meta/train_ddp.py

#!/usr/bin/env python3
"""Unified DDP trainer for FastConformer-Hybrid (RNNT+CTC), Phase A or Phase B.
Works single- or multi-GPU. For streaming bases the cache-aware encoder cfg is left
untouched (only data/opt/joint-fuse are set), so streaming behavior is preserved.
Phase A example (8xH100): --lr 1e-3 --warmup 5000 --max-steps 70000 --bd 1600 --devices 8
Phase B example:          --lr 4e-5 --warmup 500  --max-steps 15000 --bd 1600 --devices 8"""
import argparse, lightning.pytorch as pl
from omegaconf import OmegaConf, open_dict
from nemo.collections.asr.models import ASRModel
from nemo.utils.exp_manager import exp_manager
ap=argparse.ArgumentParser()
ap.add_argument("--base", required=True, help="restore_from .nemo (persianized base for Phase A, Phase-A model for Phase B)")
ap.add_argument("--train", required=True); ap.add_argument("--val", required=True)
ap.add_argument("--exp", required=True);   ap.add_argument("--name", required=True)
ap.add_argument("--lr", type=float, default=1e-3)
ap.add_argument("--min-lr", type=float, default=1e-5)
ap.add_argument("--warmup", type=int, default=5000)
ap.add_argument("--max-steps", type=int, default=70000)
ap.add_argument("--bd", type=int, default=1600, help="PER-GPU batch_duration; global ~= bd*devices")
ap.add_argument("--devices", type=int, default=8)
ap.add_argument("--workers", type=int, default=12, help="per-GPU dataloader workers (devices*workers <= cores)")
ap.add_argument("--fuse", type=int, default=16)
ap.add_argument("--maxdur", type=float, default=45.0)
ap.add_argument("--val-interval", type=int, default=2000)
ap.add_argument("--patience", type=int, default=8)
ap.add_argument("--ddp-find-unused", action="store_true", help="set if DDP complains about unused params")
ap.add_argument("--att-context", default=None, help='streaming look-ahead, e.g. "[70,1]" (80ms), "[70,0]" (0ms), or multi "[[70,13],[70,6],[70,1],[70,0]]"')
a=ap.parse_args()
m=ASRModel.restore_from(a.base, map_location="cpu")
cfg=m.cfg
if a.att_context:
    import ast as _ast
    ac=_ast.literal_eval(a.att_context)
    with open_dict(cfg): cfg.encoder.att_context_size=ac
    try:
        single = ac if (ac and isinstance(ac[0],int)) else ac[-1]   # lowest-latency for multi
        m.encoder.set_default_att_context_size(single)
        print(f"[latency] att_context_size set -> {ac}  (default infer ctx {single}; right*80ms = look-ahead)",flush=True)
    except Exception as e:
        print("att_context note:",e)
with open_dict(cfg):
    for ds,mf,sh in [(cfg.train_ds,a.train,True),(cfg.validation_ds,a.val,False)]:
        ds.manifest_filepath=mf; ds.is_tarred=False
        ds.use_lhotse=True; ds.batch_size=None
        ds.batch_duration=a.bd; ds.quadratic_duration=15
        ds.use_bucketing=True; ds.num_buckets=30; ds.bucket_buffer_size=20000; ds.shuffle_buffer_size=10000
        ds.shuffle=sh; ds.num_workers=a.workers; ds.pin_memory=True; ds.pretokenize=False
        ds.max_duration=a.maxdur; ds.min_duration=0.3
        if "tarred_audio_filepaths" in ds: ds.tarred_audio_filepaths=None
        if "augmentor" in ds: ds.augmentor=None
    cfg.train_ds.perturb_speed=True
    cfg.joint.fuse_loss_wer=True; cfg.joint.fused_batch_size=a.fuse
m.setup_training_data(cfg.train_ds); m.setup_validation_data(cfg.validation_ds)
try: m.joint.fuse_loss_wer=True; m.joint.fused_batch_size=a.fuse
except Exception as e: print("joint-fuse note:",e)
m.setup_optimization(OmegaConf.create({"name":"adamw","lr":a.lr,"weight_decay":1e-3,"betas":[0.9,0.98],
    "sched":{"name":"CosineAnnealing","warmup_steps":a.warmup,"min_lr":a.min_lr,"max_steps":a.max_steps}}))
strat = "auto" if a.devices==1 else ("ddp_find_unused_parameters_true" if a.ddp_find_unused else "ddp")
tr=pl.Trainer(logger=False, devices=a.devices, accelerator="gpu", strategy=strat,
    precision="bf16-mixed", max_steps=a.max_steps, accumulate_grad_batches=1,
    val_check_interval=a.val_interval, limit_val_batches=1.0, gradient_clip_val=1.0,
    log_every_n_steps=200, enable_checkpointing=False, enable_progress_bar=True, num_sanity_val_steps=0)
m.set_trainer(tr)
exp_manager(tr, OmegaConf.create({"exp_dir":a.exp,"name":a.name,"create_checkpoint_callback":True,
    "checkpoint_callback_params":{"monitor":"val_wer","mode":"min","save_top_k":3,"always_save_nemo":True},
    "resume_if_exists":True,"resume_ignore_no_checkpoint":True,
    "create_early_stopping_callback":True,"early_stopping_callback_params":{"monitor":"val_wer","mode":"min","patience":a.patience,"min_delta":0.001}}))
print(f"=== DDP fit: devices={a.devices} strategy={strat} lr={a.lr} bd={a.bd} (global~={a.bd*a.devices}s) base={a.base} ===",flush=True)
tr.fit(m); print("=== fit returned ===",flush=True)