Add mention of cuDNN in README

README.md

@@ -141,10 +141,14 @@ generate(model, prompt,
 
 ## CUDA GPU
 
+In addition to `JSON3` and `Jjama3`, ensure `CUDA.jl` and `Flux.jl` are installed.
+
 ```julia
 using CUDA, Flux, JSON3, Jjama3
 ```
 
+You might also need to install `cuDNN.jl` and run `using cuDNN` on some systems.
+
 For sampling, you can pass `device = gpu` to the `generate` function:
 
 ```julia
@@ -168,4 +172,4 @@ train_toks = encode(tkn, "This is a test.")
 gpu_train_toks = gpu(train_toks)
 
 forward_loss(model, gpu_train_toks[1:end-1,:], gpu_train_toks[2:end,:])
-```
+```