	IF hybrid verifier (`...-hybrid-smlp`)	This drafter
Purpose	Instruction-following answers	Speed up generation
Weights	~16.4 GiB	~2.1 GiB
Standalone?	Yes	No
Trained on	—	Hidden states from this verifier (5k ShareGPT, seq 8192)

Why a separate drafter?

The verifier keeps the shared MLP in BF16 and NVFP4 on MoE experts only. That shifts hidden states vs BF16 base — a drafter trained on the wrong verifier gets ~0% acceptance.

This checkpoint:

Extracted hidden states from ...-nvfp4-hybrid-smlp
Validation warm-start (50 samples) → scaled 5k train, 3 epochs
Validation loss 1.713 at checkpoint_best

Do not pair with any other Gemma 4 26B drafter (wrong verifier → ~0% acceptance).

vLLM — load both together

bash

vllm serve undeadindustries/gemma-4-26b-a4b-it-nvfp4-hybrid-smlp \
  --port 8000 --max-model-len 8192 \
  --gpu-memory-utilization 0.90 --max-num-seqs 8 \
  --max-num-batched-tokens 8192 \
  --enforce-eager --trust-remote-code --generation-config vllm \
  --speculative-config '{"model":"undeadindustries/gemma-4-26b-a4b-it-dflash-drafter-nvfp4-smlp","method":"dflash","num_speculative_tokens":15}'

Requirements: vLLM ≥ 0.20.2, TP=1.

Eval: use /v1/chat/completions for the -it verifier.

License

Derived from google/gemma-4-26B-A4B-it (Apache 2.0).
Subject to Gemma Terms of Use.