Contents Menu Expand Light mode Dark mode Auto light/dark, in light mode Auto light/dark, in dark mode Skip to content
FlowEdge
Logo
FlowEdge

Start

  • Overview
  • Getting Started
  • Python
  • LeRobot deployment adapter
  • Performance

Guides

  • Converter
  • Diffusion Policy
  • Fixed-shape Transformer baseline
  • Model import paths
  • Checkpoint preflight
  • Verification
  • Policy evaluation
  • Reproducible policy reports
  • Relay quickstart
  • Action delivery
  • DeadlineFlow execution plans
  • Deadline profile
  • Observability
  • Cooperative jobs
  • Generic job daemon
  • Sanitizer verification

Architecture

  • Overview
  • Memory Arena
  • Loader
  • Kernels
  • CUDA backend
  • Backbones
  • Heads
  • Deployment profile
  • Cooperative Execution
  • Model Porting Analysis

API

  • C-ABI
  • Python
  • Capabilities

Contribute

  • Contribute
  • Add a Head
  • Add a Backbone
  • Benchmark map

Project

  • Roadmap
  • Product Direction
  • FlowEdge on Tenstorrent: active program
  • Decisions
    • ADR 0001: CPU kernel strategy
    • ADR 0002: Mamba block and zero-alloc forward
    • ADR 0003: Public API boundary
    • ADR 0004: Decouple the head from the backbone
    • ADR 0005: Hand-assembled architectures, not a graph runtime
    • ADR 0006: Observation encoders are out of scope
    • ADR 0007: Optimize weight traffic, not the scan
    • ADR 0008: Lock-Free Thread Pool
    • ADR 0009: BF16 Weight Storage and Inline Widening
    • ADR 0010: External Conditions and Cooperative Solvers
    • ADR 0011: Versioned State and Request Contracts
    • ADR 0012: Keep the local inference relay beside Core
    • ADR 0013: Expose a typed SPSC client and test the process boundary
    • ADR 0014: Write canonical portable Relay traces
    • ADR 0015: Make deadline admission calibrated and observable
    • ADR 0016: Run independent head engines in a preallocated worker pool
    • ADR 0017: Share immutable weights and place outer workers explicitly
    • ADR 0018: Simulate actual worker lanes during deadline admission
    • ADR 0019: Record Relay metrics in fixed memory and export off the compute path
    • ADR 0020: Generalize Relay with non-owning cooperative jobs and portable capsules
    • ADR 0021: Add validated generic messages and frozen adapter registration
    • ADR 0022: Route generic jobs through caller-provisioned worker lanes
    • ADR 0023: Buffer generic job events before export
    • ADR 0024: Reuse bounded shared memory for generic job services
    • ADR 0025: Mamba streaming Relay adapter
    • ADR 0026: Bounded worker draining
    • ADR 0027: Gate action chunks at the controller boundary
    • ADR 0028: Separate generic-job data and administration planes
    • ADR 0029: Bound job QoS and worker supervision
  • FlowEdge Relay
  • LeRobot edge-inference RFC draft
Back to top
Copyright © 2026, FlowEdge
Made with Furo