SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[docs] Redesign (#31757)

Browse Source 
* toctree

* not-doctested.txt

* collapse sections

* feedback

* update

* rewrite get started sections

* fixes

* fix

* loading models

* fix

* customize models

* share

* fix link

* contribute part 1

* contribute pt 2

* fix toctree

* tokenization pt 1

* Add new model (#32615)

* v1 - working version

* fix

* fix

* fix

* fix

* rename to correct name

* fix title

* fixup

* rename files

* fix

* add copied from on tests

* rename to `FalconMamba` everywhere and fix bugs

* fix quantization + accelerate

* fix copies

* add `torch.compile` support

* fix tests

* fix tests and add slow tests

* copies on config

* merge the latest changes

* fix tests

* add few lines about instruct

* Apply suggestions from code review

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* fix

* fix tests

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* "to be not" -> "not to be" (#32636)

* "to be not" -> "not to be"

* Update sam.md

* Update trainer.py

* Update modeling_utils.py

* Update test_modeling_utils.py

* Update test_modeling_utils.py

* fix hfoption tag

* tokenization pt. 2

* image processor

* fix toctree

* backbones

* feature extractor

* fix file name

* processor

* update not-doctested

* update

* make style

* fix toctree

* revision

* make fixup

* fix toctree

* fix

* make style

* fix hfoption tag

* pipeline

* pipeline gradio

* pipeline web server

* add pipeline

* fix toctree

* not-doctested

* prompting

* llm optims

* fix toctree

* fixes

* cache

* text generation

* fix

* chat pipeline

* chat stuff

* xla

* torch.compile

* cpu inference

* toctree

* gpu inference

* agents and tools

* gguf/tiktoken

* finetune

* toctree

* trainer

* trainer pt 2

* optims

* optimizers

* accelerate

* parallelism

* fsdp

* update

* distributed cpu

* hardware training

* gpu training

* gpu training 2

* peft

* distrib debug

* deepspeed 1

* deepspeed 2

* chat toctree

* quant pt 1

* quant pt 2

* fix toctree

* fix

* fix

* quant pt 3

* quant pt 4

* serialization

* torchscript

* scripts

* tpu

* review

* model addition timeline

* modular

* more reviews

* reviews

* fix toctree

* reviews reviews

* continue reviews

* more reviews

* modular transformers

* more review

* zamba2

* fix

* all frameworks

* pytorch

* supported model frameworks

* flashattention

* rm check_table

* not-doctested.txt

* rm check_support_list.py

* feedback

* updates/feedback

* review

* feedback

* fix

* update

* feedback

* updates

* update

---------

Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com>
This commit is contained in:
Steven Liu
2025-03-03 10:33:46 -08:00
committed by GitHub
parent 6aa9888463
commit c0f8d055ce
423 changed files with 10925 additions and 14569 deletions
Download Patch File Download Diff File Expand all files Collapse all files

405
docs/source/en/llm_tutorial.md
View File

@@ -1,4 +1,4 @@
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
@@ -14,269 +14,276 @@ rendered properly in your Markdown viewer.
-->
# Generation with LLMs
# Text generation
[[open-in-colab]]
LLMs, or Large Language Models, are the key component behind text generation. In a nutshell, they consist of large pretrained transformer models trained to predict the next word (or, more precisely, token) given some input text. Since they predict one token at a time, you need to do something more elaborate to generate new sentences other than just calling the model -- you need to do autoregressive generation.
Text generation is the most popular application for large language models (LLMs). A LLM is trained to generate the next word (token) given some initial text (prompt) along with its own generated outputs up to a predefined length or when it reaches an end-of-sequence (`EOS`) token.
Autoregressive generation is the inference-time procedure of iteratively calling a model with its own generated outputs, given a few initial inputs. In 🤗 Transformers, this is handled by the [`~generation.GenerationMixin.generate`] method, which is available to all models with generative capabilities.
In Transformers, the [`~GenerationMixin.generate`] API handles text generation, and it is available for all models with generative capabilities.
<Tip>
This guide will show you the basics of text generation with [`~GenerationMixin.generate`] and some common pitfalls to avoid.
If you want to jump straight to chatting with a model, [try our chat CLI](quicktour#chat-with-text-generation-models).
## Default generate
</Tip>
This tutorial will show you how to:
* Generate text with an LLM
* Avoid common pitfalls
* Next steps to help you get the most out of your LLM
Before you begin, make sure you have all the necessary libraries installed:
Before you begin, it's helpful to install [bitsandbytes](https://hf.co/docs/bitsandbytes/index) to quantize really large models to reduce their memory usage.
```bash
pip install transformers bitsandbytes>=0.39.0 -q
!pip install -U transformers bitsandbytes
```
Bitsandbytes supports multiple backends in addition to CUDA-based GPUs. Refer to the multi-backend installation [guide](https://huggingface.co/docs/bitsandbytes/main/en/installation#multi-backend) to learn more.
Load a LLM with [`~PreTrainedModel.from_pretrained`] and add the following two parameters to reduce the memory requirements.
## Generate text
A language model trained for [causal language modeling](tasks/language_modeling) takes a sequence of text tokens as input and returns the probability distribution for the next token.
<!-- [GIF 1 -- FWD PASS] -->
<figure class="image table text-center m-0 w-full">
<video
style="max-width: 90%; margin: auto;"
autoplay loop muted playsinline
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_1_1080p.mov"
></video>
<figcaption>"Forward pass of an LLM"</figcaption>
</figure>
A critical aspect of autoregressive generation with LLMs is how to select the next token from this probability distribution. Anything goes in this step as long as you end up with a token for the next iteration. This means it can be as simple as selecting the most likely token from the probability distribution or as complex as applying a dozen transformations before sampling from the resulting distribution.
<!-- [GIF 2 -- TEXT GENERATION] -->
<figure class="image table text-center m-0 w-full">
<video
style="max-width: 90%; margin: auto;"
autoplay loop muted playsinline
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_2_1080p.mov"
></video>
<figcaption>"Autoregressive generation iteratively selects the next token from a probability distribution to generate text"</figcaption>
</figure>
The process depicted above is repeated iteratively until some stopping condition is reached. Ideally, the stopping condition is dictated by the model, which should learn when to output an end-of-sequence (`EOS`) token. If this is not the case, generation stops when some predefined maximum length is reached.
Properly setting up the token selection step and the stopping condition is essential to make your model behave as you'd expect on your task. That is why we have a [`~generation.GenerationConfig`] file associated with each model, which contains a good default generative parameterization and is loaded alongside your model.
Let's talk code!
<Tip>
If you're interested in basic LLM usage, our high-level [`Pipeline`](pipeline_tutorial) interface is a great starting point. However, LLMs often require advanced features like quantization and fine control of the token selection step, which is best done through [`~generation.GenerationMixin.generate`]. Autoregressive generation with LLMs is also resource-intensive and should be executed on a GPU for adequate throughput.
</Tip>
First, you need to load the model.
- `device_map="auto"` enables Accelerates' [Big Model Inference](./models#big-model-inference) feature for automatically initiating the model skeleton and loading and dispatching the model weights across all available devices, starting with the fastest device (GPU).
- `quantization_config` is a configuration object that defines the quantization settings. This examples uses bitsandbytes as the quantization backend (see the [Quantization](./quantization/overview) section for more available backends) and it loads the model in [4-bits](./quantization/bitsandbytes).
```py
>>> from transformers import AutoModelForCausalLM
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
>>> model = AutoModelForCausalLM.from_pretrained(
... "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
... )
quantization_config = BitsAndBytesConfig(load_in_4bit=True)
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", device_map="auto", quantization_config=quantization_config)
```
You'll notice two flags in the `from_pretrained` call:
Tokenize your input, and set the [`~PreTrainedTokenizer.padding_side`] parameter to `"left"` because a LLM is not trained to continue generation from padding tokens. The tokenizer returns the input ids and attention mask.
- `device_map` ensures the model is moved to your GPU(s)
- `load_in_4bit` applies [4-bit dynamic quantization](main_classes/quantization) to massively reduce the resource requirements
There are other ways to initialize a model, but this is a good baseline to begin with an LLM.
Next, you need to preprocess your text input with a [tokenizer](tokenizer_summary).
> [!TIP]
> Process more than one prompt at a time by passing a list of strings to the tokenizer. Batch the inputs to improve throughput at a small cost to latency and memory.
```py
>>> from transformers import AutoTokenizer
>>> from accelerate.test_utils.testing import get_backend
>>> DEVICE, _, _ = get_backend() # automatically detects the underlying device type (CUDA, CPU, XPU, MPS, etc.)
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", padding_side="left")
>>> model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to(DEVICE)
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", padding_side="left")
model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to("cuda")
```
The `model_inputs` variable holds the tokenized text input, as well as the attention mask. While [`~generation.GenerationMixin.generate`] does its best effort to infer the attention mask when it is not passed, we recommend passing it whenever possible for optimal results.
After tokenizing the inputs, you can call the [`~generation.GenerationMixin.generate`] method to returns the generated tokens. The generated tokens then should be converted to text before printing.
Pass the inputs to [`~GenerationMixin.generate`] to generate tokens, and [`~PreTrainedTokenizer.batch_decode`] the generated tokens back to text.
```py
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'A list of colors: red, blue, green, yellow, orange, purple, pink,'
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
"A list of colors: red, blue, green, yellow, orange, purple, pink,"
```
Finally, you don't need to do it one sequence at a time! You can batch your inputs, which will greatly improve the throughput at a small latency and memory cost. All you need to do is to make sure you pad your inputs properly (more on that below).
## Generation configuration
All generation settings are contained in [`GenerationConfig`]. In the example above, the generation settings are derived from the `generation_config.json` file of [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1). A default decoding strategy is used when no configuration is saved with a model.
Inspect the configuration through the `generation_config` attribute. It only shows values that are different from the default configuration, in this case, the `bos_token_id` and `eos_token_id`.
```py
>>> tokenizer.pad_token = tokenizer.eos_token # Most LLMs don't have a pad token by default
>>> model_inputs = tokenizer(
... ["A list of colors: red, blue", "Portugal is"], return_tensors="pt", padding=True
... ).to(DEVICE)
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
['A list of colors: red, blue, green, yellow, orange, purple, pink,',
'Portugal is a country in southwestern Europe, on the Iber']
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1", device_map="auto")
model.generation_config
GenerationConfig {
"bos_token_id": 1,
"eos_token_id": 2
}
```
And that's it! In a few lines of code, you can harness the power of an LLM.
## Common pitfalls
There are many [generation strategies](generation_strategies), and sometimes the default values may not be appropriate for your use case. If your outputs aren't aligned with what you're expecting, we've created a list of the most common pitfalls and how to avoid them.
You can customize [`~GenerationMixin.generate`] by overriding the parameters and values in [`GenerationConfig`]. Some of the most commonly adjusted parameters are [max_new_tokens](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig.max_new_tokens), [num_beams](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig.num_beams), [do_sample](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig.do_sample), and [num_return_sequences](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig.num_return_sequences).
```py
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
>>> tokenizer.pad_token = tokenizer.eos_token # Most LLMs don't have a pad token by default
>>> model = AutoModelForCausalLM.from_pretrained(
... "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
... )
# enable beam search sampling strategy
model.generate(**inputs, num_beams=4, do_sample=True)
```
### Generated output is too short/long
[`~GenerationMixin.generate`] can also be extended with external libraries or custom code. The `logits_processor` parameter accepts custom [`LogitsProcessor`] instances for manupulating the next token probability distribution. `stopping_criteria` supports custom [`StoppingCriteria`] to stop text generation. Check out the [logits-processor-zoo](https://github.com/NVIDIA/logits-processor-zoo) for more examples of external [`~GenerationMixin.generate`]-compatible extensions.
If not specified in the [`~generation.GenerationConfig`] file, `generate` returns up to 20 tokens by default. We highly recommend manually setting `max_new_tokens` in your `generate` call to control the maximum number of new tokens it can return. Keep in mind LLMs (more precisely, [decoder-only models](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt)) also return the input prompt as part of the output.
Refer to the [Generation strategies](./generation_strategies) guide to learn more about search, sampling, and decoding strategies.
### Saving
Create an instance of [`GenerationConfig`] and specify the decoding parameters you want.
```py
>>> model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to(DEVICE)
from transformers import AutoModelForCausalLM, GenerationConfig
>>> # By default, the output will contain up to 20 tokens
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
model = AutoModelForCausalLM.from_pretrained("my_account/my_model")
generation_config = GenerationConfig(
max_new_tokens=50, do_sample=True, top_k=50, eos_token_id=model.config.eos_token_id
)
```
Use [`~GenerationConfig.save_pretrained`] to save a specific generation configuration and set the `push_to_hub` parameter to `True` to upload it to the Hub.
```py
generation_config.save_pretrained("my_account/my_model", push_to_hub=True)
```
Leave the `config_file_name` parameter empty. This parameter should be used when storing multiple generation configurations in a single directory. It gives you a way to specify which generation configuration to load. You can create different configurations for different generative tasks (creative text generation with sampling, summarization with beam search) for use with a single model.
```py
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, GenerationConfig
tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")
model = AutoModelForSeq2SeqLM.from_pretrained("google-t5/t5-small")
translation_generation_config = GenerationConfig(
num_beams=4,
early_stopping=True,
decoder_start_token_id=0,
eos_token_id=model.config.eos_token_id,
pad_token=model.config.pad_token_id,
)
translation_generation_config.save_pretrained("/tmp", config_file_name="translation_generation_config.json", push_to_hub=True)
generation_config = GenerationConfig.from_pretrained("/tmp", config_file_name="translation_generation_config.json")
inputs = tokenizer("translate English to French: Configuration files are easy to use!", return_tensors="pt")
outputs = model.generate(**inputs, generation_config=generation_config)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
```
## Pitfalls
The section below covers some common issues you may encounter during text generation and how to solve them.
### Output length
[`~GenerationMixin.generate`] returns up to 20 tokens by default unless otherwise specified in a models [`GenerationConfig`]. It is highly recommended to manually set the number of generated tokens with the [`max_new_tokens`] parameter to control the output length. [Decoder-only](https://hf.co/learn/nlp-course/chapter1/6?fw=pt) models returns the initial prompt along with the generated tokens.
```py
model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to("cuda")
```
<hfoptions id="output-length">
<hfoption id="default length">
```py
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'A sequence of numbers: 1, 2, 3, 4, 5'
```
>>> # Setting `max_new_tokens` allows you to control the maximum length
>>> generated_ids = model.generate(**model_inputs, max_new_tokens=50)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
</hfoption>
<hfoption id="max_new_tokens">
```py
generated_ids = model.generate(**model_inputs, max_new_tokens=50)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'A sequence of numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,'
```
### Incorrect generation mode
</hfoption>
</hfoptions>
By default, and unless specified in the [`~generation.GenerationConfig`] file, `generate` selects the most likely token at each iteration (greedy decoding). Depending on your task, this may be undesirable; creative tasks like chatbots or writing an essay benefit from sampling. On the other hand, input-grounded tasks like audio transcription or translation benefit from greedy decoding. Enable sampling with `do_sample=True`, and you can learn more about this topic in this [blog post](https://huggingface.co/blog/how-to-generate).
### Decoding strategy
The default decoding strategy in [`~GenerationMixin.generate`] is *greedy search*, which selects the next most likely token, unless otherwise specified in a models [`GenerationConfig`]. While this decoding strategy works well for input-grounded tasks (transcription, translation), it is not optimal for more creative use cases (story writing, chat applications).
For example, enable a [multinomial sampling](./generation_strategies#multinomial-sampling) strategy to generate more diverse outputs. Refer to the [Generation strategy](./generation_strategies) guide for more decoding strategies.
```py
>>> # Set seed for reproducibility -- you don't need this unless you want full reproducibility
>>> from transformers import set_seed
>>> set_seed(42)
>>> model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to(DEVICE)
>>> # LLM + greedy decoding = repetitive, boring output
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'I am a cat. I am a cat. I am a cat. I am a cat'
>>> # With sampling, the output becomes more creative!
>>> generated_ids = model.generate(**model_inputs, do_sample=True)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'I am a cat. Specifically, I am an indoor-only cat. I'
model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to("cuda")
```
### Wrong padding side
LLMs are [decoder-only](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt) architectures, meaning they continue to iterate on your input prompt. If your inputs do not have the same length, they need to be padded. Since LLMs are not trained to continue from pad tokens, your input needs to be left-padded. Make sure you also don't forget to pass the attention mask to generate!
<hfoptions id="decoding">
<hfoption id="greedy search">
```py
>>> # The tokenizer initialized above has right-padding active by default: the 1st sequence,
>>> # which is shorter, has padding on the right side. Generation fails to capture the logic.
>>> model_inputs = tokenizer(
... ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
... ).to(DEVICE)
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'1, 2, 33333333333'
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```
>>> # With left-padding, it works as expected!
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", padding_side="left")
>>> tokenizer.pad_token = tokenizer.eos_token # Most LLMs don't have a pad token by default
>>> model_inputs = tokenizer(
... ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
... ).to(DEVICE)
>>> generated_ids = model.generate(**model_inputs)
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
</hfoption>
<hfoption id="multinomial sampling">
```py
generated_ids = model.generate(**model_inputs, do_sample=True)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```
</hfoption>
</hfoptions>
### Padding side
Inputs need to be padded if they don't have the same length. But LLMs aren't trained to continue generation from padding tokens, which means the [`~PreTrainedTokenizer.padding_side`] parameter needs to be set to the left of the input.
<hfoptions id="padding">
<hfoption id="right pad">
```py
model_inputs = tokenizer(
["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
).to("cuda")
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'1, 2, 33333333333'
```
</hfoption>
<hfoption id="left pad">
```py
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1", padding_side="left")
tokenizer.pad_token = tokenizer.eos_token
model_inputs = tokenizer(
["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
).to("cuda")
generated_ids = model.generate(**model_inputs)
tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
'1, 2, 3, 4, 5, 6,'
```
### Wrong prompt
</hfoption>
</hfoptions>
Some models and tasks expect a certain input prompt format to work properly. When this format is not applied, you will get a silent performance degradation: the model kinda works, but not as well as if you were following the expected prompt. More information about prompting, including which models and tasks need to be careful, is available in this [guide](tasks/prompting). Let's see an example with a chat LLM, which makes use of [chat templating](chat_templating):
### Prompt format
```python
>>> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-alpha")
>>> model = AutoModelForCausalLM.from_pretrained(
... "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", load_in_4bit=True
... )
>>> set_seed(0)
>>> prompt = """How many helicopters can a human eat in one sitting? Reply as a thug."""
>>> model_inputs = tokenizer([prompt], return_tensors="pt").to(DEVICE)
>>> input_length = model_inputs.input_ids.shape[1]
>>> generated_ids = model.generate(**model_inputs, max_new_tokens=20)
>>> print(tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0])
"I'm not a thug, but i can tell you that a human cannot eat"
>>> # Oh no, it did not follow our instruction to reply as a thug! Let's see what happens when we write
>>> # a better prompt and use the right template for this model (through `tokenizer.apply_chat_template`)
Some models and tasks expect a certain input prompt format, and if the format is incorrect, the model returns a suboptimal output. You can learn more about prompting in the [prompt engineering](./tasks/prompting) guide.
>>> set_seed(0)
>>> messages = [
... {
... "role": "system",
... "content": "You are a friendly chatbot who always responds in the style of a thug",
... },
... {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
... ]
>>> model_inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(DEVICE)
>>> input_length = model_inputs.shape[1]
>>> generated_ids = model.generate(model_inputs, do_sample=True, max_new_tokens=20)
>>> print(tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0])
'None, you thug. How bout you try to focus on more useful questions?'
>>> # As we can see, it followed a proper thug style 😎
For example, a chat model expects the input as a [chat template](./chat_templating). Your prompt should include a `role` and `content` to indicate who is participating in the conversation. If you try to pass your prompt as a single string, the model doesn't always return the expected output.
```py
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-alpha")
model = AutoModelForCausalLM.from_pretrained(
"HuggingFaceH4/zephyr-7b-alpha", device_map="auto", load_in_4bit=True
)
```
## Further resources
<hfoptions id="format">
<hfoption id="no format">
While the autoregressive generation process is relatively straightforward, making the most out of your LLM can be a challenging endeavor because there are many moving parts. For your next steps to help you dive deeper into LLM usage and understanding:
```py
prompt = """How many cats does it take to change a light bulb? Reply as a pirate."""
model_inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
input_length = model_inputs.input_ids.shape[1]
generated_ids = model.generate(**model_inputs, max_new_tokens=50)
print(tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0])
"Aye, matey! 'Tis a simple task for a cat with a keen eye and nimble paws. First, the cat will climb up the ladder, carefully avoiding the rickety rungs. Then, with"
```
### Advanced generate usage
</hfoption>
<hfoption id="chat template">
1. Guide on how to [control different generation methods](generation_strategies), how to set up the generation configuration file, and how to stream the output;
2. [Accelerating text generation](llm_optims);
3. [Prompt templates for chat LLMs](chat_templating);
4. [Prompt design guide](tasks/prompting);
5. API reference on [`~generation.GenerationConfig`], [`~generation.GenerationMixin.generate`], and [generate-related classes](internal/generation_utils). Most of the classes, including the logits processors, have usage examples!
```py
messages = [
{
"role": "system",
"content": "You are a friendly chatbot who always responds in the style of a pirate",
},
{"role": "user", "content": "How many cats does it take to change a light bulb?"},
]
model_inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")
input_length = model_inputs.shape[1]
generated_ids = model.generate(model_inputs, do_sample=True, max_new_tokens=50)
print(tokenizer.batch_decode(generated_ids[:, input_length:], skip_special_tokens=True)[0])
"Arr, matey! According to me beliefs, 'twas always one cat to hold the ladder and another to climb up it an’ change the light bulb, but if yer looking to save some catnip, maybe yer can
```
### LLM leaderboards
</hfoption>
</hfoptions>
1. [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), which focuses on the quality of the open-source models;
2. [Open LLM-Perf Leaderboard](https://huggingface.co/spaces/optimum/llm-perf-leaderboard), which focuses on LLM throughput.
## Resources
### Latency, throughput and memory utilization
Take a look below for some more specific and specialized text generation libraries.
1. Guide on how to [optimize LLMs for speed and memory](llm_tutorial_optimization);
2. Guide on [quantization](main_classes/quantization) such as bitsandbytes and autogptq, which shows you how to drastically reduce your memory requirements.
### Related libraries
1. [`optimum`](https://github.com/huggingface/optimum), an extension of 🤗 Transformers that optimizes for specific hardware devices;
2. [`outlines`](https://github.com/outlines-dev/outlines), a library where you can constrain text generation (e.g. to generate JSON files);
3. [`SynCode`](https://github.com/uiuc-focal-lab/syncode), a library for context-free grammar guided generation (e.g. JSON, SQL, Python);
4. [`text-generation-inference`](https://github.com/huggingface/text-generation-inference), a production-ready server for LLMs;
5. [`text-generation-webui`](https://github.com/oobabooga/text-generation-webui), a UI for text generation;
6. [`logits-processor-zoo`](https://github.com/NVIDIA/logits-processor-zoo), containing additional options to control text generation with 🤗 Transformers. See our related [blog post](https://huggingface.co/blog/logits-processor-zoo).
- [Optimum](https://github.com/huggingface/optimum): an extension of Transformers focused on optimizing training and inference on specific hardware devices
- [Outlines](https://github.com/dottxt-ai/outlines): a library for constrained text generation (generate JSON files for example).
- [SynCode](https://github.com/uiuc-focal-lab/syncode): a library for context-free grammar guided generation (JSON, SQL, Python).
- [Text Generation Inference](https://github.com/huggingface/text-generation-inference): a production-ready server for LLMs.
- [Text generation web UI](https://github.com/oobabooga/text-generation-webui): a Gradio web UI for text generation.
- [logits-processor-zoo](https://github.com/NVIDIA/logits-processor-zoo): additional logits processors for controlling text generation.