danielhanchen commited on
Commit
cd3117a
•
1 Parent(s): c389abe

Upload 8 files

Browse files
Alpaca_+_Codellama_34b_full_example.ipynb CHANGED
@@ -3,13 +3,11 @@
3
  {
4
  "cell_type": "markdown",
5
  "source": [
6
- "To run this, press \"Runtime\" and press \"Run all\" on a A100 Colab instance!\n",
7
- "\n",
8
- "**[NOTE]** You might be lucky if an A100 is free!. If not, try our Mistral 7b notebook on a free Tesla T4 [here](https://colab.research.google.com/drive/1Dyauq4kTZoLewQ1cApceUQVNcnnNTzg_?usp=sharing).\n",
9
  "<div class=\"align-center\">\n",
10
- " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"110\"></a>\n",
11
- " <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"150\"></a>\n",
12
- " <a href=\"https://huggingface.co/docs/trl/main/en/index\"><img src=\"https://github.com/huggingface/blog/blob/main/assets/133_trl_peft/thumbnail.png?raw=true\" width=\"100\"></a> Join our Discord if you need help!\n",
13
  "</div>\n",
14
  "\n",
15
  "To install Unsloth on your own computer, follow the installation instructions on our Github page [here](https://github.com/unslothai/unsloth#installation-instructions---conda).\n",
@@ -31,12 +29,14 @@
31
  "%%capture\n",
32
  "import torch\n",
33
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
34
  "if major_version >= 8:\n",
35
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
36
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
37
  "else:\n",
38
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
39
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
40
  "pass"
41
  ]
42
  },
@@ -47,7 +47,8 @@
47
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
48
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
49
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
50
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models."
 
51
  ],
52
  "metadata": {
53
  "id": "r2v_X2fA0Df5"
@@ -1502,17 +1503,20 @@
1502
  "source": [
1503
  "And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/u54VK8m8tk) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
1504
  "\n",
1505
- "We also have other notebooks on:\n",
1506
- "1. Zephyr DPO [free Colab](https://colab.research.google.com/drive/15vttTpzzVXv_tJwEk-hIcQ0S9FcEWvwP?usp=sharing)\n",
1507
- "2. Llama 7b [free Colab](https://colab.research.google.com/drive/1lBzz5KeZJKXjvivbYvmGarix9Ao6Wxe5?usp=sharing)\n",
1508
- "3. TinyLlama full Alpaca 52K in under 80 hours [free Colab](https://colab.research.google.com/drive/1AZghoNBQaMDgWJpi4RbffGM1h6raLUj9?usp=sharing)\n",
1509
- "4. Mistral 7b [free Colab](https://colab.research.google.com/drive/1Dyauq4kTZoLewQ1cApceUQVNcnnNTzg_?usp=sharing)\n",
1510
- "5. Llama 7b [free Kaggle](https://www.kaggle.com/danielhanchen/unsloth-alpaca-t4-ddp)\n",
 
 
 
1511
  "\n",
1512
  "<div class=\"align-center\">\n",
1513
- " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"110\"></a>\n",
1514
- " <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"150\"></a>\n",
1515
- " <a href=\"https://huggingface.co/docs/trl/main/en/index\"><img src=\"https://github.com/huggingface/blog/blob/main/assets/133_trl_peft/thumbnail.png?raw=true\" width=\"100\"></a>\n",
1516
  "</div>"
1517
  ],
1518
  "metadata": {
 
3
  {
4
  "cell_type": "markdown",
5
  "source": [
6
+ "To run this, press \"*Runtime*\" and press \"*Run all*\" on a **free** Tesla T4 Google Colab instance!\n",
 
 
7
  "<div class=\"align-center\">\n",
8
+ " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
9
+ " <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord button.png\" width=\"145\"></a>\n",
10
+ " <a href=\"https://ko-fi.com/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Kofi button.png\" width=\"145\"></a></a> Join Discord if you need help + support us if you can!\n",
11
  "</div>\n",
12
  "\n",
13
  "To install Unsloth on your own computer, follow the installation instructions on our Github page [here](https://github.com/unslothai/unsloth#installation-instructions---conda).\n",
 
29
  "%%capture\n",
30
  "import torch\n",
31
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
32
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
33
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
34
  "if major_version >= 8:\n",
35
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
36
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
37
  "else:\n",
38
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
39
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
40
  "pass"
41
  ]
42
  },
 
47
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
48
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
49
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
50
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
51
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
52
  ],
53
  "metadata": {
54
  "id": "r2v_X2fA0Df5"
 
1503
  "source": [
1504
  "And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/u54VK8m8tk) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
1505
  "\n",
1506
+ "Some other links:\n",
1507
+ "1. Zephyr DPO 2x faster [free Colab](https://colab.research.google.com/drive/15vttTpzzVXv_tJwEk-hIcQ0S9FcEWvwP?usp=sharing)\n",
1508
+ "2. Llama 7b 2x faster [free Colab](https://colab.research.google.com/drive/1lBzz5KeZJKXjvivbYvmGarix9Ao6Wxe5?usp=sharing)\n",
1509
+ "3. TinyLlama 4x faster full Alpaca 52K in 1 hour [free Colab](https://colab.research.google.com/drive/1AZghoNBQaMDgWJpi4RbffGM1h6raLUj9?usp=sharing)\n",
1510
+ "4. CodeLlama 34b 2x faster [A100 on Colab](https://colab.research.google.com/drive/1y7A0AxE3y8gdj4AVkl2aZX47Xu3P1wJT?usp=sharing)\n",
1511
+ "5. Mistral 7b [free Kaggle version](https://www.kaggle.com/code/danielhanchen/kaggle-mistral-7b-unsloth-notebook)\n",
1512
+ "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1513
+ "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
1514
+ "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
1515
  "\n",
1516
  "<div class=\"align-center\">\n",
1517
+ " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
1518
+ " <a href=\"https://discord.gg/u54VK8m8tk\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"145\"></a>\n",
1519
+ " <a href=\"https://ko-fi.com/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Kofi button.png\" width=\"145\"></a></a> Support our work if you can! Thanks!\n",
1520
  "</div>"
1521
  ],
1522
  "metadata": {
Alpaca_+_Gemma_7b_full_example.ipynb ADDED
The diff for this file is too large to render. See raw diff
 
Alpaca_+_Llama_7b_full_example.ipynb CHANGED
@@ -29,12 +29,14 @@
29
  "%%capture\n",
30
  "import torch\n",
31
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
32
  "if major_version >= 8:\n",
33
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
34
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
35
  "else:\n",
36
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
37
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
38
  "pass"
39
  ]
40
  },
@@ -45,7 +47,8 @@
45
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
46
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
47
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
48
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models."
 
49
  ],
50
  "metadata": {
51
  "id": "r2v_X2fA0Df5"
@@ -283,6 +286,8 @@
283
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
284
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
285
  " \"unsloth/tinyllama-bnb-4bit\",\n",
 
 
286
  "] # More models at https://huggingface.co/unsloth\n",
287
  "\n",
288
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
@@ -1348,6 +1353,7 @@
1348
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1349
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
1350
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
 
1351
  "\n",
1352
  "<div class=\"align-center\">\n",
1353
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
 
29
  "%%capture\n",
30
  "import torch\n",
31
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
32
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
33
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
34
  "if major_version >= 8:\n",
35
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
36
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
37
  "else:\n",
38
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
39
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
40
  "pass"
41
  ]
42
  },
 
47
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
48
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
49
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
50
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
51
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
52
  ],
53
  "metadata": {
54
  "id": "r2v_X2fA0Df5"
 
286
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
287
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
288
  " \"unsloth/tinyllama-bnb-4bit\",\n",
289
+ " \"unsloth/gemma-7b-bnb-4bit\", # New Google 6 trillion tokens model 2.5x faster!\n",
290
+ " \"unsloth/gemma-2b-bnb-4bit\",\n",
291
  "] # More models at https://huggingface.co/unsloth\n",
292
  "\n",
293
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
 
1353
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1354
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
1355
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
1356
+ "9. Gemma 6 trillion tokens is 2.5x faster! [free Colab](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)\n",
1357
  "\n",
1358
  "<div class=\"align-center\">\n",
1359
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
Alpaca_+_Mistral_7b_full_example.ipynb CHANGED
@@ -29,12 +29,14 @@
29
  "%%capture\n",
30
  "import torch\n",
31
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
32
  "if major_version >= 8:\n",
33
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
34
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
35
  "else:\n",
36
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
37
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
38
  "pass"
39
  ]
40
  },
@@ -45,7 +47,8 @@
45
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
46
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
47
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
48
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models."
 
49
  ],
50
  "metadata": {
51
  "id": "r2v_X2fA0Df5"
@@ -282,6 +285,8 @@
282
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
283
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
284
  " \"unsloth/tinyllama-bnb-4bit\",\n",
 
 
285
  "] # More models at https://huggingface.co/unsloth\n",
286
  "\n",
287
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
@@ -1260,6 +1265,7 @@
1260
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1261
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
1262
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
 
1263
  "\n",
1264
  "<div class=\"align-center\">\n",
1265
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
 
29
  "%%capture\n",
30
  "import torch\n",
31
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
32
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
33
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
34
  "if major_version >= 8:\n",
35
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
36
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
37
  "else:\n",
38
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
39
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
40
  "pass"
41
  ]
42
  },
 
47
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
48
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
49
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
50
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
51
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
52
  ],
53
  "metadata": {
54
  "id": "r2v_X2fA0Df5"
 
285
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
286
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
287
  " \"unsloth/tinyllama-bnb-4bit\",\n",
288
+ " \"unsloth/gemma-7b-bnb-4bit\", # New Google 6 trillion tokens model 2.5x faster!\n",
289
+ " \"unsloth/gemma-2b-bnb-4bit\",\n",
290
  "] # More models at https://huggingface.co/unsloth\n",
291
  "\n",
292
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
 
1265
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1266
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
1267
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
1268
+ "9. Gemma 6 trillion tokens is 2.5x faster! [free Colab](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)\n",
1269
  "\n",
1270
  "<div class=\"align-center\">\n",
1271
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
Alpaca_+_TinyLlama_+_RoPE_Scaling_full_example.ipynb CHANGED
@@ -31,12 +31,14 @@
31
  "%%capture\n",
32
  "import torch\n",
33
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
34
  "if major_version >= 8:\n",
35
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
36
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
37
  "else:\n",
38
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
39
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
40
  "pass"
41
  ]
42
  },
@@ -50,7 +52,8 @@
50
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
51
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
52
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
53
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models."
 
54
  ]
55
  },
56
  {
@@ -282,6 +285,8 @@
282
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
283
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
284
  " \"unsloth/tinyllama-bnb-4bit\",\n",
 
 
285
  "] # More models at https://huggingface.co/unsloth\n",
286
  "\n",
287
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
@@ -2528,6 +2533,7 @@
2528
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
2529
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
2530
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
 
2531
  "\n",
2532
  "<div class=\"align-center\">\n",
2533
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
 
31
  "%%capture\n",
32
  "import torch\n",
33
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
34
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
35
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
36
  "if major_version >= 8:\n",
37
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
38
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
39
  "else:\n",
40
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
41
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
42
  "pass"
43
  ]
44
  },
 
52
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
53
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
54
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
55
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
56
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
57
  ]
58
  },
59
  {
 
285
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
286
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
287
  " \"unsloth/tinyllama-bnb-4bit\",\n",
288
+ " \"unsloth/gemma-7b-bnb-4bit\", # New Google 6 trillion tokens model 2.5x faster!\n",
289
+ " \"unsloth/gemma-2b-bnb-4bit\",\n",
290
  "] # More models at https://huggingface.co/unsloth\n",
291
  "\n",
292
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
 
2533
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
2534
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
2535
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
2536
+ "9. Gemma 6 trillion tokens is 2.5x faster! [free Colab](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)\n",
2537
  "\n",
2538
  "<div class=\"align-center\">\n",
2539
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
ChatML_+_chat_templates_+_Mistral_7b_full_example.ipynb CHANGED
@@ -31,12 +31,14 @@
31
  "%%capture\n",
32
  "import torch\n",
33
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
34
  "if major_version >= 8:\n",
35
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
36
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
37
  "else:\n",
38
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
39
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
40
  "pass"
41
  ]
42
  },
@@ -47,7 +49,8 @@
47
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
48
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
49
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
50
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models."
 
51
  ],
52
  "metadata": {
53
  "id": "r2v_X2fA0Df5"
@@ -284,6 +287,8 @@
284
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
285
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
286
  " \"unsloth/tinyllama-bnb-4bit\",\n",
 
 
287
  "] # More models at https://huggingface.co/unsloth\n",
288
  "\n",
289
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
@@ -1392,6 +1397,7 @@
1392
  "5. Mistral 7b [free Kaggle version](https://www.kaggle.com/code/danielhanchen/kaggle-mistral-7b-unsloth-notebook)\n",
1393
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1394
  "7. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
 
1395
  "\n",
1396
  "<div class=\"align-center\">\n",
1397
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
 
31
  "%%capture\n",
32
  "import torch\n",
33
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
34
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
35
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
36
  "if major_version >= 8:\n",
37
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
38
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
39
  "else:\n",
40
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
41
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
42
  "pass"
43
  ]
44
  },
 
49
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
50
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
51
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
52
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
53
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
54
  ],
55
  "metadata": {
56
  "id": "r2v_X2fA0Df5"
 
287
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
288
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
289
  " \"unsloth/tinyllama-bnb-4bit\",\n",
290
+ " \"unsloth/gemma-7b-bnb-4bit\", # New Google 6 trillion tokens model 2.5x faster!\n",
291
+ " \"unsloth/gemma-2b-bnb-4bit\",\n",
292
  "] # More models at https://huggingface.co/unsloth\n",
293
  "\n",
294
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
 
1397
  "5. Mistral 7b [free Kaggle version](https://www.kaggle.com/code/danielhanchen/kaggle-mistral-7b-unsloth-notebook)\n",
1398
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1399
  "7. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
1400
+ "9. Gemma 6 trillion tokens is 2.5x faster! [free Colab](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)\n",
1401
  "\n",
1402
  "<div class=\"align-center\">\n",
1403
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
DPO_Zephyr_Unsloth_Example.ipynb CHANGED
@@ -30,12 +30,14 @@
30
  "%%capture\n",
31
  "import torch\n",
32
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
33
  "if major_version >= 8:\n",
34
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
35
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
36
  "else:\n",
37
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
38
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
39
  "pass"
40
  ]
41
  },
@@ -49,8 +51,9 @@
49
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
50
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
51
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
52
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
53
- "* DPO requires a model already trained by SFT on a similar dataset that is used for DPO. We use `HuggingFaceH4/mistral-7b-sft-beta` as the SFT model. Use this [notebook](https://colab.research.google.com/drive/1Dyauq4kTZoLewQ1cApceUQVNcnnNTzg_?usp=sharing) first to train a SFT model."
 
54
  ]
55
  },
56
  {
@@ -2564,6 +2567,7 @@
2564
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
2565
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
2566
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
 
2567
  "\n",
2568
  "<div class=\"align-center\">\n",
2569
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
 
30
  "%%capture\n",
31
  "import torch\n",
32
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
33
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
34
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
35
  "if major_version >= 8:\n",
36
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
37
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
38
  "else:\n",
39
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
40
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
41
  "pass"
42
  ]
43
  },
 
51
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
52
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
53
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
54
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
55
+ "* DPO requires a model already trained by SFT on a similar dataset that is used for DPO. We use `HuggingFaceH4/mistral-7b-sft-beta` as the SFT model. Use this [notebook](https://colab.research.google.com/drive/1Dyauq4kTZoLewQ1cApceUQVNcnnNTzg_?usp=sharing) first to train a SFT model.\n",
56
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
57
  ]
58
  },
59
  {
 
2567
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
2568
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
2569
  "8. Text completions like novel writing [notebook](https://colab.research.google.com/drive/1ef-tab5bhkvWmBOObepl1WgJvfvSzn5Q?usp=sharing)\n",
2570
+ "9. Gemma 6 trillion tokens is 2.5x faster! [free Colab](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)\n",
2571
  "\n",
2572
  "<div class=\"align-center\">\n",
2573
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
Mistral_7b_Text_Completion_Raw_Text_training_full_example.ipynb CHANGED
@@ -41,12 +41,14 @@
41
  "%%capture\n",
42
  "import torch\n",
43
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
 
 
44
  "if major_version >= 8:\n",
45
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
46
- " !pip install \"unsloth[colab-ampere] @ git+https://github.com/unslothai/unsloth.git\"\n",
47
  "else:\n",
48
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
49
- " !pip install \"unsloth[colab] @ git+https://github.com/unslothai/unsloth.git\"\n",
50
  "pass"
51
  ]
52
  },
@@ -60,7 +62,8 @@
60
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
61
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
62
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
63
- "* [**NEW**] With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models."
 
64
  ]
65
  },
66
  {
@@ -296,6 +299,8 @@
296
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
297
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
298
  " \"unsloth/tinyllama-bnb-4bit\",\n",
 
 
299
  "] # More models at https://huggingface.co/unsloth\n",
300
  "\n",
301
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
@@ -1324,6 +1329,7 @@
1324
  "5. Mistral 7b [free Kaggle version](https://www.kaggle.com/code/danielhanchen/kaggle-mistral-7b-unsloth-notebook)\n",
1325
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1326
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
 
1327
  "\n",
1328
  "<div class=\"align-center\">\n",
1329
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
 
41
  "%%capture\n",
42
  "import torch\n",
43
  "major_version, minor_version = torch.cuda.get_device_capability()\n",
44
+ "# Must install separately since Colab has torch 2.2.1, which breaks packages\n",
45
+ "!pip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\n",
46
  "if major_version >= 8:\n",
47
  " # Use this for new GPUs like Ampere, Hopper GPUs (RTX 30xx, RTX 40xx, A100, H100, L40)\n",
48
+ " !pip install --no-deps packaging ninja einops flash-attn xformers trl peft accelerate bitsandbytes\n",
49
  "else:\n",
50
  " # Use this for older GPUs (V100, Tesla T4, RTX 20xx)\n",
51
+ " !pip install --no-deps xformers trl peft accelerate bitsandbytes\n",
52
  "pass"
53
  ]
54
  },
 
62
  "* And Yi, Qwen ([llamafied](https://huggingface.co/models?sort=trending&search=qwen+llama)), Deepseek, all Llama, Mistral derived archs.\n",
63
  "* We support 16bit LoRA or 4bit QLoRA. Both 2x faster.\n",
64
  "* `max_seq_length` can be set to anything, since we do automatic RoPE Scaling via [kaiokendev's](https://kaiokendev.github.io/til) method.\n",
65
+ "* With [PR 26037](https://github.com/huggingface/transformers/pull/26037), we support downloading 4bit models **4x faster**! [Our repo](https://huggingface.co/unsloth) has Llama, Mistral 4bit models.\n",
66
+ "* [**NEW**] We make Gemma 6 trillion tokens **2.5x faster**! See our [Gemma notebook](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)"
67
  ]
68
  },
69
  {
 
299
  " \"unsloth/llama-2-13b-bnb-4bit\",\n",
300
  " \"unsloth/codellama-34b-bnb-4bit\",\n",
301
  " \"unsloth/tinyllama-bnb-4bit\",\n",
302
+ " \"unsloth/gemma-7b-bnb-4bit\", # New Google 6 trillion tokens model 2.5x faster!\n",
303
+ " \"unsloth/gemma-2b-bnb-4bit\",\n",
304
  "] # More models at https://huggingface.co/unsloth\n",
305
  "\n",
306
  "model, tokenizer = FastLanguageModel.from_pretrained(\n",
 
1329
  "5. Mistral 7b [free Kaggle version](https://www.kaggle.com/code/danielhanchen/kaggle-mistral-7b-unsloth-notebook)\n",
1330
  "6. We also did a [blog](https://huggingface.co/blog/unsloth-trl) with 🤗 HuggingFace, and we're in the TRL [docs](https://huggingface.co/docs/trl/main/en/sft_trainer#accelerate-fine-tuning-2x-using-unsloth)!\n",
1331
  "7. `ChatML` for ShareGPT datasets, [conversational notebook](https://colab.research.google.com/drive/1Aau3lgPzeZKQ-98h69CCu1UJcvIBLmy2?usp=sharing)\n",
1332
+ "8. Gemma 6 trillion tokens is 2.5x faster! [free Colab](https://colab.research.google.com/drive/10NbwlsRChbma1v55m8LAPYG15uQv6HLo?usp=sharing)\n",
1333
  "\n",
1334
  "<div class=\"align-center\">\n",
1335
  " <a href=\"https://github.com/unslothai/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",