainable: ['transformer\.text_embed\..*']

optim:
  learning_rate: 1.0e-4
  weight_decay: 0.0
  warmup_updates: 800           # ~10% of max_updates
  max_updates: 8000             # ~96 epochs over 9.4h; checkpoints every 1000 so it can be cut short
  grad_accumulation_steps: 1
  max_grad_norm: 1.0
  mixed_precision: bf16
  ema_decay: 0.999

logging:
  log_every: 25
  val_every: 500
  sample_every: 1000
  save_every: 1000
  keep_last: 4

sampling:
  num_per_speaker: 2
  nfe_step: 32
  cfg_strength: 2.0
  sway_sampling_coef: -1.0

# Measured 2026-09-21 on the A100: 0.244 s/update compiled vs 0.43 eager (1.8x). Same validation loss at update 500
# (0.6744 vs 0.6749). Costs ~6 min of one-off compile per run (slow first step + two recompiles), so it pays for runs >~2,000 updates.
perf:
  compile: true