# torchao

[![Open In Colab: Torchao Demo](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/transformers_doc/en/quantization/torchao.ipynb)

[torchao](https://github.com/pytorch/ao) es una biblioteca de optimización de arquitecturas de PyTorch con soporte para tipos de datos de alto rendimiento personalizados, cuantización y esparsidad. Es componible con funciones nativas de PyTorch, como [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html), para lograr una inferencia y un entrenamiento aún más rápidos.

Consulte la tabla siguiente para conocer las funciones adicionales de torchao.

| Característica | Descripción |
|--------|-------------|
| **Entrenamiento consciente de la cuantización (QAT)** | Entrena modelos cuantizados con una pérdida mínima de precisión (consulta el [README de QAT](https://github.com/pytorch/ao/blob/main/torchao/quantization/qat/README.md)) |
| **Entrenamiento en Float8** | Entrenamiento de alto rendimiento con formatos float8 (consulte la documentación de [torchtitan](https://github.com/pytorch/torchtitan/blob/main/docs/float8.md) y [Accelerate](https://huggingface.co/docs/accelerate/usage_guides/low_precision_training#configuring-torchao)) |
| **Soporte de esparsidad** | Esparsidad semiestructurada (2:4) para una inferencia más rápida (ver la publicación del blog [Acelerando el entrenamiento de redes neuronales con esparsidad semiestructurada (2:4)](https://pytorch.org/blog/accelerating-neural-network-training/)) |
| **Cuantización del optimizador** | Reduce la memoria del estado del optimizador con las variantes de 4 y 8 bits de Adam |
| **Cuantización de la caché KV** | Permite la inferencia de contexto largo con menor uso de memoria (ver [Cuantización de KV Cache](https://github.com/pytorch/ao/blob/main/torchao/_models/llama/README.md)) |
| **Soporte para kernels personalizados** | utilice sus propias operaciones compatibles con `torch.compile` |
| **FSDP2** | Componible con FSDP2 para el entrenamiento|

> [!TIP]
> Consulte el [README.md](https://github.com/pytorch/ao#torchao-pytorch-architecture-optimization) de torchao para obtener más detalles sobre la biblioteca.

torchao admite las [técnicas de cuantización](https://github.com/pytorch/ao/blob/main/torchao/quantization/README.md) siguientes.

- A16W8 Cuantización dinámica Float8
- Cuantización WeightOnly Float8 A16W8
- A8W8 Cuantización dinámica Int8
- Cuantización solo de pesos Int8 A16W8
- Cuantización solo de pesos Int4 A16W4
- A16W4 Cuantización solo de pesos Int4 + Esparsidad 2:4
- Autoquantización

torchao también permite la configuración a nivel de módulo especificando un diccionario que asocia el nombre totalmente cualificado de cada módulo con su configuración de cuantización correspondiente. Esto permite omitir la cuantización de ciertas capas y utilizar configuraciones de cuantización diferentes para distintos módulos.

Consulte la tabla siguiente para ver si su hardware es compatible.

| Componente | Compatibilidad |
|----------|----------------|
| Versiones de CUDA | ✅ cu118, cu126, cu128 |
| Versiones de XPU | ✅ pytorch2.8 |
| CPU | ✅ cambiar `device_map="cpu"` (ver los ejemplos a continuación) |

Instale torchao desde PyPi o el índice de PyTorch con los siguientes comandos.

```bash
# Updating 🤗 Transformers to the latest version, as the example script below uses the new auto compilation
# Stable release from Pypi which will default to CUDA 12.6
pip install --upgrade torchao transformers
```

Lanzamiento estable desde el índice de PyTorch

```bash
pip install torchao --index-url https://download.pytorch.org/whl/cu126 # options are cpu/cu118/cu126/cu128
```

torchao >= 0.15.0 is required. The string-based API (e.g., `TorchAoConfig("int4_weight_only")`) has been removed — use `AOBaseConfig` objects instead (see examples below).

## Ejemplos de cuantización

TorchAO ofrece una variedad de configuraciones de cuantización. Cada configuración se puede personalizar aún más con parámetros como `group_size`, `scheme` y `layout` para optimizar el rendimiento en hardware y arquitecturas de modelos específicos.

Para obtener una lista completa de las configuraciones disponibles, consulte la [documentación de la API de cuantización](https://github.com/pytorch/ao/blob/main/torchao/quantization/quant_api.py).

Puede elegir manualmente los tipos y ajustes de cuantización o seleccionar automáticamente los tipos de cuantización.

Cree una [TorchAoConfig](/docs/transformers/main/es/main_classes/quantization#transformers.TorchAoConfig) y especifique el tipo de cuantización y el `group_size` de los pesos a cuantizar (para int8 weight only e int4 weight only). Establezca `cache_implementation` en `"static"` para [torch.compile](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) automáticamente el método forward.

Mostraremos ejemplos de los métodos de cuantización recomendados según el hardware, por ejemplo, GPU A100, GPU H100, CPU.

> [!WARNING]
> torchao compila automáticamente el modelo durante la primera inferencia si establecemos `cache_implementation="static"`. El modelo se recompila cada vez que se modifica el tamaño del lote o `max_new_tokens`. Pasa `disable_compile=True` en [generate()](/docs/transformers/main/es/main_classes/text_generation#transformers.GenerationMixin.generate) para cuantizar sin compilación.

### GPU H100

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, Float8WeightOnlyConfig

quant_config = Float8DynamicActivationFloat8WeightConfig()
# or float8 weight only quantization
# quant_config = Float8WeightOnlyConfig()
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import GemliteUIntXWeightOnlyConfig

# We integrated with gemlite, which optimizes for batch size N on A100 and H100
quant_config = GemliteUIntXWeightOnlyConfig(group_size=128)
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

### GPU A100

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int8DynamicActivationInt8WeightConfig, Int8WeightOnlyConfig

quant_config = Int8DynamicActivationInt8WeightConfig()
# or int8 weight only quantization
# quant_config = Int8WeightOnlyConfig()
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import GemliteUIntXWeightOnlyConfig, Int4WeightOnlyConfig

# For batch size N, we recommend gemlite, which may require autotuning
# default is 4 bit, 8 bit is also supported by passing `bit_width=8`
quant_config = GemliteUIntXWeightOnlyConfig(group_size=128)

# For batch size 1, we also have custom tinygemm kernel that's only optimized for this
# We can set `use_hqq` to `True` for better accuracy
# quant_config = Int4WeightOnlyConfig(group_size=128, use_hqq=True)

quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

### Intel XPU

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int8DynamicActivationInt8WeightConfig, Int8WeightOnlyConfig

quant_config = Int8DynamicActivationInt8WeightConfig()
# or int8 weight only quantization
# quant_config = Int8WeightOnlyConfig()
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int4WeightOnlyConfig

quant_config = Int4WeightOnlyConfig(group_size=128, int4_packing_format="plain_int32")
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="auto",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device).to(quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

### CPU

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int8DynamicActivationInt8WeightConfig, Int8WeightOnlyConfig

quant_config = Int8DynamicActivationInt8WeightConfig()
# quant_config = Int8WeightOnlyConfig()
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="cpu",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

> [!TIP]
> La cuantización de solo pesos int4 en CPU requiere la versión 0.15.0 o posterior de torchao.

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.prototype.quantization.int4 import PrototypeInt4WeightOnlyConfig

quantization_config = TorchAoConfig(PrototypeInt4WeightOnlyConfig(group_size=128, int4_choose_qparams_algorithm="tinygemm"))

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="cpu",
    quantization_config=quantization_config
)

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(quantized_model.device).to(quantized_model.dtype)

# auto-compile the quantized model with `cache_implementation="static"` to get speed up
output = quantized_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static")
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

### Cuantización por módulo

#### 1. Omitir la cuantización para ciertas capas

Con `FqnToConfig` podemos especificar una configuración predeterminada para todas las capas, omitiendo la cuantización para ciertas capas.

```py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

model_id = "meta-llama/Llama-3.1-8B-Instruct"

from torchao.quantization import Int4WeightOnlyConfig, FqnToConfig
config = Int4WeightOnlyConfig(group_size=128)

# set default to int4 (for linears), and skip quantizing `model.layers.0.self_attn.q_proj`
quant_config = FqnToConfig({"_default": config, "model.layers.0.self_attn.q_proj": None})
quantization_config = TorchAoConfig(quant_type=quant_config)
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype=torch.bfloat16, quantization_config=quantization_config)
# lm_head is not quantized and model.layers.0.self_attn.q_proj is not quantized
print("quantized model:", quantized_model)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Manual Testing
prompt = "Hey, are you conscious? Can you talk to me?"
inputs = tokenizer(prompt, return_tensors="pt").to(quantized_model.device, quantized_model.dtype)
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
```

#### 2. Cuantizar capas diferentes con configuraciones de cuantización diferentes (sin regex)

```py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

model_id = "facebook/opt-125m"

from torchao.quantization import Int4WeightOnlyConfig, FqnToConfig, Int8DynamicActivationInt4WeightConfig, IntxWeightOnlyConfig, PerAxis, MappingType

weight_dtype = torch.int8
granularity = PerAxis(0)
mapping_type = MappingType.ASYMMETRIC
embedding_config = IntxWeightOnlyConfig(
    weight_dtype=weight_dtype,
    granularity=granularity,
    mapping_type=mapping_type,
)
linear_config = Int8DynamicActivationInt4WeightConfig(group_size=128)
quant_config = FqnToConfig({"_default": linear_config, "model.decoder.embed_tokens": embedding_config, "model.decoder.embed_positions": None})
# set `include_embedding` to True in order to include embedding in quantization
# when `include_embedding` is True, we'll remove input embedding from `modules_not_to_convert` as well
quantization_config = TorchAoConfig(quant_type=quant_config, include_embedding=True)
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cpu", dtype=torch.bfloat16, quantization_config=quantization_config)
print("quantized model:", quantized_model)
# make sure embedding is quantized
print("embed_tokens weight:", quantized_model.model.decoder.embed_tokens.weight)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Manual Testing
prompt = "Hey, are you conscious? Can you talk to me?"
inputs = tokenizer(prompt, return_tensors="pt").to("cpu", quantized_model.dtype)
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128, cache_implementation="static")
output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
```

#### 3. Cuantizar capas diferentes con configuraciones de cuantización distintas (con regex)

También podemos usar expresiones regulares para especificar la configuración de todos los módulos cuyo `module_fqn` coincida con la expresión regular; todas las expresiones regulares deben comenzar con `re:`. Por ejemplo, `re:layers\..*\.gate_proj` coincidirá con todas las capas como `layers.0.gate_proj`. Consulta la documentación [aquí](https://github.com/pytorch/ao/blob/2fe0ca0899c730c528efdbec8886feaa38879f39/torchao/quantization/quant_api.py#L2392).

```py
import logging

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

# Configure logging to see warnings and debug information
logging.basicConfig(
    level=logging.INFO, format="%(name)s - %(levelname)s - %(message)s"
)

# Enable specific loggers that might contain the serialization warnings
logging.getLogger("transformers").setLevel(logging.INFO)
logging.getLogger("torchao").setLevel(logging.INFO)
logging.getLogger("safetensors").setLevel(logging.INFO)
logging.getLogger("huggingface_hub").setLevel(logging.INFO)

model_id = "facebook/opt-125m"

from torchao.quantization import (
    Float8DynamicActivationFloat8WeightConfig,
    Int4WeightOnlyConfig,
    IntxWeightOnlyConfig,
    PerRow,
    PerAxis,
    FqnToConfig,
    Float8Tensor,
    Int4TilePackedTo4dTensor,
    IntxUnpackedToInt8Tensor,
)

float8dyn = Float8DynamicActivationFloat8WeightConfig(granularity=PerRow())
int4wo = Int4WeightOnlyConfig(int4_packing_format="tile_packed_to_4d")
intxwo = IntxWeightOnlyConfig(weight_dtype=torch.int8, granularity=PerAxis(0))

qconfig_dict = {
    # highest priority
    "model.decoder.layers.3.self_attn.q_proj": int4wo,
    "model.decoder.layers.3.self_attn.k_proj": int4wo,
    "model.decoder.layers.3.self_attn.v_proj": int4wo,
    # vllm
    "model.decoder.layers.3.self_attn.qkv_proj": int4wo,

    "re:model\.decoder\.layers\..+\.self_attn\.q_proj": float8dyn,
    "re:model\.decoder\.layers\..+\.self_attn\.k_proj": float8dyn,
    "re:model\.decoder\.layers\..+\.self_attn\.v_proj": float8dyn,
    # this should not take effect and we'll fallback to _default
    # since no full mach (missing `j` in the end)
    "re:model\.decoder\.layers\..+\.self_attn\.out_pro": float8dyn,
    # vllm
    "re:model\.decoder\.layers\..+\.self_attn\.qkv_proj": float8dyn,

    "_default": intxwo,
}
quant_config = FqnToConfig(qconfig_dict)
quantization_config = TorchAoConfig(quant_type=quant_config)
quantized_model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    quantization_config=quantization_config,
)
print("quantized model:", quantized_model)
tokenizer = AutoTokenizer.from_pretrained(model_id)
for i in range(12):
    if i == 3:
        assert isinstance(quantized_model.model.decoder.layers[i].self_attn.q_proj.weight, Int4TilePackedTo4dTensor)
        assert isinstance(quantized_model.model.decoder.layers[i].self_attn.k_proj.weight, Int4TilePackedTo4dTensor)
        assert isinstance(quantized_model.model.decoder.layers[i].self_attn.v_proj.weight, Int4TilePackedTo4dTensor)
    else:
        assert isinstance(quantized_model.model.decoder.layers[i].self_attn.q_proj.weight, Float8Tensor)
        assert isinstance(quantized_model.model.decoder.layers[i].self_attn.k_proj.weight, Float8Tensor)
        assert isinstance(quantized_model.model.decoder.layers[i].self_attn.v_proj.weight, Float8Tensor)
    assert isinstance(quantized_model.model.decoder.layers[i].self_attn.out_proj.weight, IntxUnpackedToInt8Tensor)

# Manual Testing
prompt = "What are we having for dinner?"
print("Prompt:", prompt)
inputs = tokenizer(
    prompt,
    return_tensors="pt",
).to(quantized_model.device, quantized_model.dtype)
# setting temperature to 0 to make sure result deterministic
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128, temperature=0)

correct_output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", correct_output_text[0][len(prompt) :])

# Load model from saved checkpoint
reloaded_model = AutoModelForCausalLM.from_pretrained(
    save_to,
    device_map=torch.accelerator.current_accelerator(),
    torch_dtype=torch.bfloat16,
    # quantization_config=quantization_config,
)

generated_ids = reloaded_model.generate(**inputs, max_new_tokens=128, temperature=0)
output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt) :])

assert(correct_output_text == output_text)
```

## Serialización

Guardar el modelo cuantizado con `save_pretrained` (en formato [safetensors](https://huggingface.co/docs/safetensors/en/index)) solo se admite para torchao >= v0.15. En cualquier versión inferior, solo es posible guardar manualmente como puntos de control no seguros `.bin` con [torch.save](https://docs.pytorch.org/docs/stable/generated/torch.save.html).

```py
# torchao >= 0.15
output_dir = "llama3-8b-int4wo-128"
quantized_model.save_pretrained("llama3-8b-int4wo-128")
```

```py
# torchao >= 0.15
USER_ID = "your_huggingface_user_id"
REPO_ID = "llama3-8b-int4wo-128"
quantized_model.push_to_hub(f"{USER_ID}/llama3-8b-int4wo-128")
tokenizer.push_to_hub(f"{USER_ID}/llama3-8b-int4wo-128")
```

```py
# torchao < 0.15 -> unsafe serialization
filename = "llama3-8b-int4wo-128/pytorch_model.bin"
torch.save(quantized_model.state_dict(), filename)
```

## Carga de modelos cuantizados

La carga de un modelo cuantizado depende del esquema de cuantización. En el caso de esquemas de cuantización como int8 y float8, se puede cuantizar el modelo en cualquier dispositivo y también cargarlo en cualquier dispositivo. El ejemplo siguiente demuestra cómo cuantizar un modelo en la CPU y luego cargarlo en CUDA o XPU.

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int8WeightOnlyConfig

quant_config = Int8WeightOnlyConfig(group_size=128)
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="cpu",
    quantization_config=quantization_config
)
# save the quantized model
output_dir = "llama-3.1-8b-torchao-int8"
quantized_model.save_pretrained(output_dir)

# reload the quantized model
reloaded_model = AutoModelForCausalLM.from_pretrained(
    output_dir,
    device_map="auto",
    dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(reloaded_model.device.type)

output = reloaded_model.generate(**input_ids, max_new_tokens=10)
print(tokenizer.decode(output[0], skip_special_tokens=True))

```

En el caso de int4, el modelo solo se puede cargar en el mismo dispositivo en el que se cuantizó, ya que el layout es específico del dispositivo. El siguiente ejemplo demuestra cómo cuantizar y cargar un modelo en la CPU.

```py
import torch
from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
from torchao.quantization import Int4WeightOnlyConfig
from torchao.quantization.quantize_.workflows import Int4PackingFormat

quant_config = Int4WeightOnlyConfig(group_size=128, int4_packing_format=Int4PackingFormat.PLAIN_INT32)
quantization_config = TorchAoConfig(quant_type=quant_config)

# Load and quantize the model
quantized_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-8B-Instruct",
    dtype="auto",
    device_map="cpu",
    quantization_config=quantization_config
)
# save the quantized model
output_dir = "llama-3.1-8b-torchao-int4-cpu"
quantized_model.save_pretrained(output_dir)

# reload the quantized model
reloaded_model = AutoModelForCausalLM.from_pretrained(
    output_dir,
    device_map="cpu",
    dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
input_text = "What are we having for dinner?"
input_ids = tokenizer(input_text, return_tensors="pt").to(reloaded_model.device.type)

output = reloaded_model.generate(**input_ids, max_new_tokens=10)
print(tokenizer.decode(output[0], skip_special_tokens=True))

```

## Recursos

Para tener una mejor idea del rendimiento esperado, consulte los [benchmarks](https://github.com/pytorch/ao/tree/main/torchao/quantization#benchmarks) de varios modelos con los backends CUDA y XPU. También puede ejecutar el siguiente código para realizar el benchmark de un modelo por su cuenta.

```py
from torch._inductor.utils import do_bench_using_profiling
from typing import Callable

def benchmark_fn(func: Callable, *args, **kwargs) -> float:
    """Thin wrapper around do_bench_using_profiling"""
    no_args = lambda: func(*args, **kwargs)
    time = do_bench_using_profiling(no_args)
    return time * 1e3

MAX_NEW_TOKENS = 1000
print("int4wo-128 model:", benchmark_fn(quantized_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))

bf16_model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", dtype=torch.bfloat16)
output = bf16_model.generate(**input_ids, max_new_tokens=10, cache_implementation="static") # auto-compile
print("bf16 model:", benchmark_fn(bf16_model.generate, **input_ids, max_new_tokens=MAX_NEW_TOKENS, cache_implementation="static"))
```

> [!TIP]
> Para obtener el mejor rendimiento, puede usar la configuración recomendada llamando a `torchao.quantization.utils.recommended_inductor_config_setter()`

Consulte [Otras técnicas de cuantización disponibles](https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques) para más ejemplos y documentación.

## Problemas

Si encuentras algún problema con la integración de Transformers, abre un issue en el repositorio de [Transformers](https://github.com/huggingface/transformers/issues). Para los issues directamente relacionados con torchao, abre un issue en el repositorio de [torchao](https://github.com/pytorch/ao/issues).