Unsloth Fine-Tuning

Overview

Unsloth trains LLMs with custom kernels that cut VRAM use and step time without changing the math, which makes single-GPU fine-tuning practical on hardware that would otherwise OOM. This skill covers Unsloth Core — the Python API — because that is what an agent can drive programmatically; the Desktop app and Studio web UI are interactive and out of scope.

The hard parts of an Unsloth run are not the training call. They are sizing the job against available VRAM, getting the chat template and loss masking right, and choosing an export format the target runtime can actually load. This skill covers those three.

When to Use This Skill

• Use when fine-tuning an LLM on one GPU and VRAM is the binding constraint.
• Use when a training run OOMs and needs to be resized rather than rewritten.
• Use when doing preference or RL post-training (GRPO, DPO) on consumer hardware.
• Use when a fine-tuned model must be exported to GGUF, vLLM, or merged 16-bit weights.
• Use when a fine-tune "ran fine" but the model's output format is wrong — usually a chat template or loss-masking bug, not a hyperparameter one.

Do not use this skill when

• The training is multi-node or large-scale multi-GPU. Use plain TRL with Accelerate/DeepSpeed.
• The architecture is unsupported by Unsloth. Fall back to TRL; do not force it.
• The user wants managed cloud training. That is Hugging Face Jobs, not local Unsloth.
• The user wants the Desktop or Studio GUI. Point them at the installer, not this skill.

How It Works

Step 1: Size the run before writing code

VRAM is the constraint that decides everything else. Estimate weights first, then leave room for activations and optimizer state:

Load mode	Weight cost	8B model	Use when
load_in_4bit (QLoRA)	~0.55 GB per 1B params	~4.5 GB	Default. Under 16 GB VRAM.
load_in_8bit	~1.1 GB per 1B params	~9 GB	Quality-sensitive, 16-24 GB.
load_in_16bit	~2 GB per 1B params	~16 GB	LoRA at full precision, 24 GB+.
full_finetuning=True	~2 GB weights + ~12 GB optimizer	~112 GB	Rarely justified. Prefer LoRA.

Add roughly 2-6 GB for activations, scaling with max_seq_length and batch size. Treat these as planning figures and confirm against nvidia-smi on the first run — they vary by architecture, attention implementation and vocabulary size.

If the estimate does not fit, reduce in this order: max_seq_length, then batch size (raising gradient_accumulation_steps to hold the effective batch constant), then LoRA rank, then model size. Cutting rank before sequence length usually costs more quality than it saves memory.

Step 2: Load the model

import unsloth must come before transformers, trl or peft. Unsloth patches those libraries at import time; importing them first silently disables the optimizations.

import unsloth  # must be first
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-8B",
    max_seq_length = 2048,
    load_in_4bit = True,
    dtype = None,  # auto-detects bf16 where supported
)

Pick the loader that matches the modality: FastLanguageModel for text-only causal LMs, FastVisionModel for vision-language models, FastModel when the modality is decided at runtime.

The unsloth/ Hub namespace holds pre-quantized copies that download faster and skip a local quantization pass. Upstream repos such as Qwen/ or meta-llama/ work identically.

Step 3: Fix the chat template before training

This is the most common silent failure. A run with the wrong template converges cleanly and produces a model that ignores its stop tokens or emits prompt scaffolding at inference.

from unsloth.chat_templates import (
    get_chat_template,
    standardize_data_formats,
    train_on_responses_only,
)

tokenizer = get_chat_template(tokenizer, chat_template = "qwen3")
dataset = standardize_data_formats(dataset)  # normalizes ShareGPT/OpenAI column names

Then mask the prompt so loss is computed on assistant turns only. Without this, the model is also trained to generate user messages:

trainer = train_on_responses_only(
    trainer,
    instruction_part = "<|im_start|>user\n",
    response_part = "<|im_start|>assistant\n",
)

The two part strings must match the template's actual delimiters. Verify by decoding one batch and confirming the masked region covers exactly the prompt.

Step 4: Attach LoRA adapters

model = FastLanguageModel.get_peft_model(
    model,
    r = 16,
    lora_alpha = 16,
    lora_dropout = 0.0,
    target_modules = [
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    use_gradient_checkpointing = "unsloth",  # Unsloth's variant, lower VRAM than True
    random_state = 3407,
)

Rank guidance: r=8-16 for style and format adaptation, r=32-64 when teaching genuinely new capability. Setting lora_alpha to 1-2x r is a safe default. Keep lora_dropout = 0.0 — Unsloth's fast path is only taken when dropout is zero.

Train all seven projection modules unless VRAM forces otherwise; attention-only LoRA underperforms noticeably on instruction data. For MoE models, expert layers are nn.Parameter rather than nn.Linear and need target_parameters instead of target_modules.

Step 5: Train

Unsloth returns standard PEFT-wrapped models, so TRL's trainers work unmodified.

from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 8,  # effective batch 16
        warmup_steps = 5,
        num_train_epochs = 1,
        learning_rate = 2e-4,
        optim = "adamw_8bit",
        output_dir = "outputs",
    ),
)
trainer.train()

2e-4 suits LoRA; full fine-tuning needs roughly 10x lower. One to three epochs is typical — LoRA overfits small datasets quickly, so watch eval loss rather than trusting an epoch count.

Step 6: Export to the target runtime

The right format depends entirely on where the model will run:

Target	Call	Notes
llama.cpp / Ollama / LM Studio	model.save_pretrained_gguf(dir, tokenizer, quantization_method="q4_k_m")	Builds llama.cpp on first use.
vLLM / TGI / Transformers	model.save_pretrained_merged(dir, tokenizer, save_method="merged_16bit")	Full-size weights.
Adapter only (swapped at runtime)	model.save_pretrained_merged(dir, tokenizer, save_method="lora")	Megabytes, not gigabytes.
Hugging Face Hub	model.push_to_hub_gguf(...) / model.push_to_hub_merged(...)	Needs a write token.

quantization_method accepts a list, so several GGUF quants can be produced in one conversion pass: ["q4_k_m", "q5_k_m", "q8_0"]. q4_k_m is the usual quality/size compromise. The iq* importance-matrix quants additionally require imatrix_file=.

Avoid save_method="merged_4bit" for anything redistributed — it bakes in the quantization and cannot be cleanly re-quantized afterwards.

Examples

Example 1: QLoRA SFT on a 16 GB GPU

import unsloth
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template, train_on_responses_only
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-8B",
    max_seq_length = 2048,
    load_in_4bit = True,
)
model = FastLanguageModel.get_peft_model(model, r = 16, lora_alpha = 16)

tokenizer = get_chat_template(tokenizer, chat_template = "qwen3")
dataset = load_dataset("mlabonne/FineTome-100k", split = "train[:5000]")

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 8,
        num_train_epochs = 1,
        learning_rate = 2e-4,
        optim = "adamw_8bit",
        output_dir = "outputs",
    ),
)
trainer = train_on_responses_only(
    trainer,
    instruction_part = "<|im_start|>user\n",
    response_part = "<|im_start|>assistant\n",
)
trainer.train()

model.save_pretrained_gguf("qwen3-tuned", tokenizer, quantization_method = "q4_k_m")

Example 2: GRPO with vLLM-backed generation

GRPO samples several completions per prompt at every step, so generation dominates step time. Load with fast_inference=True to route sampling through vLLM in the same process.

import unsloth
from unsloth import FastLanguageModel
from trl import GRPOTrainer, GRPOConfig

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-4B",
    max_seq_length = 1024,
    load_in_4bit = True,
    fast_inference = True,       # vLLM sampling backend
    max_lora_rank = 32,          # must be >= the r used below
    gpu_memory_utilization = 0.6,
)
model = FastLanguageModel.get_peft_model(model, r = 32, lora_alpha = 32)

def reward_length(completions, **kwargs):
    """Placeholder. Replace with a task-specific verifier."""
    return [min(len(c) / 200.0, 1.0) for c in completions]

trainer = GRPOTrainer(
    model = model,
    processing_class = tokenizer,
    reward_funcs = [reward_length],
    train_dataset = dataset,
    args = GRPOConfig(
        num_generations = 8,
        max_prompt_length = 256,
        max_completion_length = 512,
        learning_rate = 5e-6,
        output_dir = "grpo-outputs",
    ),
)
trainer.train()

gpu_memory_utilization splits VRAM between vLLM's KV cache and training. Raise it if generation is the bottleneck, lower it if training OOMs. max_lora_rank is fixed at load time and must be at least the r passed later, or adapter loading fails.

GRPO learning rates sit roughly two orders of magnitude below SFT. Reward functions receive completions plus any dataset columns as keyword arguments, and return one float per completion.

Best Practices

• ✅ Set random_state so a promising run can be reproduced.
• ✅ Log peak VRAM on the first run and reuse it to size later jobs on the same hardware.
• ✅ Evaluate the exported artifact, not just the adapter — quantization shifts behaviour.
• ❌ Don't change max_seq_length between training and export; the GGUF inherits it.
• ❌ Don't tune hyperparameters before the loss mask has been verified once.

Limitations

• The VRAM figures above are planning heuristics, not benchmarks. Confirm on target hardware.
• Architecture support changes between releases. Check upstream before assuming a model works.
• Unsloth's speed and memory claims are the project's own published figures, measured on their own benchmarks; they are not independently verified here.
• This skill does not replace environment-specific validation, testing, or expert review.
• Stop and ask for clarification if the GPU, model, dataset format or export target is unknown — every step above depends on those four.

Security & Safety Notes

• Training commands are long-running and hold the GPU exclusively. Confirm before launching on a shared or remote machine.
• push_to_hub_gguf and push_to_hub_merged publish weights to a public Hub repo by default. Confirm intent and pass private=True when the model is not meant to be public.
• Read Hugging Face tokens from the environment (HF_TOKEN), never inline in a script. A committed token grants write access to every model the account owns.
• Fine-tuning reproduces the training data's content and biases in the weights. Confirm the dataset is licensed for training and free of secrets before starting.
• GGUF export builds llama.cpp from source on first use, compiling third-party code and requiring network access. Expect it to be slow and to need a working toolchain.
• Unsloth is dual-licensed: the core package is Apache-2.0, while optional components such as the Studio UI are AGPL-3.0. Check a component's license before redistributing it.

Common Pitfalls

• Problem: Trained model ignores stop tokens or echoes the prompt format. Solution: Wrong chat template, or train_on_responses_only was never applied. Verify the mask on a decoded batch before blaming hyperparameters.

• Problem: CUDA OOM partway through the first epoch rather than at step 0. Solution: A long sample exceeded the activation budget. Lower max_seq_length or filter outliers — peak memory tracks the longest sequence, not the mean.

• Problem: Training runs, but at ordinary unaccelerated speed. Solution: transformers or trl was imported before unsloth, so the patches never applied. Move import unsloth to the top of the file.

• Problem: save_pretrained_gguf appears to hang on first call. Solution: It is building llama.cpp. Ensure a compiler and network access are available, or export merged_16bit and convert separately.

• Problem: GRPO fails with a LoRA rank mismatch. Solution: max_lora_rank at from_pretrained is below the r given to get_peft_model. Raise it to match.

• Problem: Loss collapses to near zero within a few hundred steps. Solution: Overfitting a small dataset, or the loss mask is leaking the answer into the prompt. Check dataset size against epoch count, then re-verify masking.

Related Skills

• @trl-training - Use for the TRL CLI, multi-GPU runs, or architectures Unsloth lacks.
• @hugging-face-model-trainer - Use for managed training on Hugging Face Jobs instead of local hardware.
• @local-llm-expert - Use to serve the exported GGUF via Ollama, llama.cpp or vLLM.

Additional Resources

• Unsloth documentation
• Reinforcement learning guide (GRPO, DPO)
• Saving to GGUF
• unslothai/unsloth on GitHub