mimiry-cli 0.1.3

Python SDK for the Mimiry GPU Cloud API

Mimiry Python SDK

Python client for the Mimiry GPU Cloud API. Deploy GPU instances, submit batch jobs, and manage cloud resources programmatically.

Installation

Prerequisites (Linux/WSL)

Modern Linux distributions (Debian 12+, Ubuntu 23.04+, WSL) require a virtual environment due to PEP 668:

# Create and activate a virtual environment
python3 -m venv ~/.venvs/mimiry
source ~/.venvs/mimiry/bin/activate

Install from PyPI

pip install mimiry-cli          # SDK only
pip install "mimiry-cli[cli]"   # SDK + CLI

Verify installation

mimiry version
# mimiry 0.1.3

Quick Start

from mimiry import MimiryClient

client = MimiryClient(api_key="mky_your_key_here")

# List available GPUs (defaults to Verda)
currency = "eur"
symbol = {"eur": "€", "usd": "$"}.get(currency, currency.upper())
for gpu in client.list_instance_types(currency=currency):
    print(f"{gpu['instance_type']} — {symbol}{gpu['price_per_hour']}/hr")

# List Scaleway GPUs
for gpu in client.list_instance_types(provider="scaleway"):
    print(f"{gpu['instance_type']} — €{gpu['price_per_hour']}/hr")

# Submit a job
job = client.submit_job(
    name="training-run",
    instance_type="1V100.6V",
    image="ubuntu-22.04-cuda-12.0",
    location="FIN-01",
    ssh_key_ids=["your-key-uuid"],
    startup_script="#!/bin/bash\npython train.py",
    auto_shutdown=True,
)
print(f"Job {job['id']} submitted — status: {job['status']}")

Authentication

1. Create an API key from the Mimiry Dashboard → API Keys
2. Pass it to the client:

client = MimiryClient(api_key="mky_your_key_here")

API keys require scopes for the endpoints you want to access:

Scope	Endpoints
jobs:read	List/get jobs
jobs:write	Submit/cancel jobs
instances:read	List GPUs, locations, availability, images, providers
ssh_keys:read	List SSH keys
ssh_keys:write	Add/delete SSH keys
registry:read	List registry credentials
registry:write	Add/delete registry credentials

Supported Providers

The SDK is provider-agnostic. Pass provider="..." to target a specific backend.

Provider	Slug	GPU Types	Locations
Verda	verda (default)	V100, A100, H100, etc.	FIN-01 (Helsinki)
Scaleway	scaleway	H100, L4, L40S, B300	fr-par-2 (Paris), nl-ams-1 (Amsterdam), pl-waw-2 (Warsaw), + more

Note: The legacy slug datacrunch is still accepted as an alias for verda for backward compatibility.

Scaleway Instance Types

Scaleway GPU types follow the pattern {GPU}-{count}-{memory}:

Instance Type	GPU	VRAM	Example
H100-1-80G	1× H100	80 GB	Single H100
H100-2-80G	2× H100	160 GB	Dual H100
L40S-1-48G	1× L40S	48 GB	Single L40S
L4-1-24G	1× L4	24 GB	Single L4

Scaleway Locations

Code	Name
fr-par-1	Paris 1
fr-par-2	Paris 2 (GPU)
fr-par-3	Paris 3
nl-ams-1	Amsterdam 1
nl-ams-2	Amsterdam 2
pl-waw-2	Warsaw 2
pl-waw-3	Warsaw 3

API Reference

Jobs

# List all jobs
jobs = client.list_jobs()

# Get job details
job = client.get_job("job-uuid")

# Submit a job (Verda — default)
job = client.submit_job(
    name="my-job",
    instance_type="1V100.6V",
    image="ubuntu-22.04-cuda-12.0",
    location="FIN-01",
    ssh_key_ids=["key-uuid"],
    startup_script="#!/bin/bash\nnvidia-smi",
    auto_shutdown=True,
    heartbeat_timeout_seconds=1800,  # optional, default 600
    max_runtime_seconds=7200,        # optional, no default
)

# Submit a job on Scaleway
job = client.submit_job(
    name="scaleway-training",
    instance_type="H100-1-80G",
    image="ubuntu_jammy",
    location="fr-par-2",
    ssh_key_ids=["key-uuid"],
    startup_script="#!/bin/bash\nnvidia-smi",
    provider="scaleway",
    auto_shutdown=True,
)

# Cancel a job
client.cancel_job("job-uuid")

# Wait for a job to finish (polls every 10s, timeout 1h)
result = client.wait_for_job("job-uuid", poll_interval=10, timeout=3600)

# Submit and wait in one call
result = client.submit_job_and_wait(
    name="my-job",
    instance_type="1V100.6V",
    image="ubuntu-22.04-cuda-12.0",
    location="FIN-01",
    ssh_key_ids=["key-uuid"],
    startup_script="#!/bin/bash\npython train.py",
)

Streaming Logs & Metrics

Jobs automatically stream stdout/stderr to the dashboard in real-time (every 15s). You can also emit structured metrics (loss, accuracy, etc.) that will appear as interactive charts.

Emitting Metrics (File Convention)

Write JSONL to /tmp/mimiry_metrics.jsonl — no SDK dependency required:

import json

# In your training loop:
for epoch in range(num_epochs):
    loss = train_one_epoch()
    accuracy = evaluate()
    
    # Write metrics — they appear as live charts in the dashboard
    with open("/tmp/mimiry_metrics.jsonl", "a") as f:
        f.write(json.dumps({
            "step": epoch,
            "loss": float(loss),
            "accuracy": float(accuracy),
            "learning_rate": optimizer.param_groups[0]["lr"],
        }) + "\n")

Rules:

• Each line must be valid JSON
• Include a step or epoch field for the X-axis
• All numeric fields are automatically plotted
• The agent watches the file every 10s and streams new entries to the dashboard
• No SDK import needed — works with any framework (PyTorch, TensorFlow, JAX, etc.)

PyTorch Example

import json, torch

metrics_file = "/tmp/mimiry_metrics.jsonl"

for epoch in range(100):
    model.train()
    total_loss = 0
    for batch in train_loader:
        loss = train_step(model, batch)
        total_loss += loss.item()
    
    avg_loss = total_loss / len(train_loader)
    val_acc = evaluate(model, val_loader)
    
    with open(metrics_file, "a") as f:
        f.write(json.dumps({
            "step": epoch,
            "train_loss": avg_loss,
            "val_accuracy": val_acc,
            "gpu_memory_mb": torch.cuda.max_memory_allocated() / 1e6,
        }) + "\n")

Viewing Logs & Metrics

• Dashboard: Click any job → Logs tab shows streaming output, Metrics tab shows interactive charts
• API: Logs and metrics are stored in job_logs and job_metrics tables, queryable via the Supabase client

Instance Types

# List all GPU types with pricing (EUR default, Verda default)
gpus = client.list_instance_types()

# List in USD
gpus = client.list_instance_types(currency="usd")

# List Scaleway GPU types
gpus = client.list_instance_types(provider="scaleway")

Availability

# Check all availability (Verda)
available = client.check_availability()

# Check specific instance type
available = client.check_availability(instance_type="1V100.6V")

# Check Scaleway availability
available = client.check_availability(provider="scaleway")
available = client.check_availability(instance_type="H100-1-80G", provider="scaleway")

Locations

# Verda locations
locations = client.list_locations()

# Scaleway locations
locations = client.list_locations(provider="scaleway")

OS Images

# All images (Verda)
images = client.list_images()

# Images compatible with a specific GPU type
images = client.list_images(instance_type="1V100.6V")

# Scaleway images
images = client.list_images(provider="scaleway")

Providers

providers = client.list_providers()
# Returns: [{"name": "Verda", "slug": "verda"}, {"name": "Scaleway", "slug": "scaleway"}]

SSH Keys

SSH keys are synced to all active providers automatically when created via the API.

# List keys
keys = client.list_ssh_keys()

# Add a key (synced to Verda + Scaleway)
key = client.add_ssh_key("my-laptop", open("~/.ssh/id_ed25519.pub").read())

# Delete a key (removed from all providers)
client.delete_ssh_key("key-uuid")

Container Registry Credentials

Store credentials for private container registries (Docker Hub, AWS ECR, GHCR, etc.). When you submit a job with a container_image and registry_credential_id, the platform automatically runs docker login + docker pull before your startup script.

# List saved credentials
creds = client.list_registry_credentials()

# Add a generic credential (Docker Hub, GHCR, etc.)
cred = client.add_registry_credential(
    name="Docker Hub",
    registry_url="docker.io",
    username="myuser",
    password="dckr_pat_xxxxxxxxxxxx",
    is_default=True,
)

# Add an AWS ECR credential
# Your IAM credentials are stored securely. At job dispatch time, the platform
# exchanges them for a short-lived ECR token (valid 12h) server-side.
# Your AWS credentials never touch the compute node.
ecr_cred = client.add_registry_credential(
    name="My ECR",
    registry_url="123456789.dkr.ecr.eu-west-1.amazonaws.com",
    username="AKIAIOSFODNN7EXAMPLE",        # AWS Access Key ID
    password="wJalrXUtnFEMI/K7MDENG/bPx",   # AWS Secret Access Key
    registry_type="aws_ecr",
    aws_region="eu-west-1",
)

# Delete a credential
client.delete_registry_credential("credential-uuid")

# Submit a job with a private container image
job = client.submit_job(
    name="bio-pipeline",
    instance_type="1V100.6V",
    image="ubuntu-22.04-cuda-12.0",
    location="FIN-01",
    ssh_key_ids=["key-uuid"],
    container_image="ghcr.io/myorg/pipeline:v2",
    registry_credential_id=cred["id"],
    startup_script="docker run ghcr.io/myorg/pipeline:v2 --data /mnt/data",
    auto_shutdown=True,
)

CLI

For terminal usage, see the CLI Guide.

Error Handling

The SDK raises typed exceptions for API errors:

from mimiry import MimiryClient, AuthenticationError, InsufficientCreditsError

client = MimiryClient(api_key="mky_...")

try:
    job = client.submit_job(...)
except AuthenticationError:
    print("Invalid API key")
except InsufficientCreditsError:
    print("Not enough credits — top up at the dashboard")
except MimiryError as e:
    print(f"API error [{e.status_code}]: {e.message}")

Exception	HTTP Status	Meaning
AuthenticationError	401	Invalid or missing API key
InsufficientCreditsError	402	Not enough credits
InsufficientScopeError	403	API key lacks required scope
NotFoundError	404	Resource not found
RateLimitError	429	Too many requests
ServerError	5xx	Server-side error
MimiryError	other	Catch-all base exception

Context Manager

The client can be used as a context manager to ensure connections are closed:

with MimiryClient(api_key="mky_...") as client:
    jobs = client.list_jobs()

Configuration

client = MimiryClient(
    api_key="mky_...",
    base_url="https://custom-endpoint.example.com",  # override API URL
    timeout=60.0,       # request timeout in seconds (default 30)
    max_retries=5,      # retry count for transient failures (default 3)
)

Requirements

• Python ≥ 3.8
• httpx ≥ 0.24.0

License

MIT