Fireworks fine-tuned models only serve on on-demand deployments, never serverless

Fireworks fine-tuned models only serve on on-demand deployments, never serverless

TL;DR

Context: Fireworks supervised fine-tuning: trained models can only be served on on-demand deployments Fireworks' supervised fine-tuning guide states it plainly: deploying a trained model to an on-demand (dedicated) deployment is the only supported method for serving trained models. A LoRA you just trained will not serve on serverless at all. After training completes, evaluate the model before holding dedicated capacity for production serving, then deploy with a deployment shape (or default).

Use this when you run into the situation in the title.

When not to use this skill: unrelated tasks. It covers only the procedure above.

Compatibility

The steps above apply to the commands named in them. This skill does not pin a version, so if a flag looks different on your machine, check your installed version's docs first.

Details

Your Fireworks fine-tuned LoRA cannot run on serverless The most common Fireworks fine-tuning failure is assuming the trained model works like any other serverless model. It does not. The official guide says deploying a trained model to an on-demand (dedicated) deployment is the only supported method for serving trained models. Serverless is not an option for fine-tuned weights. The correct order of operations: 1. Train the SFT job. 2. Evaluate the model before you pay for dedicated capacity in production. Do not hold a deployment while you decide whether the model is any good. 3. Deploy to on-demand with a validated shape, or pass default and let Fireworks pick. People burn money by deploying first and evaluating second, and they burn time by pointing their inference client at accounts/fireworks/models/[name] expecting their LoRA to answer. It will not: only the deployment identifier serves the fine-tuned weights.

Source: Fireworks supervised fine-tuning: trained models can only be served on on-demand deployments Fireworks' supervised fine-tuning guide states it plainly: deploying a trained model to an on-demand (dedicated) deployment is the only supported method for serving trained models. A LoRA you just trained will not serve on serverless at all. After training completes, evaluate the model before holding dedicated capacity for production serving, then deploy with a deployment shape (or default).

Verified resolution: # Your Fireworks fine-tuned LoRA cannot run on serverless The most common Fireworks fine-tuning failure is assuming the trained model works like any other serverless model. It does not. The official guide says deploying a trained model to an on-demand (dedicated) deployment is the only supported method for serving trained models. Serverless is not an option for fine-tuned weights. The correct order of operations: 1. Train the SFT job. 2. Evaluate the model before you pay for dedicated capacity in production. Do not hold a deployment while you decide whether the model is any good. 3. Deploy to on-demand with a validated shape, or pass default and let Fireworks pick. People burn money by deploying first and evaluating second, and they burn time by pointing their inference client at accounts/fireworks/models/[name] expecting their LoRA to answer. It will not: only the deployment identifier serves the fine-tuned weights.

Source URL: https://docs.fireworks.ai/fine-tuning/fine-tuning-models