Clarify Accelerate usage in SFTTrainer documentation

docs/source/sft_trainer.md

@@ -25,6 +25,69 @@ trainer.train()
 
 <iframe src="https://trl-lib-trackio.hf.space/?project=trl-documentation&metrics=train*&runs=sft_qwen3-0.6B_capybara" style="width: 100%; min-width: 300px; max-width: 800px;" height="830" frameBorder="0"></iframe>
 
+## Using SFTTrainer with Accelerate
+
+`SFTTrainer` is built on top of the  Transformers `Trainer` and relies on
+ Accelerate for device placement and distributed training.
+
+When training with `SFTTrainer`, models should be loaded normally using
+`from_pretrained`, and training scripts should be launched with
+`accelerate launch`. Device placement, multi-GPU execution, and
+distributed strategies are handled by Accelerate at runtime.
+
+The `device_map="auto"` option is intended for inference-time model sharding
+and should not be used during training, as it can conflict with
+Accelerate-managed execution and lead to runtime errors.
+
+### Minimal example
+
+```bash
+accelerate launch train_sft.py
+```
+> ⚠️ **Important**
+>
+> Do not pass `device_map="auto"` to `from_pretrained()` when using
+> `SFTTrainer`, as this option is intended for inference and can conflict
+> with Accelerate-managed training.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
+from trl import SFTTrainer
+from datasets import Dataset
+
+model_name = "facebook/opt-125m"
+
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+tokenizer.pad_token = tokenizer.eos_token
+
+model = AutoModelForCausalLM.from_pretrained(model_name)
+
+dataset = Dataset.from_dict(
+    {
+        "text": [
+            "Explain supervised fine-tuning.",
+            "What is tensor parallelism?",
+        ]
+    }
+)
+
+training_args = TrainingArguments(
+    output_dir="./sft_output",
+    per_device_train_batch_size=1,
+    num_train_epochs=1,
+    logging_steps=1,
+)
+
+trainer = SFTTrainer(
+    model=model,
+    args=training_args,
+    train_dataset=dataset,
+    tokenizer=tokenizer,
+)
+
+trainer.train()
+```
+
 ## Expected dataset type and format
 
 SFT supports both [language modeling](dataset_formats#language-modeling) and [prompt-completion](dataset_formats#prompt-completion) datasets. The [`SFTTrainer`] is compatible with both [standard](dataset_formats#standard) and [conversational](dataset_formats#conversational) dataset formats. When provided with a conversational dataset, the trainer will automatically apply the chat template to the dataset.