podstack 1.3.21

Official Python SDK for Podstack GPU Notebook Platform

Podstack Python SDK

Official Python SDK for the Podstack GPU Platform. Run ML workloads on remote GPUs with simple decorators, track experiments, and manage models.

Installation

pip install podstack

With optional dependencies:

pip install podstack[torch]        # PyTorch support
pip install podstack[huggingface]  # HuggingFace Transformers
pip install podstack[all]          # All ML frameworks

Quick Start

import podstack

# Initialize the SDK
podstack.init(
    api_key="your-api-key",
    project_id="your-project-id"
)

# Run a function on a remote GPU with a single decorator
@podstack.gpu(type="L40S", fraction=100)
def train():
    import torch
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    return {"status": "done"}

result = train()  # Executes on remote GPU!

Decorators & Annotations

Podstack provides decorators that turn any Python function into a remote GPU workload with built-in experiment tracking.

@podstack.gpu - Remote GPU Execution

import podstack

# Basic GPU execution
@podstack.gpu(type="L40S")
def train_model():
    import torch
    model = torch.nn.Linear(768, 10).cuda()
    return {"params": sum(p.numel() for p in model.parameters())}

result = train_model()

# Specify GPU type, count, and fraction
@podstack.gpu(type="A100-80G", count=2, fraction=100)
def train_large_model():
    import torch
    print(f"GPUs available: {torch.cuda.device_count()}")

# Install pip packages on the fly
@podstack.gpu(type="L40S", pip=["transformers", "datasets", "accelerate"])
def finetune_llm():
    from transformers import AutoModelForCausalLM, AutoTokenizer
    model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
    ...

# Use uv for faster package installation
@podstack.gpu(type="L40S", uv=["torch", "transformers"])
def fast_setup():
    ...

# Install from requirements.txt
@podstack.gpu(type="L40S", requirements="requirements.txt", use_uv=True)
def train_with_deps():
    ...

# Use conda packages
@podstack.gpu(type="L40S", conda="cudatoolkit=11.8")
def train_with_conda():
    ...

# Use a pre-built environment
@podstack.gpu(type="L40S", env="nlp")
def nlp_task():
    ...

# Set execution timeout (default: 3600s)
@podstack.gpu(type="L40S", timeout=7200)
def long_training():
    ...

# Disable remote execution (run locally for debugging)
@podstack.gpu(type="L40S", remote=False)
def debug_locally():
    print("This runs on your local machine")

# Use as a context manager
with podstack.gpu(type="A100-80G", count=2) as cfg:
    print(f"GPU config set: {cfg.type}")

Available GPU types: T4, L4, A10, L40S, A100-40G, A100-80G, H100

Available environments: ml, nlp, cv, audio, tabular, rl, scientific

@podstack.experiment - Experiment Tracking

import podstack

# As a decorator
@podstack.experiment(name="transformer-experiments")
def run_experiment():
    ...

# As a context manager
with podstack.experiment(name="transformer-experiments") as exp:
    print(f"Experiment ID: {exp.id}")

@podstack.run - Run Tracking

Automatically tracks execution time and GPU configuration.

import podstack

# As a decorator
@podstack.experiment(name="my-experiment")
@podstack.run(name="training-v1", track_gpu=True)
def train():
    podstack.registry.log_params({"lr": 0.001, "batch_size": 32})
    for epoch in range(10):
        loss = 1.0 / (epoch + 1)
        podstack.registry.log_metrics({"loss": loss}, step=epoch)

# As a context manager
with podstack.run(name="training-v1") as run:
    podstack.registry.log_params({"lr": 0.001})
    podstack.registry.log_metrics({"loss": 0.5}, step=1)
    print(f"Run ID: {run.id}")

# With tags
@podstack.run(name="ablation-study", tags={"variant": "no-dropout"})
def ablation():
    ...

@podstack.model - Model Registration

import podstack

# Register model after function completes
@podstack.experiment(name="my-experiment")
@podstack.run(name="training-v1")
@podstack.model.register(name="my-classifier")
def train_and_save():
    import torch
    model = torch.nn.Linear(768, 10)
    torch.save(model.state_dict(), "model.pt")
    podstack.registry.log_artifact("model.pt", "model")

# Promote model to production after validation
@podstack.model.promote(name="my-classifier", version=1, stage="production")
def validate_and_promote():
    # Run validation checks
    accuracy = 0.95
    assert accuracy > 0.90, "Model doesn't meet threshold"

Combining Decorators

Stack decorators for a complete ML workflow:

import podstack

podstack.init(api_key="your-api-key", project_id="your-project-id")

@podstack.gpu(type="L40S", pip=["transformers", "datasets"])
@podstack.experiment(name="sentiment-analysis")
@podstack.run(name="bert-finetune-v1", track_gpu=True)
@podstack.model.register(name="sentiment-bert")
def full_pipeline():
    from transformers import AutoModelForSequenceClassification, Trainer

    model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")

    # Log hyperparameters
    podstack.registry.log_params({
        "model": "bert-base-uncased",
        "learning_rate": 2e-5,
        "epochs": 3
    })

    # Train...
    podstack.registry.log_metrics({"accuracy": 0.92, "f1": 0.89})

    return {"accuracy": 0.92}

result = full_pipeline()  # Runs on remote L40S GPU with full tracking

Registry - Experiment Tracking & Model Management

Initialize

from podstack import registry

registry.init(
    api_key="your-api-key",
    project_id="your-project-id"
)

Track Experiments and Runs

from podstack import registry

# Set experiment
registry.set_experiment("my-experiment")

# Start a tracked run
with registry.start_run(name="training-v1") as run:
    # Log hyperparameters
    registry.log_params({
        "learning_rate": 0.001,
        "batch_size": 32,
        "epochs": 10,
        "optimizer": "adam"
    })

    # Log metrics at each step
    for epoch in range(10):
        loss = train_epoch()
        accuracy = evaluate()
        registry.log_metrics({"loss": loss, "accuracy": accuracy}, step=epoch)

    # Set tags
    registry.set_tag("framework", "pytorch")

    # Upload artifacts to cloud artifact store
    registry.log_artifact("model.pt")
    registry.log_artifact("training_curves.png", artifact_path="plots/curves.png")

    # Log dataset provenance (first-class resource, deduped by content hash)
    registry.log_dataset("imdb-reviews", path="data/imdb.csv", context="training")

    # Or pass a DataFrame — schema and row/feature counts are auto-computed
    import pandas as pd
    df = pd.read_csv("data/imdb.csv")
    registry.log_dataset("imdb-reviews", df=df, context="training")

Log and Load Models

from podstack import registry

# Serialize and upload the model to the artifact store (auto-detects framework)
registry.log_model(model, artifact_path="model", framework="pytorch")

# Register in model registry
registry.register_model(
    name="my-classifier",
    run_id=run.id,
    description="BERT sentiment classifier"
)

# Promote to production
registry.set_model_stage("my-classifier", version=1, stage="production")

# Set aliases
registry.set_model_alias("my-classifier", alias="champion", version=1)

# Load model from any machine — files are downloaded automatically if missing locally
model = registry.load_model("my-classifier", stage="production")

Compare Runs

from podstack import registry

# Compare multiple runs
comparison = registry.compare_runs(
    run_ids=["run-id-1", "run-id-2", "run-id-3"],
    metric_keys=["loss", "accuracy"]
)

# Get metric history for a run
history = registry.get_metric_history("run-id-1", "loss")
for point in history:
    print(f"Step {point.step}: {point.value}")

# Search runs
runs = registry.search_runs(
    experiment_id="exp-id",
    status="completed",
    max_results=50
)

Dataset Tracking & Lineage

Podstack tracks datasets as first-class resources, linking them to runs and model versions so you can always answer "what data was this model trained on?"

The lineage chain is:

Dataset(s) ──[logged to]──▶ Run ──[run_id]──▶ ModelVersion

log_dataset() — log a dataset to the active run

dataset = registry.log_dataset(
    name="imdb-reviews",          # required — human-readable name
    path="data/imdb.csv",         # local path or URI (s3://, gcs://, https://)
    context="training",           # "training" | "validation" | "test" (default: "training")
)

The dataset is stored as a project-level resource and linked to the current run. Subsequent calls with the same file produce the same dataset record — no duplicates.

Auto-enrichment from a local file:

# SHA-256 digest is computed automatically for files ≤ 500 MB.
# This enables deduplication across runs — if two runs use the exact
# same file, they share one Dataset record in the registry.
dataset = registry.log_dataset("imdb-reviews", path="data/imdb.csv")
print(dataset.digest)  # "a3f2c1..." — hex SHA-256

Auto-enrichment from a pandas DataFrame:

import pandas as pd

df = pd.read_csv("data/imdb.csv")

dataset = registry.log_dataset(
    name="imdb-reviews",
    df=df,
    context="training",
)
# schema and profile are computed automatically:
print(dataset.schema)   # {"text": "object", "label": "int64"}
print(dataset.profile)  # {"num_rows": 50000, "num_features": 2}

Pass both path and df to get digest dedup and schema inference:

dataset = registry.log_dataset("imdb-reviews", path="data/imdb.csv", df=df)

All parameters:

Parameter	Type	Default	Description
name	str	required	Human-readable dataset name
path	str	None	Local file path or URI (s3://, gcs://, https://)
df	DataFrame	None	pandas DataFrame — schema and profile auto-computed
context	str	"training"	Role of the dataset: "training", "validation", or "test"
digest	str	None	SHA-256 hex digest. Computed from path if not provided
source_type	str	"local"	Storage backend: "local", "s3", "gcs", "url"
tags	dict	None	Arbitrary string key-value tags

Returns: Dataset object with fields:

Field	Type	Description
id	str	UUID of the dataset record
name	str	Dataset name
digest	str	SHA-256 hex digest (empty if not computed)
source_type	str	Storage backend
source	str	File path or URI
schema	dict	Column → dtype mapping
profile	dict	num_rows, num_features, and any other stats
tags	dict	Tags dict
created_at	str	ISO 8601 timestamp

Via the Run object (equivalent to calling registry.log_dataset()):

with registry.start_run("training-v1") as run:
    dataset = run.log_dataset("imdb-reviews", df=df, context="training")

Multiple datasets per run

Log validation and test sets alongside the training set:

with registry.start_run("bert-finetune") as run:
    run.log_dataset("imdb-train", df=train_df, context="training")
    run.log_dataset("imdb-val",   df=val_df,   context="validation")
    run.log_dataset("imdb-test",  df=test_df,  context="test")

get_run_datasets() — retrieve datasets logged to a run

Returns every Dataset object linked to a run, in the order they were logged.

datasets = registry.get_run_datasets(run_id)

Parameters:

Parameter	Type	Description
run_id	str	ID of the run to query

Returns: list[Dataset] — same object as returned by log_dataset().

Fields on each Dataset:

Field	Type	Description
id	str	UUID of the dataset record
name	str	Human-readable name
digest	str	SHA-256 hex digest (empty if not computed at log time)
source_type	str	"local", "s3", "gcs", or "url"
source	str	File path or URI that was passed to log_dataset()
schema	dict	Column → dtype mapping (e.g. {"text": "object", "label": "int64"})
profile	dict	Stats dict, always contains num_rows and num_features when a DataFrame was passed
tags	dict	Key-value tags
created_at	str	ISO 8601 timestamp

Examples:

from podstack import registry

registry.init(api_key="...", project_id="...")

datasets = registry.get_run_datasets("3a9f12c4-...")

# Inspect each dataset
for ds in datasets:
    print(ds.name)
    print(f"  source : {ds.source}")
    print(f"  digest : {ds.digest[:16]}…")
    print(f"  rows   : {ds.profile.get('num_rows', 'unknown')}")
    print(f"  schema : {ds.schema}")

Checking datasets on a run you have in hand:

with registry.start_run("training-v1") as run:
    run.log_dataset("train", df=train_df, context="training")
    run.log_dataset("val",   df=val_df,   context="validation")

# After the run completes, retrieve everything that was logged
datasets = registry.get_run_datasets(run.id)
assert len(datasets) == 2

Verifying deduplication — the same physical file logged across two runs returns the same dataset ID:

ds1 = registry.get_run_datasets(run_a.id)[0]
ds2 = registry.get_run_datasets(run_b.id)[0]

# Same file → same digest → same Dataset record
assert ds1.id == ds2.id
assert ds1.digest == ds2.digest

get_model_lineage() — trace a model back to its training data

Returns the full provenance chain for every version of a registered model: which datasets each version was trained on, via which run.

lineage = registry.get_model_lineage(model_id)

Parameters:

Parameter	Type	Description
model_id	str	ID of the registered model

Returns: dict with the following structure:

{
  "model_id": str,
  "versions": [
    {
      "version":  int,        # version number (1, 2, 3 …)
      "stage":    str,        # "development" | "staging" | "production" | "archived"
      "run_id":   str,        # ID of the linked training run (empty if none)
      "run_name": str,        # display name of the run
      "datasets": [Dataset]   # list of Dataset dicts logged to that run
    },
    …
  ]
}

Each datasets entry has the same fields as a Dataset object (id, name, digest, source_type, source, schema, profile, tags, created_at).

Examples:

Basic iteration:

from podstack import registry

registry.init(api_key="...", project_id="...")

model   = registry.get_model("sentiment-bert")
lineage = registry.get_model_lineage(model.id)

for version in lineage["versions"]:
    print(f"v{version['version']} · {version['stage']}")
    print(f"  Run: {version['run_name']} ({version['run_id'][:8]}…)")
    for ds in version["datasets"]:
        rows = ds["profile"].get("num_rows", "?")
        print(f"  └─ {ds['name']}  {rows} rows  sha256:{ds['digest'][:12]}…")

Example output:

v3 · production
  Run: bert-finetune-v3 (3a9f12c4…)
  └─ imdb-train  40000 rows  sha256:a3f2c1d8e9b0…
  └─ imdb-val     5000 rows  sha256:7e4b2f1a0c3d…
v2 · staging
  Run: bert-finetune-v2 (8b2e77d1…)
  └─ imdb-train  40000 rows  sha256:a3f2c1d8e9b0…
v1 · archived
  Run: bert-finetune-v1 (f1c3a0e2…)
  └─ imdb-train  40000 rows  sha256:a3f2c1d8e9b0…

Finding every unique dataset ever used to train any version of a model:

lineage  = registry.get_model_lineage(model.id)
seen     = {}
for version in lineage["versions"]:
    for ds in version["datasets"]:
        seen[ds["id"]] = ds  # dedup by ID

unique_datasets = list(seen.values())
print(f"{len(unique_datasets)} unique dataset(s) across all versions")

Checking whether the production version was trained on an approved dataset:

APPROVED_DIGEST = "a3f2c1d8e9b0..."

lineage = registry.get_model_lineage(model.id)
prod = next(v for v in lineage["versions"] if v["stage"] == "production")

approved = any(ds["digest"] == APPROVED_DIGEST for ds in prod["datasets"])
print("Production model trained on approved data:", approved)

End-to-end example

import pandas as pd
from podstack import registry

registry.init(api_key="...", project_id="...")
registry.set_experiment("sentiment-analysis")

# Load data
train_df = pd.read_csv("data/train.csv")
val_df   = pd.read_csv("data/val.csv")

with registry.start_run("bert-finetune-v3") as run:
    # Log datasets — digest is auto-computed, schema inferred
    run.log_dataset("imdb-train", path="data/train.csv", df=train_df, context="training")
    run.log_dataset("imdb-val",   path="data/val.csv",   df=val_df,   context="validation")

    # Train
    run.log_params({"lr": 2e-5, "epochs": 3})
    run.log_metrics({"accuracy": 0.93, "f1": 0.92})

# Register and promote the model
registry.register_model("sentiment-bert", run_id=run.id)
registry.set_model_stage("sentiment-bert", version=3, stage="production")

# Later — answer "what data trained v3?"
model = registry.get_model("sentiment-bert")
lineage = registry.get_model_lineage(model.id)

Artifact Storage

Podstack stores every artifact you log — model files, plots, CSV exports, anything — in the project's cloud artifact store. Artifacts are keyed by run ID, so the same file can be retrieved from any machine, by any project member, at any time.

log_artifact() — upload a file for the active run

# Upload a single file (uses the filename as the artifact path)
registry.log_artifact("model.pt")

# Upload with an explicit path inside the artifact store
registry.log_artifact("training_curves.png", artifact_path="plots/curves.png")
registry.log_artifact("feature_importance.csv", artifact_path="analysis/features.csv")

Parameters:

Parameter	Type	Default	Description
local_path	str	required	Path to the local file to upload
artifact_path	str	filename	Relative path inside the artifact store. Defaults to os.path.basename(local_path)

If the artifact store is temporarily unreachable, the SDK saves the file to a local fallback cache (~/.podstack/artifacts/<run_id>/) so your run is never interrupted.

Via the Run object — equivalent to calling registry.log_artifact():

with registry.start_run("training-v1") as run:
    run.log_artifact("confusion_matrix.png", artifact_path="plots/confusion_matrix.png")
    run.log_artifact("model.pkl")

list_artifacts() — list all artifacts for a run

artifacts = registry.list_artifacts(run_id)
for a in artifacts:
    print(f"{a['path']:40s}  {a['size'] / 1e6:.1f} MB  {a['last_modified']}")

Parameters:

Parameter	Type	Description
run_id	str	ID of the run to query

Returns: list[dict] — one entry per artifact:

Key	Type	Description
path	str	Relative artifact path (e.g. "plots/curves.png")
size	int	File size in bytes
etag	str	Content hash for integrity verification
last_modified	str	ISO 8601 upload timestamp

download_artifact() — retrieve an artifact

Downloads a specific artifact from the cloud store into a local directory. Falls back to the local cache when the store is unreachable.

# Download a single file
dest = registry.download_artifact("run-id", "model/model.pkl", "./downloads/")
print(f"Saved to: {dest}")

# Download a whole model directory
dest = registry.download_artifact("run-id", "model", "./local_models/")

Parameters:

Parameter	Type	Description
run_id	str	ID of the run that logged the artifact
artifact_path	str	Relative artifact path as logged (e.g. "model/model.pkl")
local_path	str	Destination directory

Returns: str — absolute path to the downloaded file or directory.

Raises: ArtifactNotFoundError if the artifact cannot be found in the store or the local cache.

Models as artifacts: log_model() and load_model()

log_model() serializes your model to disk and uploads every resulting file to the artifact store in one call. load_model() resolves the registered model version, downloads any missing files from the store, then deserializes the model — so it works correctly from any machine regardless of where training happened.

# ── Training machine ──────────────────────────────────────────────────────────
with registry.start_run("bert-finetune-v3") as run:
    # train...
    registry.log_model(model, artifact_path="model", framework="pytorch")

registry.register_model("sentiment-bert", run_id=run.id)
registry.set_model_stage("sentiment-bert", version=3, stage="production")

# ── Any machine (CI, inference server, colleague's laptop) ───────────────────
# Model files are downloaded automatically from the artifact store if not cached
model = registry.load_model("sentiment-bert", stage="production")

log_model() parameters:

Parameter	Type	Default	Description
model	any	required	Model object (PyTorch, TensorFlow, sklearn, HuggingFace, or any picklable object)
artifact_path	str	"model"	Sub-path inside the artifact store
framework	str	auto-detected	"pytorch", "tensorflow", "sklearn", "huggingface", or "pickle"
metadata	dict	None	Arbitrary key-value metadata stored as run params

load_model() parameters:

Parameter	Type	Default	Description
model_name	str	required	Registered model name
version	int	None	Specific version to load. Mutually exclusive with stage
stage	str	None	Stage to load from: "development", "staging", "production", "archived"
framework	str	from run params	Override framework for deserialization

Viewing artifacts in the dashboard

Every artifact logged with log_artifact() or log_model() appears automatically in the Artifacts tab of the run's detail page in the Podstack dashboard. No extra steps are needed — the tab populates from the same store the SDK writes to.

The Artifacts tab shows:

Column	Description
Path	The relative artifact path as logged (e.g. model/model.pkl, plots/curves.png)
Type badge	File extension, color-coded by category — model weights, data files, images, configs, etc.
Size	Formatted file size (B / KB / MB)
Uploaded	Timestamp of when the file was stored
Download	One-click download button — opens a short-lived direct download link in the browser

A footer below the list shows the combined size of all artifacts for the run.

# Everything logged here shows up in the dashboard Artifacts tab
with registry.start_run("bert-finetune-v3") as run:
    registry.log_params({"lr": 2e-5, "epochs": 3})
    registry.log_metrics({"accuracy": 0.93})

    # These all appear as separate rows in the Artifacts tab
    registry.log_artifact("confusion_matrix.png", artifact_path="plots/confusion_matrix.png")
    registry.log_artifact("feature_importance.csv", artifact_path="analysis/features.csv")
    registry.log_model(model, artifact_path="model", framework="pytorch")
    # ↳ each model file (model.pkl, config.json, etc.) appears as its own row

Access control

Artifact upload and download URLs are issued by the registry API and require a valid API key and project membership. The URLs are short-lived, ensuring that access always reflects the current state of your project — a revoked key can no longer generate new URLs. Any member of a project can upload and download artifacts for runs within that project.

List and Browse

from podstack import registry

# List experiments
experiments = registry.list_experiments()

# List models
models = registry.list_models()

# List artifacts for a specific run
artifacts = registry.list_artifacts(run_id)

# Download a specific artifact to a local directory
dest = registry.download_artifact("run-id", "model/model.pt", "./downloads/")
print(f"Saved to: {dest}")

GPU Runner - Direct Code Execution

For running code strings directly on GPUs without decorators:

import podstack

podstack.init(api_key="your-api-key", project_id="your-project-id")

# Run code on a remote GPU
result = podstack.run_on_gpu('''
import torch
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"Memory: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB")
''', gpu="L40S")

print(result.output)
print(f"Success: {result.success}")
print(f"Duration: {result.duration_seconds}s")

Client API

For direct API access to notebooks and executions:

from podstack import Client

client = Client(api_key="your-api-key")

# Create a notebook
notebook = client.sync_create_notebook(name="experiment", gpu_type="L40S")
print(f"JupyterLab: {notebook.jupyter_url}")

# Run code
result = client.sync_run("print('Hello GPU!')", gpu_type="L40S")
print(result.output)

Error Handling

from podstack import (
    PodstackError,
    AuthenticationError,
    GPUNotAvailableError,
    RateLimitError,
    ExecutionTimeoutError
)

try:
    result = train()
except AuthenticationError:
    print("Invalid API key")
except GPUNotAvailableError as e:
    print(f"GPU not available")
except RateLimitError as e:
    print(f"Rate limited, retry after {e.retry_after}s")
except ExecutionTimeoutError as e:
    print(f"Execution timed out: {e.execution_id}")
except PodstackError as e:
    print(f"Error: {e.message}")

Configuration

import podstack

# Option 1: Initialize explicitly
podstack.init(
    api_key="your-api-key",
    project_id="your-project-id",
    api_url="https://api.podstack.ai/v1",       # optional
    registry_url="https://registry.podstack.ai"  # optional
)

# Option 2: Environment variables
# PODSTACK_API_KEY=your-api-key
# PODSTACK_PROJECT_ID=your-project-id
# PODSTACK_API_URL=https://api.podstack.ai/v1
# PODSTACK_REGISTRY_URL=https://registry.podstack.ai

# Option 3: Auto-init (set PODSTACK_AUTO_INIT=1)
# SDK auto-initializes from env vars at import time

License

MIT License - see LICENSE for details.