Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.
update
HenryNdubuakufea8d9588fb8
diff --git a/src/cli.py b/src/cli.pyindex 29aedbe..44c7c31 100644--- a/src/cli.py+++ b/src/cli.py@@ -163,6 +163,16 @@ def main():tp.add_argument("name", type=str)tp.add_argument("--zone", type=str, default=None)+ tp = tpu_sub.add_parser("setup", add_help=False)+ tp.add_argument("name", type=str)+ tp.add_argument("--zone", type=str, default=None)++ tp = tpu_sub.add_parser("train", add_help=False)+ tp.add_argument("name", type=str)+ tp.add_argument("--zone", type=str, default=None)+ tp.add_argument("train_args", nargs="*", default=[],+ help="Extra args passed to needle train")+tp = tpu_sub.add_parser("claude", add_help=False)tp.add_argument("name", type=str)tp.add_argument("--zone", type=str, default=None)@@ -191,6 +201,8 @@ def main():from .tokenize_data import tokenizetokenize(args)elif args.command == "train":+ import jax+ jax.distributed.initialize()from .train import traintrain(args)elif args.command == "run":diff --git a/src/tpu.py b/src/tpu.pyindex e1702f6..52ada98 100644--- a/src/tpu.py+++ b/src/tpu.py@@ -53,6 +53,8 @@ TPU_HELP = """tpu commands:needle tpu create NAME [--type TYPE] [--version VER]needle tpu connect NAME [--zone ZONE]+ needle tpu setup NAME [--zone ZONE] setup all workers (multi-host)+ needle tpu train NAME [--zone ZONE] launch training on all workersneedle tpu claude NAME [--zone ZONE]needle tpu stop NAME [--zone ZONE]needle tpu start NAME [--zone ZONE]@@ -114,6 +116,45 @@ def _update_ssh_config(path, host_name, new_block):f.write(new_block)+def _is_multihost(accel_type):+ """Check if a TPU accelerator type requires multi-host setup.++ v6e-1/4/8 are single-host. v6e-16+ are multi-host.+ """+ m = re.search(r"(\d+)$", accel_type)+ if not m:+ return False+ num_chips = int(m.group(1))+ return num_chips > 8+++def _get_num_workers(name, zone):+ """Get the number of workers for a TPU VM."""+ result = _run(+ ["gcloud", "compute", "tpus", "tpu-vm", "describe", name,+ "--zone", zone, "--project", PROJECT,+ "--format", "json(networkEndpoints)"],+ check=False, capture=True, quiet=True,+ )+ import json+ try:+ data = json.loads(result.stdout)+ return len(data.get("networkEndpoints", []))+ except (json.JSONDecodeError, TypeError):+ return 1+++def _run_all_workers(name, zone, command):+ """Run a command on all workers of a TPU VM."""+ return _run(+ ["gcloud", "compute", "tpus", "tpu-vm", "ssh", name,+ "--zone", zone, "--project", PROJECT,+ "--worker=all",+ "--command", command],+ check=False,+ )++def _check_tpu_health(name, zone):"""SSH into the VM and verify /dev/accel* devices exist."""print(f"[tpu] Verifying TPU devices on '{name}'...")@@ -199,9 +240,15 @@ def tpu_create(args):print(f"[tpu] SUCCESS: created '{args.name}' in {zone}")args.zone = zone_check_tpu_health(args.name, zone)- _setup_git_on_instance(args.name, zone, git_name, git_email)- tpu_claude(args)- tpu_connect(args)+ if _is_multihost(args.accel_type):+ print(f"[tpu] Multi-host TPU detected ({args.accel_type})")+ _setup_git_on_instance(args.name, zone, git_name, git_email)+ tpu_setup(args)+ tpu_claude(args)+ else:+ _setup_git_on_instance(args.name, zone, git_name, git_email)+ tpu_claude(args)+ tpu_connect(args)returnstderr = result.stderr.strip()if not stderr:@@ -372,10 +419,106 @@ def tpu_list(args):print("[tpu] No instances found.")+def tpu_setup(args):+ """Set up all workers of a multi-host TPU: sync code, install deps.++ Creates a lightweight tarball (excludes .venv, .data_cache, __pycache__,+ .git), copies it to all workers via SCP, then runs `source ./setup`+ on each worker to create a fresh venv with jax[tpu].+ """+ zone = args.zone or _detect_zone(args.name)+ num_workers = _get_num_workers(args.name, zone)++ if num_workers <= 1:+ print(f"[tpu] Single-host TPU ({num_workers} worker). Use 'needle tpu connect' instead.")+ return++ print(f"[tpu] Multi-host TPU: {num_workers} workers")++ # Find the needle project root (where this script lives)+ project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))++ # Create lightweight tarball excluding heavy dirs+ import tempfile+ tarball = os.path.join(tempfile.gettempdir(), "needle_sync.tar.gz")+ print(f"[tpu] Creating tarball (excluding .venv, .data_cache, .git)...")+ _run(+ ["tar", "-czf", tarball,+ "--exclude=.venv", "--exclude=.data_cache",+ "--exclude=__pycache__", "--exclude=.git",+ "--exclude=*.pyc", "--exclude=checkpoints",+ "-C", os.path.dirname(project_root),+ os.path.basename(project_root)],+ )+ tar_size = os.path.getsize(tarball)+ print(f"[tpu] Tarball: {tar_size / 1024:.0f} KB")++ # SCP tarball to all workers+ print(f"[tpu] Copying to all {num_workers} workers...")+ result = _run(+ ["gcloud", "compute", "tpus", "tpu-vm", "scp",+ tarball, f"{args.name}:/tmp/needle_sync.tar.gz",+ "--zone", zone, "--project", PROJECT,+ "--worker=all"],+ check=False,+ )+ os.remove(tarball)+ if result.returncode != 0:+ print("[tpu] ERROR: failed to copy tarball to workers.", file=sys.stderr)+ sys.exit(1)++ # Extract and install on all workers+ print(f"[tpu] Extracting and installing on all {num_workers} workers...")+ result = _run_all_workers(args.name, zone,+ "rm -rf ~/needle && tar -xzf /tmp/needle_sync.tar.gz -C ~ "+ "&& rm /tmp/needle_sync.tar.gz "+ "&& cd ~/needle && source ./setup")+ if result.returncode != 0:+ print("[tpu] WARNING: setup failed on some workers.", file=sys.stderr)++ # Set up git config on all workers+ git_name, git_email = _collect_git_config()+ if git_name and git_email:+ _run_all_workers(args.name, zone,+ f'git config --global user.name "{git_name}" && '+ f'git config --global user.email "{git_email}"')++ print(f"[tpu] All {num_workers} workers ready.")+++def tpu_train(args):+ """Launch training on all workers of a multi-host TPU."""+ zone = args.zone or _detect_zone(args.name)+ num_workers = _get_num_workers(args.name, zone)++ # Pass through any extra args after the name+ extra = getattr(args, "train_args", [])+ train_cmd = "cd ~/needle && .venv/bin/needle train"+ if extra:+ train_cmd += " " + " ".join(extra)++ if num_workers <= 1:+ print(f"[tpu] Single-host TPU. Running training directly...")+ _run(+ ["gcloud", "compute", "tpus", "tpu-vm", "ssh", args.name,+ "--zone", zone, "--project", PROJECT,+ "--command", train_cmd],+ )+ return++ print(f"[tpu] Launching training on {num_workers} workers...")+ result = _run_all_workers(args.name, zone, train_cmd)+ if result.returncode != 0:+ print("[tpu] WARNING: training exited with errors on some workers.",+ file=sys.stderr)++def tpu_dispatch(args):actions = {"create": tpu_create,"connect": tpu_connect,+ "setup": tpu_setup,+ "train": tpu_train,"claude": tpu_claude,"stop": tpu_stop,"start": tpu_start,diff --git a/src/train.py b/src/train.pyindex 3279d48..dac6283 100644--- a/src/train.py+++ b/src/train.py@@ -498,10 +498,6 @@ def shard_batch(batch, num_devices):def train(args):- # Multi-host TPU: initialize distributed JAX before anything else.- # On single-host setups this is a no-op (coordinator_address is auto-detected).- jax.distributed.initialize()-num_devices = jax.local_device_count()use_wandb = getattr(args, "wandb", False)