Step 01Current
Declare the service, not a benchmark.
Model, request mix, SLO, hardware inventory, power envelope and failure policy become one versioned study brief.
- Input
- Workload + SLO
- Output
- Study manifest
- Guardrail
- Verify before action
FROM SIMULATION TO AUTONOMY
One system that studies, calibrates and deploys the fabric.
The future OpenFabric agent treats SimLLM and hardware as one evolving system. It predicts before deployment, learns from measured residuals, then verifies every change before acting.
Vision page / current boundaries remain visible belowINTERACTIVE WORKFLOW
The calibration loop
OF / AGENTDecideunder constraints
THE ARCHITECTURAL MENTAL MODEL
A twin is useful only when every layer can disagree visibly.
OpenFabric does not learn one opaque correction factor. Each boundary has an owner, an interface, a measurement and a residual. The agent can update one layer without hiding a missing mechanism in another.
01
Service
Arrivals / prefix reuse / SLO / framework scheduler
current02
Execution
Kernels / GPU / HBM / DMA / NCCL / WQE / NIC
next03
Fabric
Collectives / flows / queues / packets / recovery
current04
Evidence
TTFT / TPOT / FCT / counters / traces
current05
Agent
Search / calibrate / deploy / observe / retune
future01 / Current
Trace-backed prediction
Workload queueing, real scheduler records, compute estimates, collective traffic, packet backends and request metrics.
02 / Next
Calibrated runtime twin
GPU and NIC resource queues, PD KV transfer, capture-driven kernel tables and hardware residual accounting.
03 / Future
Verified one-click operation
An agent proposes, simulates, deploys within policy, observes and rolls back when evidence leaves its support envelope.
THE PRODUCT PROMISE