[docs] MTP support

docs/source/en/assisted_decoding.md

@@ -141,6 +141,26 @@ outputs = model.generate(**inputs, assistant_early_exit=4, do_sample=False, max_
 tokenizer.batch_decode(outputs, skip_special_tokens=True)
 ```
 
+## Multi-token prediction
+
+Multi-token prediction (MTP) drafts candidate tokens with extra prediction layers trained into the model itself, so it does not need a separate assistant model. The MTP layers predict several future tokens from a single position and share the main model's embeddings and output head. The main model verifies the drafts in one forward pass, which speeds up generation without loading a second model.
+
+MTP works only with checkpoints trained with MTP layers, such as [DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3) and [GLM-4.5](https://huggingface.co/zai-org/GLM-4.5). These models have a `num_mtp_layers` config value and MTP weights. [`~GenerationMixin.generate`] raises an error if the checkpoint has no MTP layers.
+
+Pass `use_mtp=True` to [`~GenerationMixin.generate`]. Like speculative decoding, MTP supports greedy search and sampling but not batched inputs.
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("zai-org/GLM-4.5-Air")
+model = AutoModelForCausalLM.from_pretrained("zai-org/GLM-4.5-Air", device_map="auto")
+inputs = tokenizer("Hugging Face is an open-source company", return_tensors="pt").to(model.device)
+
+outputs = model.generate(**inputs, use_mtp=True, do_sample=False, max_new_tokens=50)
+tokenizer.batch_decode(outputs, skip_special_tokens=True)
+```
+
 ## Universal assisted decoding
 
 Universal assisted decoding (UAD) makes speculative decoding possible even when the main and assistant models have different tokenizers. It lets you pair any small assistant model with the main model. Candidate tokens are re-encoded and the algorithm computes the longest common subsequence so the continuation stays aligned.