Jetashree Ravi, an applied machine-learning lead at Fireworks AI, describes why moving from closed models to open alternatives often requires more than changing an endpoint. The talk separates model quality from serving performance, connecting task-specific training to inference-stack optimization for a workload’s input length, output length, cache behavior and latency targets.
The presentation explains custom CUDA kernels, speculative decoding, quantization and disaggregated serving, then contrasts three training interfaces: an assisted training agent, managed training and an API for custom loops and loss functions. The demonstrations show training and evaluation datasets, hyperparameter selection, progress inspection and metrics rather than a promise that fine-tuning automatically improves every use case.
The proposed feedback loop trains a checkpoint, deploys it, monitors the workload and uses new data for another training iteration. Hot-loading is presented as a way to swap checkpoints without downtime, while keeping training and inference numerically consistent addresses deployment-quality differences. Customer results are examples reported by Fireworks AI, not universal performance guarantees.
Watch on YouTube
