[Docs/quantization] Clearer explanation on how things works under the hood. + remove outdated info (#25216)

* clearer explanation on how things works under the hood.

* Update docs/source/en/main_classes/quantization.md

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update docs/source/en/main_classes/quantization.md

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* add `load_in_4bit` in `from_pretrained`

---------

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

docs/source/en/main_classes/quantization.md

@@ -29,6 +29,29 @@ If you want to quantize your own pytorch model, check out this [documentation](h
 
 Here are the things you can do using `bitsandbytes` integration
 
+### General usage
+
+You can quantize a model by using the `load_in_8bit` or `load_in_4bit` argument when calling the [`~PreTrainedModel.from_pretrained`] method as long as your model supports loading with 🤗 Accelerate and contains `torch.nn.Linear` layers. This should work for any modality as well.
+
+```python
+from transformers import AutoModelForCausalLM
+
+model_8bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_8bit=True)
+model_4bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_4bit=True)
+```
+
+By default all other modules (e.g. `torch.nn.LayerNorm`) will be converted in `torch.float16`, but if you want to change their `dtype` you can overwrite the `torch_dtype` argument:
+
+```python
+>>> import torch
+>>> from transformers import AutoModelForCausalLM
+
+>>> model_8bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_8bit=True, torch_dtype=torch.float32)
+>>> model_8bit.model.decoder.layers[-1].final_layer_norm.weight.dtype
+torch.float32
+```
+
+
 ### FP4 quantization 
 
 #### Requirements
@@ -41,7 +64,7 @@ Make sure that you have installed the requirements below before running any of t
 - Install latest `accelerate`
 `pip install --upgrade accelerate`
 
-- Install latest `transformers` from source 
+- Install latest `transformers`
 `pip install --upgrade transformers`
 
 #### Tips and best practices

src/transformers/modeling_utils.py

@@ -2126,10 +2126,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 `True` when there is some disk offload.
             load_in_8bit (`bool`, *optional*, defaults to `False`):
                 If `True`, will convert the loaded model into mixed-8bit quantized model. To use this feature please
-                install `bitsandbytes` compiled with your CUDA version by running `pip install -i
+                install `bitsandbytes` (`pip install -U bitsandbytes`).
-                https://test.pypi.org/simple/ bitsandbytes-cudaXXX` where XXX is your CUDA version (e.g. 11.6 = 116).
+            load_in_4bit (`bool`, *optional*, defaults to `False`):
-                Make also sure that you have enough GPU RAM to store half of the model size since the 8bit modules are
+                If `True`, will convert the loaded model into 4bit precision quantized model. To use this feature
-                not compiled and adapted for CPUs.
+                install the latest version of `bitsandbytes` (`pip install -U bitsandbytes`).
             quantization_config (`Dict`, *optional*):
                 A dictionary of configuration parameters for the `bitsandbytes` library and loading the model using
                 advanced features such as offloading in fp32 on CPU or on disk.