gbrixi/evo2_20b_vllm

Evo 2 20B, 1M context (vLLM)

Evo 2 is a state-of-the-art DNA language model trained autoregressively on trillions of DNA tokens.

For instructions, details, and examples, please refer to the GitHub and paper.

This is the vLLM checkpoint of arcinstitute/evo2_20b: the same model, converted to safetensors for the vLLM backend of Evo 2. Load it with

from evo2 import Evo2

model = Evo2('evo2_20b', backend='vllm')

It runs on GPUs with and without FP8 tensor cores. H100, H200 and L4 use its FP8 weights; on other GPUs, such as the A100, an FP8-free form is built once from surgery/.

For the Vortex backend, use arcinstitute/evo2_20b.

Model Details

• Base Model: Evo 2 20B
• Context Length: 1 million tokens
• Parameters: 20B
• Architecture: 50 layers

Main Evo 2 Checkpoints

Evo 2 20B and 7B vLLM checkpoints, trained up to 1 million sequence length, are available here:

Checkpoint name	Num layers	Num parameters
evo2_20b	50	20B
evo2_7b	32	7B

We also share 7B and 1B base vLLM checkpoints trained on 8192 context length:

Checkpoint name	Num layers	Num parameters
evo2_7b_base	32	7B
evo2_1b_base	25	1B

Also available: evo2_7b_262k (7B, 262,144-token context) and evo2_7b_microviridae (7B base fine-tuned on Microviridae genomes).

Usage

Please refer to the Evo 2 GitHub repository for detailed usage instructions and examples.