CLI-first
Queueing, control, and inspection from familiar commands.
Skip to content
gflow01 02 03 04
gflowSingle-node scheduling for shared machines
Give one Linux machine a real job scheduler
Queue, run, and inspect GPU or CPU jobs on a shared workstation with a small CLI and a local daemon.
- Single-node by design
- GPU-aware scheduling
- tmux-backed execution
- MCP-ready for agents
Recoverable
Attach, follow logs, and redo failed work.
Explicit policy
Declare GPUs, VRAM limits, priorities, reservations, and dependencies.
$ gflowd up
daemon ready on 127.0.0.1:5577
$ gbatch --gpus 1 --project vision python train.py
submitted batch job 184
$ gqueue
JOBID NAME ST TIME GPU NODELIST(REASON)
184 train R 00:12:41 1 0
185 eval PD - 0 (WaitingForDependency)
$ gjob log 184
step=420 loss=0.184 throughput=178 img/s
Queue snapshot
Running04
Queued09
Held01
GPU policy
Visible0,1,2,3
SharedVRAM-aware
Why it exists
When one machine is no longer just yours
Ad hoc tmux sessions and manual GPU etiquette break down once a workstation is shared.
Without queue discipline
- Long jobs collide with interactive work.
- Failures are hard to recover because state and logs are scattered.
- GPU requests become tribal knowledge instead of policy.
With gflow
- A local daemon keeps state and queue control consistent.
- Every job has a lifecycle you can inspect and recover.
- Resources and dependencies are declared at submission time.
How it works
A familiar flow in four commands
gflowd upStart the local scheduler.
gbatch --gpus 1 python train.pySubmit a command or script with explicit resources.
gqueueInspect running, queued, or completed jobs.
gjob log <job_id>Follow logs or attach when a run needs attention.
Capabilities
Built for daily workstation use
Queueing and lifecycle
Submit, hold, release, cancel, update, and redo jobs with an inspectable state model.
GPU-aware scheduling
Request GPUs directly, enable shared mode, and set VRAM limits.
Workflow composition
Use dependencies, arrays, and parameter sweeps for multi-stage runs.
Operational visibility
Read queue state through tables, trees, JSON, CSV, or YAML.
Recoverable execution
Run each job in its own tmux session for direct logs and recovery.
Automation and AI integration
Expose scheduler actions through a local MCP server.
Where it fits
Common scenarios
Shared lab GPU server
Coordinate multiple researchers on one box with explicit scheduling rules.
Solo research machine
Keep long-running experiments structured and restartable.
Local evaluation pipelines
Chain prep, train, benchmark, and reporting jobs with dependencies.
Documentation paths
Start where you are
Install and launch
Install the binary, create config defaults, and start the daemon.
Open InstallationLearn the first workflow
Start the daemon, submit a job, inspect the queue, and read logs.
Open Quick StartJump to command reference
Use the cheat sheet for `gflowd`, `gbatch`, `gqueue`, `gjob`, `gctl`, and more.
Open ReferenceConnect your agents
Run `gflow` as a local MCP server for agent tooling.
Open AI IntegrationAI integration
Make scheduler operations available as tools
Run gflow as a local stdio MCP server so agent CLIs can inspect queues and drive scheduler workflows.
gflow mcp serveStart scheduling the machine you already have
Use the docs as an operator handbook.