#!/usr/bin/env python3 """Current joint launcher. --print-config is CPU-only and creates no run.""" import json import os from pathlib import Path import sys from psi.hfd_training_launcher import resolve, runtime_environment, to_args def main(): root = Path(__file__).resolve().parents[1] sys.path.insert(0, str(root)) os.chdir(root) # Match the old shell: relative paths are repo-relative. extra = sys.argv[1:] prepare_data = "--prepare-data" in extra config_only = os.environ.get("HFD_CONFIG_ONLY", "0") in {"1", "json"} or "--print-config" in extra extra = [arg for arg in extra if arg not in {"--print-config", "--prepare-data"}] options, mode = resolve(root, os.environ, extra) env = runtime_environment(options, os.environ) args = to_args(options) if config_only: # No dataset scan, network request, monitor, CUDA import or output dir. print(json.dumps(options, ensure_ascii=False, indent=2)) return # The same parser as train.py checks every option before starting workers. import tyro from psi.config.train.posttrain_full_duplex_audio_command_psi0_config import DynamicLaunchConfig parsed = tyro.cli(DynamicLaunchConfig, config=(tyro.conf.ConsolidateSubcommandArgs,), args=args[1:]) if prepare_data: # Exactly the training configuration, no models, CUDA, WandB or new run. os.environ.update(env) os.environ["CUDA_VISIBLE_DEVICES"] = "" from scripts.data.prepare_full_duplex_indexes import warm_indexes warm_indexes(parsed.data) return from psi.hfd_training_preflight import validate_inputs, validate_wandb, validate_output_storage if mode != 'contract': validate_output_storage(options['train.output-dir'],env) validate_inputs(root, options, env, mode) if mode == "contract": print("[ready] Data contracts verified; contract mode does not start training.") return validate_wandb(options, env) import torch count = torch.cuda.device_count() requested = int(env.get("HFD_NPROC_PER_NODE", count)) if not 1 <= requested <= count: raise ValueError(f"HFD_NPROC_PER_NODE={requested}; allocated CUDA device_count={count}") o = options print("[HFD] action input MLP precision=" + ("FP32" if o.get("model.action-input-projection-fp32", False) else "legacy AMP (saved run)"), flush=True) print(f"[HFD] architecture={o.get('model.interaction-architecture', 'legacy')} " "(a1_c_deferred: shared frozen layers 0-27 + private tail8/cross; " "anchored state token prepended to next AV packet)", flush=True) ratios = [o[f"data.{name}-task-session-ratio"] for name in ("interruption", "control", "backchannel")] print(f"[HFD] joint semantics=24 | GPUs={requested} B/GPU={o['train.train-batch-size']} " f"steps={o['train.max-training-steps']} warmup={o['train.warmup-steps']}", flush=True) print(f"[HFD] audio session mix: turn-taking={1-sum(ratios):.1%} interruption={ratios[0]:.1%} " f"stop={ratios[1]:.1%} backchannel={ratios[2]:.1%}; hold=off", flush=True) print(f"[HFD] audio window shares: hesitation={o['data.hesitation-task-window-ratio']:.1%} " f"pause={o['data.pause-task-window-ratio']:.1%}; AV={mode}", flush=True) if o.get('data.third-party-root'): print('[HFD] third-party: append approved split selections once; preserve original pools; ' 'current listen-hold/root40 + shared visual overlay; total ratio reported by dataset.', flush=True) if o.get('data.multimodal-overlay-manifests'): print(f"[HFD] all-task overlay: QA={o.get('data.multimodal-overlay-qa-ratio',.75):.1%}, " f"visual-only={o.get('data.multimodal-overlay-visual-only-ratio',.10):.1%}; " "remaining audio-only; same voice per session, no independent SWA task. " "These are selection probabilities; unsafe insertion falls back to audio-only.",flush=True) print(f"[HFD] save={o['train.checkpointing-steps']} eval={o['train.validation-steps']} " f"eval B={o['train.val-batch-size']}; action/Qwen strides=1/20; RTC=30/20/10", flush=True) print(f"[HFD] timestamp={o['timestamp']} resume={o['train.resume-from-checkpoint']} " f"full_state={o['train.resume-load-full-state']} wandb={o['wandb.resume']}", flush=True) command = [sys.executable, "-m", "torch.distributed.run", "--standalone", "--local_addr=" + env.get("HFD_TORCHRUN_LOCAL_ADDR", "127.0.0.1"), f"--nproc_per_node={requested}", "scripts/train.py", *args] if env.get("HFD_MANAGE_GPU_MONITOR", "0") == "1": command = ["bash", str(root / "scripts/train_with_gpu_monitor.sh"), *command] os.chdir(root) os.execvpe(command[0], command, env) if __name__ == "__main__": try: main() except (ValueError, FileNotFoundError) as exc: raise SystemExit(f"[error] {exc}") from None