Arthur Zucker(@art_zucker):Your are making claims that extrapolate the actual impact of this. 1. There is no mega kernel equivalent for GPU 2. Its a best case with 6 tokens accepted from draft model 3. Only valid for low volume of request. Is that a real use case? vLLM is cooking, yes, but not it’s not as easy as that

Arthur Zucker

@art_zucker

Head of transformers @huggingface 🤗

Joined October 2021

679 Following 8.4K Followers

Arthur Zucker@art_zucker

2026.09.24 11:59

Your are making claims that extrapolate the actual impact of this. 1. There is no mega kernel equivalent for GPU 2. Its a best case with 6 tokens accepted from draft model 3. Only valid for low volume of request. Is that a real use case? vLLM is cooking, yes, but not it’s not as easy as that