---
license: cc-by-nc-4.0
language:
- en
datasets:
- nvidia/Nemotron-ClimbMix
tags:
- llama
- matformer
- matformer-m
- vllm
- causal-lm
- pretraining
- climbmix
---

# atrost/climbmix-matformer-353m-1p2b-h100-m-vllm

Native `LlamaForCausalLM` conversion of `atrost/climbmix-matformer-353m-1p2b-h100-m`.

The source checkpoint was an extracted fixed-size MatFormer `M` submodel.
Because that submodel has fixed attention and MLP dimensions, its tensors can be
remapped losslessly into the standard Hugging Face Llama layout. This repo does
not require remote code and is intended to load through vLLM's native Llama path.

## Source

- Source checkpoint: `atrost/climbmix-matformer-353m-1p2b-h100-m`
- Source revision: `cb222bb9353c30c372c7478533a5cc6f9feac1d7`
- Local parity max logits diff before save: `0.000e+00`

## Loading

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "atrost/climbmix-matformer-353m-1p2b-h100-m-vllm"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, torch_dtype="auto")
```

```python
from vllm import LLM

llm = LLM(model="atrost/climbmix-matformer-353m-1p2b-h100-m-vllm")
```