florianvoss/Llama-3.1-8B-DFlash-N8-LLiMa

Llama 3.1 8B DFlash N8 for LLiMa

This repository contains a deployed DFlash target/draft pair compiled with LLiMa for SiMa.ai Modalix.

Models

• —Target: simaai/Llama-3.1-8B-Instruct-GPTQ-Safetensors
• —Draft: z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat

The parent directory contains one deployed target and one deployed draft model. LLiMa detects their roles from each devkit/vlm_config.json file and dispatches generation through the DFlash runtime.

Compilation configuration

• —DFlash verification width: 8 tokens (one anchor plus up to seven draft proposals)
• —Maximum sequence length: 2048
• —Language prefill group size: 128
• —Target precision: BF16 activations, INT4 weights
• —Draft precision: BF16 activations, INT8 weights
• —Quantized embeddings and KV cache
• —Filter sharing enabled
• —LLiMa source commit: aaff418a1d4fb5293a7a12504ef8a717673f297c on add_dflash

Directory layout

Llama-3.1-8B-DFlash-N8-LLiMa/
├── models--simaai--Llama-3.1-8B-Instruct-GPTQ-Safetensors/
│   ├── devkit/
│   └── elf_files/
└── models--z-lab--LLaMA3.1-8B-Instruct-DFlash-UltraChat/
    ├── devkit/
    └── elf_files/

These are compiled Modalix artifacts, not portable Hugging Face model weights. Use them with a DFlash-enabled LLiMa runtime.