| --- |
| tags: |
| - masked-diffusion |
| - diffusion |
| - qwen2 |
| - code |
| language: |
| - en |
| pipeline_tag: text-generation |
| --- |
| |
| # dlm-code0.6b-exp |
|
|
| **Experimental** masked-diffusion language model for code, from the [thunder-fast](https://github.com/AdrianTuci1/thunder-fast) project. |
|
|
| This is a **discrete masked-diffusion (MDM)** adaptation of a Qwen2-based pretrained code model (~0.5B). It is *not* an autoregressive next-token model: generation is a progressive un-masking loop under **bidirectional attention**, so a whole output window is filled in parallel over 24 diffusion steps. |
|
|
| ## Details |
|
|
| | | | |
| |---|---| |
| | Backbone | Qwen2 (`model_type: qwen2`), 24 layers, hidden 896, 14 attention / 2 KV heads, GQA | |
| | Vocabulary | 151,936 + `<M>` mask token (id 151665) | |
| | Diffusion | discrete `[MASK]`, 24-step progressive un-masking, entropy-based position confidence (`alg_temp` 0.6, `top_k` 500) | |
| | Context | window-based (256 tokens) — not left-to-right; supports infilling | |
| | Weights | bf16 (`model.safetensors`) | |
|
|
| ## How to run |
|
|
| This checkpoint is intended to be driven by a **diffusion decoder** (bidirectional attention + the iterative un-masking loop), e.g. the thunder-fast runtime. Loading it as a plain causal `Qwen2ForCausalLM` will *not* reproduce diffusion generation. |
|
|
| ``` |
| # thunder-fast runtime (ggml engine) or the reference DiffusionLM |
| # e.g. runtime --model dlm-code0.6b-exp.gguf --prompt "..." |
| ``` |
|
|
| ## Status |
|
|
| Experimental / work-in-progress (`.exp`). It is a diffusion code-generation checkpoint published for the thunder-fast project; see the repo for training, the masking schedule, and evaluation details. |
|
|
