luke
luke
> I added a simple matrix multiplication and addition calculation: https://github.com/passchaos/ml-bench/blob/main/rust-ml/src/main.rs#L44, but it takes much longer than PyTorch. I wonder if it's a problem with my usage. > > framework...
great. it would be awesome to have more training code examples and workflows with candle
this sorta thing is great. i wonder if there is enough people with merge authority to handle this inside of the candle repo...
I thought it might be quantized thing. So running the same test on the Non-Quantized Qwen 3 8B i get the same error.
compilation errors are sometimes because missing libraries in the path. do you have the proper coda paths set.
i have these in debian 12 ``` export CUDA_ROOT=/usr/local/cuda export CUDA_PATH=/usr/local/cuda export CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda export CUDA_HOME=/usr/local/cuda export LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/nvidia/current ```
adding the bin to to the cudnn helped me run the program but the problem is that compiled app still needs the cudnn to dynamically link... isnt there someway like...
for deployment i needed to bundle cuda and cudnn with my app.
when i run the 4b and 12b like ``` cargo run --example gemma_vision --release --features cuda,cudnn -- --prompt "describe this image" --which v3_4b --image /home/pro/Desktop/sd_final.png ``` it downloads the models...
is this working? its been sitting here for a while would be great to have Gemma vision models in candle especially if the work is done..?