Improve documentation for the all-linear flag (#1357)

* added docs for all-linear

* added doc in quantization section

* added doc in lora section

* minor edit

* minor edit

docs/source/developer_guides/lora.md

@@ -179,4 +179,12 @@ model.unload()
 
 # delete adapter
 model.delete_adapter("dpo")
+```
+
+## QLoRA-style training
+
+The default LoRA settings in 🤗PEFT follow the [original paper](https://hf.co/papers/2106.09685) and add trainable weights to the query and value layers of each attention block. However, in [QLoRA](https://hf.co/papers/2305.14314), it was found that adding trainable weights to all the linear layers of a transformer model is beneficial to match full-finetuning performance. Since the list of modules to add will vary depending on the architecture, we provided a convenient shorthand : simple specify `target_modules='all-linear'` and let 🤗PEFT handle the rest: 
+
+```py
+config = LoraConfig(target_modules="all-linear", ...) # adds LoRA to all linear layers like in QLoRA
 ```

docs/source/developer_guides/quantization.md

@@ -125,6 +125,12 @@ lora_config = LoraConfig(
 
 model = get_peft_model(model, lora_config)
 ```
+### QLoRA-style training
+QLoRA adds trainable weights to all the linear layers in the transformer architecture. Since the attribute names for these linear layers can vary across architectures, we provide a convenient flag `'all-linear'` for this setting:
+
+```py
+config = LoraConfig(target_modules="all-linear", ...) # adds LoRA to all linear layers like in QLoRA
+```
 
 ## Next steps