From 27eb7f0941e000fcff4ee0f0d693e724951acb82 Mon Sep 17 00:00:00 2001
From: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Date: Fri, 22 Dec 2023 13:45:26 +0100
Subject: [PATCH] =?UTF-8?q?[`Docs`]=C2=A0Add=20unsloth=20optimizations=20i?=
 =?UTF-8?q?n=20TRL's=20documentation=20(#1119)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* add unsloth

* Update sft_trainer.mdx (#1124)

Co-authored-by: Daniel Han <danielhanchen@gmail.com>

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 docs/source/sft_trainer.mdx | 55 +++++++++++++++++++++++++++++++++++++
 1 file changed, 55 insertions(+)

diff --git a/docs/source/sft_trainer.mdx b/docs/source/sft_trainer.mdx
index a6ca3e439c..4c0c1abeac 100644
--- a/docs/source/sft_trainer.mdx
+++ b/docs/source/sft_trainer.mdx
@@ -410,6 +410,61 @@ We have tested NEFTune by training `mistralai/Mistral-7B-v0.1` on the [OpenAssis
 </div>
 
 Note however, that the amount of performance gain is _dataset dependent_ and in particular, applying NEFTune on synthetic datasets like [UltraChat](https://huggingface.co/datasets/stingning/ultrachat) typically produces smaller gains.
+
+### Accelerate fine-tuning 2x using `unsloth`
+
+You can further accelerate QLoRA / LoRA (2x faster, 60% less memory) and even full-finetuning (1.1x faster) using the [`unsloth`](https://github.com/unslothai/unsloth) library that is compatible with `SFTTrainer`. Currently `unsloth` supports only Llama (Yi, TinyLlama as well) and Mistral architectures.
+First install `unsloth` according to the [official documentation](https://github.com/unslothai/unsloth#installation-instructions---conda). Once installed, you can incorporate unsloth into your workflow in a very simple manner; instead of loading `AutoModelForCausalLM`, you just need to load a `FastLlamaModel` or `FastMistralModel` as follows:
+
+```python
+import torch
+
+from transformers import TrainingArguments
+from trl import SFTTrainer
+from unsloth import FastLlamaModel, FastMistralModel
+
+max_seq_length = 2048 # Supports automatic RoPE Scaling, so choose any number.
+dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
+load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
+
+# Load Llama model
+model, tokenizer = FastLlamaModel.from_pretrained(
+    model_name = "unsloth/llama-2-7b", # Supports any llama model eg meta-llama/Llama-2-7b-hf
+    max_seq_length = max_seq_length,
+    dtype = dtype,
+    load_in_4bit = load_in_4bit,
+    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
+)
+
+# Do model patching and add fast LoRA weights
+model = FastLlamaModel.get_peft_model(
+    model,
+    r = 16,
+    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
+                      "gate_proj", "up_proj", "down_proj",],
+    lora_alpha = 16,
+    lora_dropout = 0, # Currently only supports dropout = 0
+    bias = "none",    # Currently only supports bias = "none"
+    use_gradient_checkpointing = True,
+    random_state = 3407,
+    max_seq_length = max_seq_length,
+)
+
+args = TrainingArguments(output_dir="./output")
+
+trainer = SFTTrainer(
+    model = model,
+    args = args,
+    train_dataset = dataset,
+    dataset_text_field = "text",
+    max_seq_length = max_seq_length,
+)
+
+trainer.train()
+```
+
+The saved model is fully compatible with Hugging Face's transformers library. Learn more about unsloth in their [official repository](https://github.com/unslothai/unsloth).
+
 ## Best practices
 
 Pay attention to the following best practices when training a model with that trainer: