Mimiry 0.2.0

Python SDK for Mimiry GPU compute (softlaunch). v1 wraps the existing /api/compute/v1/sessions API.

mimiry — Python SDK for Mimiry GPU compute

Status: v0.2.0 (alpha) Backend: softlaunch.mimiry.com (early beta)

Python-native interface for running GPU jobs on Mimiry: decorate a function, call .remote(), get the result back. v1 wraps the existing /api/compute/v1/sessions API — no backend changes required.

Install (editable, from this directory)

pip install -e .

Auth

The SDK uses SSH-JWT auth — the same SSH key registered on your Mimiry account. The fastest way to get set up is the interactive wizard, which generates a key (if needed), walks you through registering it in the portal, writes MIMIRY_SSH_KEY to your shell profile, and verifies the connection:

mimiry setup   # alias: mimiry init

To configure auth manually instead, point the SDK at your private key:

export MIMIRY_SSH_KEY=~/.ssh/mimiry

Or pass ssh_key_path= explicitly to mimiry.configure().

GPU types and providers

Not every GPU type is offered by every provider. Today (2026-05-31) T4 is only available on gcp (europe-west4-a); Verda offers higher-end GPUs (V100_16G, A100_80G, A100_160G, B200/B300/H200). Quota cap on the softlaunch account is 2 concurrent sessions.

Always check /availability before assuming a provider supports a given GPU:

mimiry availability --gpu-family T4

Pass provider="gcp" explicitly when the SDK's default routing is wrong for the GPU you want. v2 will auto-resolve this via /availability.

Python version compatibility

cloudpickle serializes the user's function as a Python code object. Code objects are not portable across major.minor Python versions — a function pickled by Python 3.12 cannot be unpickled by Python 3.10 (TypeError: code expected at most 16 arguments, got 18).

The SDK defaults @mimiry.function to an image based on nvcr.io/nvidia/cuda:12.6.2-runtime-ubuntu24.04, which ships Python 3.12 — matching recent Ubuntu / Debian / Fedora local installs. If you override the image, the container's python3 must be the same major.minor as your local interpreter, otherwise the bootstrap will write /tmp/mimiry_bootstrap_error: failed to unpickle payload and exit non-zero.

The simplest way to confirm a match is python3 --version locally and inside your image. v2 will auto-select a matching base image.

Quickstart — one-shot function

import mimiry

@mimiry.function(
    gpu="T4",
    provider="gcp",  # required for T4 today
    image="nvcr.io/nvidia/cuda:12.6.2-runtime-ubuntu24.04",
)
def gpu_info():
    import subprocess
    return subprocess.check_output(
        ["nvidia-smi", "--query-gpu=name,memory.total", "--format=csv"],
        text=True,
    )

print(gpu_info.remote())

Quickstart — raw bash command

import mimiry

result = mimiry.run(
    image="nvcr.io/nvidia/cuda:12.6.2-runtime-ubuntu24.04",
    gpu="T4",
    provider="gcp",
    command="nvidia-smi",
)
print(result.logs)

What works in v1

• @mimiry.function(gpu=..., image=...) decorator
• .remote(*args, **kwargs) — sync call, returns the function's return value
• .map(iterable) — sequential fanout (respects the 2-concurrent-session quota)
• Image.from_registry(uri).pip_install(...).apt_install(...) — basic image customisation (installs at container start; no real Dockerfile build)
• mimiry.run(image, gpu, command) — raw bash entrypoint
• SSH-JWT auth via existing key

What's deferred to v2 (needs backend changes)

• @mimiry.cls + @enter + @method (warm container, persistent state)
• Sub-second warm-call latency (needs per-region image cache + warm pool)
• .map() with real parallelism > 2 (needs quota raise)
• Native Python-object result return (currently via stdout-sentinel, 64KB-ish limit)
• Pre-built image catalog / fast image layer caching

Every .remote() call today provisions a fresh VM. Measured cold start on 2026-05-31: ~3 min wall clock end-to-end (submit → result returned), of which ~2 min is VM provisioning and image pull, ~30 s is the bootstrap apt / pip install layer inside the container.

Examples

See examples/:

• 01_hello.py — minimal nvidia-smi
• 02_cuda_probe.py — port of scripts/run_quick_test.sh to SDK
• 03_bash_command.py — mimiry.run() for ffmpeg-style jobs