Self-Speculative Decoding

Speculative Decoding in Practice

In speculative decoding, a small (draft) model quickly generates multiple draft tokens, which are then verified in parallel by the large (target) model. While speculative decoding can effectively speed up the target model, in practice it is difficult to maintain or even obtain a proper draft model, especially when the target model is finetuned with customized data.

Self-Speculative Decoding

Built on top of the concept of “self-speculative decoding”, IPEX-LLM can now accelerate the original FP16 or BF16 model without the need of a separate draft model or model finetuning; instead, it automatically converts the original model to INT4, and uses the INT4 model as the draft model behind the scene. In practice, this brings ~30% speedup for FP16 and BF16 LLM inference latency on Intel GPU and CPU respectively.

Using IPEX-LLM Self-Speculative Decoding

Please refer to IPEX-LLM self-speculative decoding code snippets below, and the detailed GPU and CPU examples in the project repo.

model = AutoModelForCausalLM.from_pretrained(model_path,
           optimize_model=True,
           torch_dtype=torch.float16, #use bfloat16 on cpu
           load_in_low_bit="fp16", #use bf16 on cpu
           speculative=True, #set speculative to true
           trust_remote_code=True,
           use_cache=True)
output = model.generate(input_ids,
           max_new_tokens=args.n_predict,
           do_sample=False)