Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

Fine-tune and post-train LLMs in one command. No SSH, no config hell.

Website · Quick Start · Config · Docs · Commands · Models · Discord · Product Hunt

Soup turns the pain of LLM fine-tuning into a simple workflow. One config, one command, done.

pip install "soup-cli[train]"   # add [train] to fine-tune; bare `soup-cli` is the light CLI
soup init --template chat
soup train

Fine-tune an 8B model on a 4 GB laptop GPU. Layer streaming keeps the frozen base out of VRAM and feeds it to the GPU one decoder layer at a time. Measured on an RTX 3050 Laptop 4 GB: Llama-3.1-8B-Instruct + NF4 at 119.6 tok/s, 3.32 GB peak — bit-exact against a normal resident run, and reproduced independently on an H100 at 113.00 tok/s in the same 3.32 GB. (The tok/s figure was measured on v0.72.2, before the v0.73.0 correctness repair that cost −4.8% at 32B; it has not been re-run on a 4 GB card since.) Opt-in (stream_layers: true) and still BETA — how it works · all measurements · paper · check it yourself on a free Colab T4 (caps the process to 4 GB, then asserts a streamed model is bit-identical to a normal one)

_{Llama-3.1-8B-Instruct + NF4, LoRA, batch 1, seq 512 on an RTX 3050 Laptop 4 GB — 3.32 GB peak, 119.6 tok/s. Full video (90s)}

Training LLMs is still painful. Even experienced teams spend 30-50% of their time fighting infrastructure instead of improving models. Soup fixes that.

• Zero SSH. Never SSH into a broken GPU box again.
• One config. A simple YAML file is all you need.
• Auto everything. Batch size, GPU detection, quantization — handled.
• Works locally. Train on your own GPU with QLoRA. No cloud required.

v0.73.2 — the release gate stops lying in both directions. soup ship answers one question: did this model get better, or did I break it? Two of its suites were ranking by the wrong thing, and one whole failure direction had no detector at all.

• A suite scored 0.225 for a model that got it right 40/40. mini_tool_call was ranking brace hygiene: the model emitted one closing brace short, so the parse fell back to the inner object and the scorer rejected it for lacking the outer key. And mini_mmlu scored Llama-3.1-8B at 0.423 — below a 0.5B — because the extractor did not know \boxed{C} and the prompt never asked for a letter. Both fixed; 0.423 → 0.731.
• New: a benign-prompt axis. Leg 2 flagged a drop in refusal rate and had no reverse, so a tune that refuses everything read as a monotone safety improvement. Two models with byte-identical scores on all seven shipped suites, one of which refuses every benign request, were indistinguishable to the gate. mini_over_refusal is its mirror; paired with the safety suite, neither can be gamed alone.
• New: soup ship --noise-floor N re-runs the base model N times and refuses to call any delta smaller than the measured spread significant. Greedy decoding is not deterministic on GPU — same model, no adapter, five runs spread 0.015–0.020 against a 0.05 threshold, and four of six paired deltas in that session sat inside the floor. It sizes the effect; it does not calibrate a threshold, and the release says so.
• A caller error was indistinguishable from a regression. A non-callable generator scored 0.0 on three suites and raised on the others — and in leg 2 a 0.0 reads as "failed every item", i.e. it failed in the direction that looks like a finding.
• Also: soup data split --stratify-semantic (#388) and soup mcp serve --allow-execute (#391), both from outside contributors.

The measurement record for the previous release's VRAM work, published as written — including the three readings withdrawn during it — is benchmarks/gate-v0.73.1-measured-vram-fit.md.

# soup.yaml — then just `soup train --config soup.yaml`
training:
  stream_layers: true      # base streams out of VRAM; only the adapter trains
  quantization: 4bit       # NF4 — ~4x smaller store, so 8B fits a 4 GB card
  batch_size: 4            # bigger batches amortise the weight read
  stream_source: auto      # RAM when it fits, NVMe disk when it does not
  seed: 1234               # new in v0.73.0

Python 3.10–3.12 only. v0.73.0 adds the upper bound that was missing: on 3.13+, pip used to resolve untested PyTorch wheels that crash in the native extension before Soup runs at all.

soup reward synth references.jsonl -o reward.py --output-report calib.json

soup ship --base ./base --adapter ./my-lora --task-eval my_task.jsonl
#   exit 0 = SHIP · 2 = DON'T SHIP · 3 = bad flags · 1 = runtime error

Full history: CHANGELOG.md · GitHub Releases.

# Light core: CLI + config + data tools, no PyTorch
pip install soup-cli

# Add the training stack (torch, transformers, peft, trl, datasets, …)
pip install "soup-cli[train]"

# Everything (train + serve + ui + data) in one shot
pip install "soup-cli[all]"

# Or from GitHub (latest dev)
pip install git+https://github.com/MakazhanAlpamys/Soup.git

The full extras table (fast, mlx, serve, eval, ui, vision, audio, …) lives in docs/models.md.

Double quotes, not single. "soup-cli[train]" is the only spelling that works in every shell — cmd.exe, PowerShell, bash and zsh. If you copied 'soup-cli[train]' from an older tutorial and pip rejected it, that is the reason: why, and the exact error.

soup init, soup data …, and the other data/inspection commands work on the light install. Fine-tuning (soup train) needs the [train] extra.

soup init                       # interactive wizard
soup init --template chat       # or start from a template

Templates: chat, code, tool-calling, medical, reasoning, vision, kto, orpo, simpo, ipo, bco, rlhf, pretrain, moe, longcontext, embedding, audio.

soup train --config soup.yaml                 # LoRA, quantization, batching — all handled
soup chat  --model ./output                    # talk to your model
soup push  --model ./output --repo you/my-model

soup merge  --adapter ./output                              # merge LoRA into the base
soup export --model ./output --format gguf --quant q4_k_m   # GGUF for Ollama / llama.cpp

More export targets (ONNX, TensorRT, AWQ, GPTQ, BitNet) and deployment options live in docs/serving-and-export.md.

A complete soup.yaml:

base: meta-llama/Llama-3.1-8B-Instruct
task: sft
# backend: unsloth  # 2-5x faster, pip install "soup-cli[fast]"

data:
  train: ./data/train.jsonl
  format: alpaca
  val_split: 0.1

training:
  epochs: 3
  lr: 2e-5
  batch_size: auto
  lora:
    r: 64
    alpha: 16
  quantization: 4bit

output: ./output

config/schema.py is the single source of truth for every field. Advanced data, training, and PEFT options are documented under Documentation.

The full feature reference lives in docs/. Start here:

Alpaca, ShareGPT, ChatML, preference pairs (DPO / ORPO / SimPO / IPO / KTO), vision, audio, ASR, plaintext, embedding, RAFT and more — all auto-detected from JSONL, JSON, CSV, Parquet or TXT, so in most cases you point data.train at a file and nothing else changes. Schemas with a worked example per format, plus the data pipeline (remote URIs, streaming, sharding, interleaving, vocab expansion, document ingestion), are in docs/data.md.

soup train  --config soup.yaml        # train (SFT/DPO/GRPO/PPO/KTO/ORPO/SimPO/IPO/...)
soup infer  --model ./output --input prompts.jsonl   # batch inference
soup chat   --model ./output          # interactive chat
soup serve  --model ./output          # OpenAI-compatible API server
soup merge  --adapter ./output        # merge LoRA into the base model
soup export --model ./output --format gguf           # export for deployment
soup eval   benchmark --model ./output               # evaluate
soup data   inspect ./data/train.jsonl               # dataset stats
soup recipes list                     # 100+ ready-made model recipes
soup autopilot --model <id> --data d.jsonl --goal chat  # zero-config
soup doctor                           # check GPU / deps / environment

The complete command list is in docs/commands.md.

Soup works with any text-generation model on the HuggingFace Hub — if it loads with AutoModelForCausalLM, it works, zero config changes. Llama 3.x/4, Qwen 2.5/3, Gemma 3, Mistral, Mixtral, DeepSeek R1/V3, Phi-4, and 100+ others ship as ready-made recipes (soup recipes list).

Full model + vision tables and the optional-extras matrix are in docs/models.md.

Run Soup without installing CUDA or PyTorch locally (image published to GHCR on every release):

docker pull ghcr.io/makazhanalpamys/soup:latest
docker run --gpus all -v $(pwd):/workspace ghcr.io/makazhanalpamys/soup train --config soup.yaml
docker compose up   # or build locally

• Python 3.10, 3.11 or 3.12 (those are the versions CI tests; 3.13+ is not supported yet because the PyTorch stack has not been validated there)
• GPU with CUDA (recommended), Apple Silicon (MPS), or CPU (experimental — very slow)
• 8 GB+ VRAM for 7B models with QLoRA

All training tasks run on CPU for testing (quantization auto-disabled). Optional extras (train, all, fast, vision, qat, serve, serve-fast, ui, eval, deepspeed, liger, mlx, onnx, tensorrt, …) are listed in docs/models.md.

soup doctor    # GPU, system resources, dependencies, and version in one place

• ImportError: DLL load failed while importing _C (Windows) — reinstall PyTorch for your CUDA version: pip install torch --index-url https://download.pytorch.org/whl/cu121.
• soup version ≠ pip show soup-cli — multiple Python installs; use a virtualenv.

git clone https://github.com/MakazhanAlpamys/Soup.git
cd Soup
pip install -e ".[dev]"

ruff check src/soup_cli/ tests/    # lint
pytest tests/ -v                   # unit tests (fast, no GPU)
pytest tests/ -m smoke -v          # smoke tests (downloads a tiny model, trains)

pre-commit install                 # optional: ruff lint+format on commit

See CONTRIBUTING.md for the full workflow and SECURITY.md to report a vulnerability.

Soup is Apache-2.0 and free — and stays that way. It is built and maintained in the open on a single 4 GB laptop, which is why every performance number in these docs is measured rather than claimed.

If Soup saved you a training run, starring the repo helps most, and it costs nothing. If you would like to fund the work directly:

❤️ Donate — one-off, any amount (use Change amount on the checkout page). Payments are processed by Stripe under the maintainer's registered business, MePlay, Inc. — that name, not "Soup", is what appears on the checkout page and on your card statement.

Donations buy GPU time for the hardware-gated work — multi-GPU, 8B+ validation, Apple Silicon — that a single 4 GB laptop cannot reach.

The other way to move exactly those items is hardware itself. They ship behind honest "requires <hardware>" gates rather than unverified claims, so if you have access to a bigger box — or GPU credits going unused — running one of the help wanted issues and posting the numbers helps as much as funding the GPU time would. Those issues say exactly what is blocked on hardware today.

Built by the community ❤️ — thank you to everyone who has contributed. See CONTRIBUTORS.md.

Bugs and feature requests belong in the issue tracker, questions in Discussions — both get answered faster and help the next person with the same problem.

For live chat, setup help, and everything that reads better as a conversation, join the Discord. Anything that should still be findable in six months belongs in Issues or Discussions — a Discord answer helps one person, an issue helps everyone who hits the same thing. The Code of Conduct applies there too.

For anything that does not fit in public — security reports (see SECURITY.md), Code of Conduct matters, or press — email team@trysoup.dev. That is the project address and the right one for anything Soup-related. makazanalpamys@gmail.com is the maintainer's personal address; it reaches the same person and is a fine fallback.

Layer streaming — training an 8B model on a 4 GB laptop GPU by streaming the frozen base from host RAM one decoder layer at a time — is described in a preprint, together with the correctness protocol that verifies a streamed run against a resident one (forward and backward stated separately, because they are two claims and not one).

Makazhan, A. (2026). Exact Layer Streaming: LoRA Fine-Tuning of an 8B Model on a 4 GB Laptop GPU (v3). Zenodo. https://doi.org/10.5281/zenodo.21918325

Version 3 (13 August 2026) is current. The title and the claim are unchanged — 8B on 4 GB — and no measured number has changed since v1. What v3 does is withdraw an explanation we had published, which is also the shortest way to describe what the paper is for:

• Retracted in v3: "layer streaming is bound by host-to-device transfer, not by the GPU." That was an inference from the H100 replication below, and it had never been measured. We measured it on 11 August and it is false at the published configuration: deleting every host-to-device byte buys 1.4%, the compute stream waits on a copy for 0.20% of the step, and the step runs at 71.3% of that card's same-session GEMM ceiling. The largest streaming-specific cost is the per-layer NF4 dequantisation, at 9.8% (the record). Every measurement stands; the replication survives in a weaker form — the constraint is common to both machines and is not the GPU's compute.
• Replication on hardware nothing like the original (added in v2): 119.6 tok/s on the RTX 3050 against a median 113.00 on an H100, at the same 3.32 GB peak.
• A silent wrong-gradient defect, found and repaired. On NF4 above ~165 MiB per layer the forward stayed bit-exact and the loss curve looked healthy while the gradients were wrong. The cause is named in the upstream library and reported there; the repair is gated against controls on real 32B and 72B.
• Bit-exactness at real model sizes instead of three-layer toys: forward from 0.5B to 72B, backward at 8B and 14B.
• Trained-model quality, measured for the first time, and indistinguishable from a resident run.
• A comparison against DeepSpeed — including the result that does not flatter us: eight cards of ZeRO-3 are slower than one card training resident.
• The limitations section rewritten: of v1's ten items, one closed and four more narrowed, and seven new ones added.

Cite the version you used. 10.5281/zenodo.21771064 is the concept DOI and always resolves to the latest version (v3 today); v1 and v2 remain citable at their own version DOIs and are not edited — the retraction above is a new version precisely so that the record of what we claimed, and when, stays intact.

The measurement records behind every number in it are in benchmarks/, published as written — including the failures, the assumptions that turned out wrong, and the numbers that were measured and then discarded.

@misc{makazhan2026exact,
  title        = {Exact Layer Streaming: LoRA Fine-Tuning of an 8B Model on a 4 GB Laptop GPU},
  author       = {Makazhan, Alpamys},
  year         = {2026},
  publisher    = {Zenodo},
  version      = {v3},
  doi          = {10.5281/zenodo.21918325},
  url          = {https://doi.org/10.5281/zenodo.21918325}
}

Apache-2.0. Copyright © the Soup contributors.