Carmenest's picture
docs: improve model card with quickstart, benchmarks, Apache-2.0
f80c9ca verified
metadata
license: apache-2.0
tags:
  - diffusion
  - llada
  - gguf
  - cpu-inference
  - diffuse-cpp
language:
  - en
base_model: GSAI-ML/LLaDA-8B-Instruct
pipeline_tag: text-generation

LLaDA-8B-Instruct-GGUF

GGUF quantizations of GSAI-ML/LLaDA-8B-Instruct for use with diffuse-cpp, the first C++ inference engine for Diffusion Language Models.

LLaDA is a masked diffusion language model based on the Llama backbone. Unlike autoregressive models that generate one token at a time, LLaDA generates all tokens in parallel through iterative refinement — making it compute-bound rather than memory-bound on CPU.

On a 12-core CPU, LLaDA with diffuse-cpp reaches 27.7 tok/s on translation tasks — 3.3x faster than llama.cpp (8.51 tok/s) on the same hardware.

Available Quantizations

File Type Size Description
llada-8b-f16.gguf F16 ~14.9 GB Full precision, best quality
llada-8b-q8_0.gguf Q8_0 ~8.4 GB 8-bit quantization, near-lossless
llada-8b-q4km.gguf Q4_K_M ~5.1 GB 4-bit mixed, best speed/quality ratio

Recommended: Q4_K_M for most users.

Quick Start

# Download
huggingface-cli download diffuse-cpp/LLaDA-8B-Instruct-GGUF llada-8b-q4km.gguf

# Build diffuse-cpp
git clone --recursive https://github.com/iafiscal1212/diffuse-cpp.git
cd diffuse-cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

# Run
./build/diffuse-cli -m ../llada-8b-q4km.gguf \
    --tokens "128000,3923,374,279,6864,315,9822,30" \
    -n 256 -s 16 -t 12 --remasking entropy_exit

Performance

Benchmarked on AMD EPYC 4465P 12-Core, Q4_K_M, entropy_exit + inter-step cache, B=256:

Prompt No-Cache Cache Steps vs llama.cpp
Capital of France? 17.5 24.4 tok/s 3 2.9x
Translate to French 25.9 27.7 tok/s 2 3.3x
15 x 23? 12.8 15.7 tok/s 4 1.8x
Translate to Spanish 7.6 22.9 tok/s 7 2.7x
Python is_prime() 3.2 4.9 tok/s 16 0.6x
Poem about ocean 3.2 5.3 tok/s 16 0.6x
Why is sky blue? 3.3 12.0 tok/s 16 1.4x
List the planets 3.3 9.4 tok/s 15 1.1x
Average 9.6 15.3 tok/s 1.8x
  • Inter-step cache: 1.6x average speedup with no quality degradation
  • 6 of 8 prompts outperform llama.cpp (8.51 tok/s baseline)
  • LLaDA excels at translation tasks (converges in 2-5 steps)

Model Details

  • Architecture: Llama backbone with bidirectional (non-causal) attention
  • Parameters: 8B
  • Layers: 32
  • Hidden size: 4096
  • Attention: MHA (32 query heads, 32 KV heads)
  • FFN: SwiGLU, intermediate 12288
  • Vocabulary: 126,464 tokens
  • RoPE theta: 500,000
  • Mask token ID: 126336

Also Available

  • Dream-v0-Instruct-7B-GGUF — Qwen2.5 backbone, GQA. Excels at math and code (21.6 tok/s, correctly solves arithmetic in 2 steps).

Citation

@software{diffuse_cpp_2026,
  title={diffuse-cpp: High-Performance Inference for Diffusion Language Models},
  author={Carmen Esteban},
  year={2026},
  url={https://github.com/iafiscal1212/diffuse-cpp}
}

License

Apache 2.0