vllm.v1.spec_decode

Modules:

• custom_class_proposer –
• dflash –
• dynamic –
• extract_hidden_states –
• fly –

  Entropy-gated deferred verification for FLy speculative decoding.

• gemma4 –

  Gemma4 MTP (Multi-Token Prediction) proposer for speculative decoding.

• llm_base_proposer –
• metrics –
• ngram_proposer –
• ngram_proposer_gpu –

  GPU-accelerated N-gram proposer using fully async PyTorch tensor operations.

• step3p5 –
• suffix_decoding –
• utils –
• vocab_mapping –