tpuz 0.1.6

Manage GCP TPU VMs from your terminal — create, run, recover, teardown

tpuz

Manage GCP TPU & GPU VMs from your terminal.

Create, train, debug, recover, teardown — one command.

Getting Started · Docs · GPU Guide · Security

---

Why?

Training on GCP TPUs/GPUs means juggling gcloud commands, SSH sessions, preemption recovery, cost tracking, and secrets. tpuz handles all of it:

from tpuz import TPU

tpu = TPU("my-tpu", accelerator="v4-8")
tpu.up()
tpu.setup()
tpu.run("python train.py", secrets=["WANDB_API_KEY"], sync="./src")
tpu.logs()
tpu.cost_summary()  # $4.12 (2.0h × $2.06/hr)
tpu.down()

Or one command:

tpuz train my-tpu "python train.py" -a v4-8 --sync=. --recover --teardown

Install

pip install tpuz

Zero Python dependencies. Requires gcloud CLI (install).

Features

Core

tpu.up()                    # Create TPU VM (idempotent)
tpu.up_queued()             # Queued Resources (reliable spot)
tpu.setup()                 # Install JAX[TPU] + deps
tpu.verify()                # Check JAX on all workers
tpu.run("cmd", sync=".")    # Upload code + launch
tpu.logs()                  # Stream training logs
tpu.wait()                  # Poll for completion
tpu.collect(["model.pkl"])  # Download artifacts
tpu.down()                  # Delete VM

GPU VMs

from tpuz import GCE

vm = GCE.gpu("my-vm", gpu="a100")   # A100 40GB
vm = GCE.gpu("my-vm", gpu="h100x8") # 8x H100
vm = GCE.gpu("my-vm", gpu="t4")     # T4 (cheapest)
vm.up()                              # Same API as TPU

Secrets (Cloud Secret Manager)

from tpuz import SecretManager

sm = SecretManager()
sm.create("WANDB_API_KEY", "your-key")
sm.grant_tpu_access_all()

tpu.run("python train.py", secrets=["WANDB_API_KEY"])
# Secrets never leave GCP — loaded server-side via IAM

Checkpoints (GCS)

from tpuz import GCS

gcs = GCS("gs://my-bucket")
tpu.run_with_resume("python train.py", gcs=gcs, run_name="run-01")
# Auto-detects latest checkpoint → appends --resume-from-step=5000

Preemption Recovery

tpu.watch_notify("python train.py",
    notify_url="https://hooks.slack.com/...",
    max_retries=5)
# Auto: delete → recreate → setup → restart from checkpoint → Slack notify

Debugging

tpu.repl()                             # Interactive Python REPL
tpu.debug("python train.py")           # VS Code debugger attach
tpu.tunnel(6006)                       # TensorBoard
tpu.health_check()                     # Full health dashboard:
#   Process:   running
#   Heartbeat: fresh (12s ago)
#   Disk:      45% (90/200 GB)
#   GPU:       85% utilization
#   Training:  step 1234/5000 | loss 2.31 | 56,000 tok/s
#   ETA:       ~35m

Cost Control

tpu.cost_summary()                     # $4.12 (2.0h × $2.06/hr)
tpu.set_budget(50, notify_url=slack)   # Alert at $40, kill at $50
tpu.schedule("python train.py",
    start_after="22:00", max_cost=10)  # Train overnight, budget $10

Scaling & Failover

tpu.scale("v4-32")                     # Upgrade: v4-8 → v4-32
TPU.create_multi_zone("tpu", "v4-8",
    zones=["us-central2-b", "europe-west4-a"])  # Try each zone

Run-Once (Docker-like)

tpu.run_once("python train.py",
    sync="./src", collect_files=["model.pkl"],
    gcs=gcs, notify_url=slack)
# up → setup → resume → run → wait → collect → notify → down

Profiles & Audit

tpu.save_profile("big-run")            # Save config for reuse
tpu = TPU.from_profile("big-run", "new-tpu")  # Reuse later
tpu.dry_run("python train.py")         # Preview commands without executing

Multi-Host (TPU Pods)

Auto-detected. All SSH commands parallel with per-worker retries:

Accelerator	Chips	Workers	Spot $/hr
v4-8	4	1	$2.06
v4-32	16	4	$8.24
v5litepod-8	8	1	$9.60
v5litepod-64	64	8	$76.80
v6e-8	8	1	$9.60

CLI

tpuz up NAME -a v4-8          tpuz logs NAME
tpuz down NAME                tpuz logs-all NAME
tpuz status NAME              tpuz health NAME
tpuz setup NAME               tpuz tunnel NAME 6006
tpuz verify NAME              tpuz repl NAME
tpuz run NAME "cmd" --sync=.  tpuz debug NAME "cmd"
tpuz wait NAME                tpuz scale NAME v4-32
tpuz kill NAME                tpuz cost NAME
tpuz collect NAME files...    tpuz avail v4-8
tpuz watch NAME "cmd"         tpuz preflight
tpuz train NAME "cmd" -a v4-8 --recover --teardown
tpuz run-once NAME "cmd" --sync=. --collect model.pkl

Documentation

• Getting Started — Zero to training in 9 steps
• Usage Guide — Every feature explained
• GPU VMs — A100/H100/T4 management
• Secrets & Security — Cloud Secret Manager setup
• Best Practices — Production workflows

Pair with kgz

pip install kgz     # Kaggle free GPUs — execute code remotely
pip install tpuz    # GCP TPU/GPU pods — manage VM lifecycle

Claude Code

mkdir -p ~/.claude/skills/tpuz-guide
cp SKILL.md ~/.claude/skills/tpuz-guide/skill.md

Acknowledgments

Cloud TPU resources provided by Google's TPU Research Cloud (TRC) program.

License

MIT