models : Added support for RND1 Diffusion Language Model

[wp4032]

RND1 works on llama.cpp, can run:

llama-diffusion-cli -m RND1-Base-0910.gguf -p "write code to train MNIST in pytorch" -ub 256 --diffusion-algorithm 1 --diffusion-steps 256 --diffusion-visual --temp 0.5

Model Card

https://huggingface.co/radicalnumerics/RND1-Base-0910

Instructions

# Create conda env
cd llama.cpp && conda create --name rnd1 python=3.12
conda activate rnd1
pip install -r requirements.txt

# Converting to gguf
huggingface-cli download radicalnumerics/RND1-Base-0910 --local-dir RND1-Base-0910
python llama.cpp/convert_hf_to_gguf.py RND1-Base-0910/ --outfile RND1-Base-0910.gguf --outtype bf16 

# Building diffusion cli
cmake -B build    # Will build with Metal automatically
cmake --build llama.cpp/build --target llama-diffusion-cli -j
llama.cpp/build/bin/llama-diffusion-cli -m RND1-Base-0910.gguf -p "What is a GPU?" -ub 32 --temp 0.01 -ngl 999 -fa on --seed 1234 --verbose

Results

Works on GB200, non-causal, BF16, 512 context len, 256 diffusion steps:

total time: 55793.55ms, time per step: 217.94ms, sampling time per step: 52.18ms

Works on H100, non-causal, BF16, 512 context len, 256 diffusion steps:

total time: 55602.47ms, time per step: 217.20ms, sampling time per step: 63.52ms

Works on Mac Studio M3 Ultra, non-causal, BF16, 512 context len, 256 diffusion steps:

total time: 65817.30ms, time per step: 257.10ms, sampling time per step: 16.65ms

[CISC]

Nice, thank you!

[am17an]

Please update or close the issue with a comment #17291

[wp4032]

Please update or close the issue with a comment #17291

Regarding this, I am still looking for the cause of the issue, however, I am quite confident it is not to do with RND1's implementation. If you really want me to close the issue #17291, I can.

[am17an]

Regarding this, I am still looking for the cause of the issue, however, I am quite confident it is not to do with RND1's implementation. If you really want me to close the issue #17291, I can.

I don't really want you to close the issue, however if it is a correctness bug as you mention there then the outputs for RND1 will also suffer from the same issue, I'm not sure why you are saying it is not to do with RND1's implementation as that is the same implementation as Qwen2. I would rather see that resolved before merging.