Today is a big milestone for RapidFire AI: we’re officially integrated into Hugging Face’s TRL documentation as a first-class integration. That means TRL users can now discover, install, and run RapidFire AI as the fastest way to compare many fine-tuning/post-training configurations—without changing their workflow.

Why this matters

Teams don’t have the time (or budget) to train one config after another. Our TRL integration lets you launch many TRL configurations concurrently—even on a single GPU—via adaptive, chunk-based scheduling. In internal benchmarks referenced in the TRL page, this delivers ~16–24× higher experimentation throughput than sequential runs, so you reach better models dramatically faster.

What you get, out of the box

Drop-in TRL wrappers — Use RFSFTConfig, RFDPOConfig, and RFGRPOConfig as near-zero-code replacements for TRL’s SFT/DPO/GRPO configs.

• Chunk-based concurrent training — We shard data and cycle configs at chunk boundaries to maximize GPU utilization and enable early, apples-to-apples, comparisons.

• Interactive Control Ops (IC Ops) — From the dashboard, Stop, Resume, Clone, and Clone & Warm-Start any run in flight to double-down on winners and pause stragglers—no job restarts required.

• Multi-GPU orchestration — The scheduler auto-distributes configs across available GPUs; you focus on models, not plumbing.

• MLflow-based dashboard — Real-time metrics, logs, and IC Ops in one place as soon as you start your experiment.

How it works

RapidFire AI slices your dataset into “chunks” and rotates configurations through the GPU at chunk boundaries. You get incremental signal on all configs quickly, while automatic checkpointing keeps training stable. Then, use IC Ops to adapt mid-flight—stop low-performers early and clone promising ones with tweaked hyperparameters (optionally warm-starting from the parent’s weights).

Supported TRL trainers

• SFT with RFSFTConfig

• DPO with RFDPOConfig

• GRPO with RFGRPOConfig

These are designed as drop-in replacements, so you keep your TRL mental model while gaining concurrency and control.

Get started in minutes

pip install rapidfireai

# (Optional) Authenticate to access gated models, then initialize & start RapidFire AI
huggingface-cli login --token YOUR_TOKEN
rapidfireai init
rapidfireai start

Once running, open the dashboard at http://0.0.0.0:3000, point your code at our TRL-compatible configs, and launch a grid or random search across multiple configurations. You’ll see all runs stream into the dashboard immediately—and you can steer them live with IC Ops.

Where to learn more

Official TRL integration

We built RapidFire AI so practitioners can iterate faster, learn more from the same GPUs, and ship better LLMs with confidence. Being featured in TRL’s docs validates that mission, and we’re just getting started.

If you try the integration, we’d love to hear how much faster your experimentation loop gets and what you’d like us to add next.