Config-driven training setup
`soup init` creates a `soup.yaml`, and the site states that Soup writes much of the training configuration automatically, including task, quantization, learning rate, and epochs when rules determine them.
Soup CLI
ClaimSoup CLI is an open-source command-line tool for post-training and related model workflows, including fine-tuning, preference training, evaluation, merging, deployment, and data preparation. It is aimed at users who want a single offline CLI to manage model training and shipping on local or borrowed hardware.
Post-training and model ops in one CLI
Soup CLI is an offline, CLI-first toolkit for post-training language models and related workflows. It combines training setup, data preparation, evaluation, model merging, inference serving, export, and deployment commands in one interface.
Its core training workflow is designed to work from a config file generated by the tool, with rules that set many defaults automatically. The site emphasizes methods such as supervised fine-tuning and preference training, plus layer streaming so the frozen base model can stay in CPU RAM or NVMe instead of occupying full VRAM.
The product is positioned for users training on limited hardware, including cases where a model is too large to fit resident in GPU memory. The docs also show commands for migrating configs from other tools, browsing recipes, running eval gates, and pushing or deploying models to external runtimes.
Core capabilities
Layer streaming for oversized models
Soup can stream the frozen base layer-by-layer from CPU RAM or NVMe into VRAM so only one layer is resident at a time, which the site says supports training models that do not fit on the card.
Multiple training methods
The training docs and home page mention support for supervised fine-tuning and preference methods including DPO, ORPO, SimPO, and KTO, with the training page stating 23 task values in the schema.
Evaluation and save gating
`soup train` supports `--gate` to run evaluation suites at epoch boundaries and halt on regression, and the home page says Soup derives evals from the user’s own data and gates saves on a ship/don’t-ship verdict.
Migration and recipes
`soup migrate` imports configs from tools like LLaMA-Factory, Axolotl, and Unsloth, while `soup recipes` lets users list, inspect, search, and copy from 142 ready-made recipes.
Broader CLI workflow
The reference includes commands for chat, serving, export, merge, push, evaluation, data validation, deduplication, generation, and dataset publishing, making the tool a broader model ops CLI rather than a training-only utility.
Practical ways to use Soup CLI
Fine-tuning on limited GPU memory
Use Soup when a base model is too large to load fully into VRAM and you need a streaming approach that keeps the frozen base in RAM or on NVMe.
Preference training and alignment
Use it for DPO, ORPO, SimPO, or KTO runs where you want the reference model to come from the same streamed base rather than loading a second resident copy.
Config migration from another framework
Use `soup migrate` to convert existing setups from LLaMA-Factory, Axolotl, or Unsloth into Soup YAML with minimal manual rewriting.
Training with checkpoints and regression checks
Use gating and checkpoint-related flags when you want runs to stop on eval regressions or automatically resume and push checkpoints during training.
Prepare, evaluate, and ship from one CLI
Use Soup when you want a single workflow for data review, training, evaluation, adapter merging, deployment, and export to serving formats.
Pros and Cons
Pros
- Covers a wide training-and-shipping workflow in one CLI instead of splitting tasks across multiple tools.
- Supports offline use according to the site’s positioning.
- Layer streaming can make some models trainable on hardware that cannot hold the full frozen base in VRAM.
- Includes migration tools and recipes that reduce setup work for common starting points.
- Exposes evaluation gates and run-management commands that support more controlled training loops.
Cons
- The site describes layer streaming as beta, and notes it is slower than resident training.
- The strongest streaming claims are tied to specific methods and environments described on the site, so supported setups may be narrower than the full CLI surface suggests.
- Pricing information is not clearly broken out on the site text provided; the public pages mainly emphasize open-source access and documentation.
FAQ
What does Soup CLI do?
Soup CLI is a command-line tool for post-training workflows such as training, evaluation, merging, deployment, data preparation, and dataset handling.
Can it train models that do not fit fully in GPU memory?
Yes. The site says Soup can stream the frozen base from CPU RAM or NVMe layer by layer so the model does not have to stay resident in VRAM.
Which training methods are supported?
The site explicitly mentions SFT, DPO, ORPO, SimPO, and KTO, and the training docs describe 23 task values in the schema.
Can I use it with another tool’s config?
Yes. `soup migrate` is documented for importing configs from LLaMA-Factory, Axolotl, and Unsloth.
Is it only for training?
No. The CLI reference also documents chat, serving, export, merge, push, eval, and data commands.
Quick Facts
- Category
- Developer Tool
- Primary interface
- Command-line interface
- License
- Apache-2.0
- Source domain
- trysoup.dev
- Install command
- `pip install "soup-cli[train]"`
- Supported Python
- 3.10 to 3.12