Skip to main content
aevyra-forge tunes your vLLM deployment overnight. Give it a model, a GPU, and a workload trace. It runs an autonomous loop — propose a config, boot the server, benchmark against your real workload, keep or revert, repeat. By morning you have a deployment recipe that beats hand-tuned defaults, with a full audit trail of every experiment.
vLLM exposes roughly 40 serving args. The defaults are conservative — designed to work safely on any GPU, not to max out any specific one. On a T4 with a chat workload the defaults leave significant throughput on the table:
Forge finds that gain by searching the joint space of batching, caching, and memory knobs — and keeping only the changes that improve the score on your actual workload.

​
Where Forge fits

Forge operates on the infrastructure layer. Where Reflex rewrites prompts and Origin diagnoses agent failures, Forge maximises throughput and minimises latency for a model that’s already doing the right thing.

​
The loop

Each iteration: the agent reads the playbook and experiment history, proposes one targeted change, and Forge measures whether it actually helps. The audit trail captures every decision — config, result, rationale — so you can see exactly how the winning recipe was found.

​
Tuning layers

Layer 1 has the highest leverage per experiment because it requires no recompilation. Forge escalates to Layer 2 when Layer 1 converges.

​
Works with any GPU and LLM

Auto-detects NVIDIA and AMD GPUs via nvidia-smi / rocm-smi. Works with any OpenAI-compatible LLM for the agent.
Python 3.10+. Apache-2.0 licensed.

Quick start

Run your first tuning session in 15 minutes

Tutorial

Dry-run walkthrough with real log output

Concepts: Recipe

The artifact Forge proposes, mutates, and keeps or reverts

Concepts: Playbook

The agent’s instruction manual