Skip to content
Other

Gemma 4 31B LiLiCorr Drafter

generation accelerator for Gemma 4

2 min readupdated

A lightweight 2.58B-parameter draft model built to speed up google/gemma-4-31B-it via speculative decoding (where a small, fast model predicts several tokens ahead and the larger model verifies them in a single pass). Trained from scratch by MeganEFlynn, it pairs a 5-layer Qwen3 decoder backbone with a LiLiCorr head. The model generates 7 speculative tokens per step (block size 8), cutting response latency without sacrificing the base model's output quality. It is aimed at developers and server operators running Gemma 4 31B in production on vLLM who want to squeeze out maximum generation speed.

Key features

  • Speculative decoding: predicts an 8-token block (7 speculative tokens per step) using greedy proposal and verification against the base model's hidden states.
  • Hybrid architecture: combines a 5-layer Qwen3 decoder (hidden size 5376, 32 attention heads, 2048 sliding-window attention) with a 2-layer LiLiCorr head (top-k 8, width 1024).
  • High prediction accuracy: after epoch 0 validation, overall accuracy (full_acc) reached 0.6048, with an expected accepted length (lilicorr_eal) of 4.7257 tokens.
  • Compact footprint: the bf16 model.safetensors weights file takes up just 5.16 GB.

What you can do

  • Speed up Gemma 4 31B inference: plug it in as a draft model in a supported vLLM build or via the speculators library to lower latency in chatbots and APIs.
  • Fine-tune or benchmark epoch 1: use this checkpoint (checkpoint_best after epoch 0) to benchmark speed on your own prompts, or wait for the final epoch 1 release.

Example

This model cannot generate text standalone and only runs paired with the google/gemma-4-31B-it verifier. To run it in vLLM, set this repository as your draft model using the draft_model method and a block size of 8, after replacing the local verifier path in config.json with the Hugging Face Hub model ID.

Read moreDetails
Validation (13,925 prompt groups): per-position accuracy across positions 1–7 was 0.8613 / 0.7433 / 0.6528 / 0.5800 / 0.5179 / 0.4636 / 0.4133, with a loss of 1.4685.
Training: 1,774,091 rows (2.14B tokens) from the nm-research/gemma-4-31b-it-responses-open-perfectblend dataset over 13,383 steps (epoch 0 of 2).
Limitations: requires the speculators library (lilicorr-clean / gemma4-perfectblend-lilicorr-clean branch) or a patched vLLM build with the lilicorr algorithm registered.
Discuss Telegram 1 60

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 4.8 GB · ready files5.8 GB
  • FP8 / INT8recommendedweights 2.4 GB3.1 GB
  • 4-bitweights 1.4 GB2.0 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # MeganEFlynn/gemma-4-31b-it-lilicorr-drafter
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download MeganEFlynn/gemma-4-31b-it-lilicorr-drafter --local-dir ~/models/gemma-4-31b-it-lilicorr-drafter
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. gemma-4-31Bfamily root
  2. gemma-4-31B-it
  3. gemma-4-31b-it-lilicorr-drafter

New posts about versions of gemma-4-31B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
Merges 3All on Hugging Face

About the model

MeganEFlynn/gemma-4-31b-it-lilicorr-drafter
Author
MeganEFlynn
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

601

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments