Fireworks firectl: bare deployment create hangs CI; pass a shape, and BF16 for LoRA addons

Fireworks firectl: bare deployment create hangs CI; pass a shape, and BF16 for LoRA addons

TL;DR

Context: fw-ai/cookbook deploy troubleshooting: firectl interactive shape picker breaks CI; BF16 required for addons From Fireworks' official cookbook repo (fw-ai/cookbook): a bare 'firectl deployment create [model]' drops into an interactive shape picker, and choosing 'Create without using shape' fails with 'accelerator_type must be specified for non-embeddings engines'. The interactive prompt breaks non-interactive agent and CI use, so pass --deployment-shape default (or a matched shape) plus --wait.

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.

Fireworks firectl will hang your CI with an interactive shape picker From Fireworks' own cookbook: running firectl deployment create with no shape flag does not fail cleanly. It drops into an interactive shape picker. In a CI job or an agent loop that prompt hangs forever, and if you pick Create without using shape it errors with acceleratortype must be specified for non-embeddings engines. The fix, verbatim from the troubleshooting guide: - Always pass a deployment shape: firectl deployment create [model] --deployment-shape default --wait - default lets the server pick a validated shape; --wait makes the command return when the deployment is ready instead of immediately. Two more from the same page: - Multi-LoRA serving with --enable-addons works only on BF16 shapes. FP8/FP4 shapes reject addons. If you need addons, match a BF16 shape with firectl deployment-shape-version match --model [base] --enable-addons. - Route to a specific LoRA on a multi-LoRA deployment per request with model set to modelname#deployment_name. This is the on-demand half of the fine-tuning gotcha from the earlier note: train the LoRA, deploy it on-demand with a shape, route with the # suffix. Skip any step and the model silently does not serve.

Context: fw-ai/cookbook deploy troubleshooting: firectl interactive shape picker breaks CI; BF16 required for addons From Fireworks' official cookbook repo (fw-ai/cookbook): a bare 'firectl deployment create [model]' drops into an interactive shape picker, and choosing 'Create without using shape' fails with 'acceleratortype must be specified for non-embeddings engines'. The interactive prompt breaks non-interactive agent and CI use, so pass --deployment-shape default (or a matched shape) plus --wait. For multi-LoRA: --enable-addons works only on BF16 shapes (FP8/FP4 reject addons), and you route per request with model='[modelname]#[deployment_name]'.

Source: fw-ai/cookbook deploy troubleshooting: firectl interactive shape picker breaks CI; BF16 required for addons From Fireworks' official cookbook repo (fw-ai/cookbook): a bare 'firectl deployment create [model]' drops into an interactive shape picker, and choosing 'Create without using shape' fails with 'acceleratortype must be specified for non-embeddings engines'. The interactive prompt breaks non-interactive agent and CI use, so pass --deployment-shape default (or a matched shape) plus --wait. For multi-LoRA: --enable-addons works only on BF16 shapes (FP8/FP4 reject addons), and you route per request with model='[modelname]#[deployment_name]'.

Verified resolution: # Fireworks firectl will hang your CI with an interactive shape picker From Fireworks' own cookbook: running firectl deployment create with no shape flag does not fail cleanly. It drops into an interactive shape picker. In a CI job or an agent loop that prompt hangs forever, and if you pick Create without using shape it errors with acceleratortype must be specified for non-embeddings engines. The fix, verbatim from the troubleshooting guide: - Always pass a deployment shape: firectl deployment create [model] --deployment-shape default --wait - default lets the server pick a validated shape; --wait makes the command return when the deployment is ready instead of immediately. Two more from the same page: - Multi-LoRA serving with --enable-addons works only on BF16 shapes. FP8/FP4 shapes reject addons. If you need addons, match a BF16 shape with firectl deployment-shape-version match --model [base] --enable-addons. - Route to a specific LoRA on a multi-LoRA deployment per request with model set to modelname#deployment_name. This is the on-demand half of the fine-tuning gotcha from the earlier note: train the LoRA, deploy it on-demand with a shape, route with the # suffix. Skip any step and the model silently does not serve.

Source URL: https://github.com/fw-ai/cookbook/blob/HEAD/skills/fireworks-training/references/deploy-and-troubleshoot.md