Serverless GPU infrafor agents withsub-5s cold boots.
Graphene gives humans and coding agents a CLI-first GPU runtime with production-equivalent CUDA, CAS/NVMe caching, and no SDK rewrite or Docker rebuild loop.
brew install graphene-sh/tap/graphene
graphene version
graphene loginInit, add, run. The GPU path stays CLI-native.
A project locks its runtime, dependencies become cached artifacts, and the Python entrypoint starts on a live GPU.
Project config and GPU runtime are locked.
Dependencies become a content-addressed runtime artifact.
The same command starts on a production GPU.
01 from transformers import pipeline0203 classifier = pipeline(04 "zero-shot-classification",05 device=0,06 )0708 print(classifier(09 "agents need fast GPUs",10 ["ready_to_ship", "blocked"],11 ))
T4 and L4 are live. Larger accelerators are queued next.
The public runtime catalog exposes the GPUs users can actually run today. The rest of the rate card stays visible so teams can plan for heavier agent and training workloads.
These are the GPUs exposed through the CLI runtime selector and backed by the current GCP deployment path.
Do not reshape your app for someone else's SDK.
Most GPU serverless products ask you to wrap functions, decorators, or app objects. Graphene starts from the command you already run.
graphene run -- python inference.pyRun an existing script on a cloud GPU without moving it into a new SDK shape.
graphene run -- uvicorn app:appExpose the same service entrypoint you already use locally.
graphene run --train train.pySubmit longer GPU jobs while keeping CLI logs and status attached.
The platform handles runtime selection, source sync, scheduling, and output streaming around your command instead of forcing a new application model.
Startup measured in seconds, not deployment cycles.
Graphene moves slow work out of the hot path. Runtime and dependency artifacts are prepared before a GPU run, then placed close to the selected node.
base + Python + CUDA
uv lock + installed tree
small snapshot diff
model and data assets
run path: snapshot code -> schedule -> execute
Stop debugging CUDA drift after deployment.
Graphene locks Python, CUDA, framework versions, and GPU class into a runtime profile. Local development invokes the same production-equivalent runtime that serves the job.
No repeated Docker build loop on the GPU hot path.
Graphene treats heavy layers and fast-changing code differently. Runtime and dependency artifacts are content-addressed and reused; code moves as a small snapshot.