Model card: quantized relation, links to the vLLM checkpoints

README.md

@@ -4,6 +4,7 @@ tags:
 - vllm
 base_model:
 - arcinstitute/evo2_1b_base
+base_model_relation: quantized
 ---
 
 <img src="https://cdn-uploads.huggingface.co/production/uploads/649aee789fc303937a045f6a/IGUfG31MMvDzhdjRK-nlJ.jpeg" width="70%" />
@@ -26,19 +27,16 @@ It runs on GPUs with and without FP8 tensor cores. H100, H200 and L4 use its FP8
 
 For the Vortex backend, use [arcinstitute/evo2_1b_base](https://huggingface.co/arcinstitute/evo2_1b_base).
 
-Evo 2 40B and 7B checkpoints, trained up to 1 million sequence length, are available here:
+Evo 2 20B and 7B vLLM checkpoints, trained up to 1 million sequence length, are available here:
 | Checkpoint name | Num layers | Num parameters |
 |------------------------------|----|----------|
-| [evo2_40b](https://huggingface.co/arcinstitute/evo2_40b) | 50 | 40B |
-| [evo2_7b](https://huggingface.co/arcinstitute/evo2_7b) | 32 | 7B  | 
+| [evo2_20b](https://huggingface.co/gbrixi/evo2_20b_vllm) | 50 | 20B |
+| [evo2_7b](https://huggingface.co/gbrixi/evo2_7b_vllm) | 32 | 7B  |
 
-We also share 40B, 7B, and 1B base checkpoints trained on 8192 context length:
+We also share 7B and 1B base vLLM checkpoints trained on 8192 context length:
 | Checkpoint name | Num layers | Num parameters |
 |------------------------------|----|----------|
-| [evo2_40b_base](https://huggingface.co/arcinstitute/evo2_40b_base) | 50 | 40B |
-| [evo2_7b_base](https://huggingface.co/arcinstitute/evo2_7b_base) | 32 | 7B  | 
-| [evo2_1b_base](https://huggingface.co/arcinstitute/evo2_1b_base) | 25 | 1B  |
+| [evo2_7b_base](https://huggingface.co/gbrixi/evo2_7b_base_vllm) | 32 | 7B  |
+| [evo2_1b_base](https://huggingface.co/gbrixi/evo2_1b_base_vllm) | 25 | 1B  |
 
-## Conversion
-
-Converted from `evo2_1b_base.pt` in [arcinstitute/evo2_1b_base](https://huggingface.co/arcinstitute/evo2_1b_base) (sha256 `8ffba7d0e6445a8f2c92d9ff1c4e772c7f73ca9179f5e0c699b8b0ca1b966f64`). The FP8 input projections are kept as trained. `surgery/` holds the precomputed tensors for GPUs without FP8 tensor cores, and the sha256 of the weights they belong to. Check it with `python -m evo2.test.test_evo2 --model_name evo2_1b_base --backend vllm` (see the [Evo 2 GitHub](https://github.com/ArcInstitute/evo2)).
+Also available: [evo2_7b_262k](https://huggingface.co/gbrixi/evo2_7b_262k_vllm) (7B, 262,144-token context) and [evo2_7b_microviridae](https://huggingface.co/gbrixi/evo2_7b_microviridae_vllm) (7B base fine-tuned on Microviridae genomes).