setup of ChatRWKV

[bello7777]

Hey guys great stuff, can we have very easy setup step process to install ChatRWKV on a ubuntu server for example?

[BlinkDL]

python 3.8/3.9/3.10

pip install numpy tokenizers prompt_toolkit ninja
pip install torch --extra-index-url https://download.pytorch.org/whl/cu117 --upgrade (use 1.13.1)
pip install rwkv --upgrade

:)

[soulteary]

There is no need to toss the environment, just use the container @bello7777

#58

[bello7777]

Thanks mate, I will do it .
for the moment I just launch it on AWS ubuntu 18.04 ,

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 200.00 MiB (GPU 0; 14.62 GiB total capacity; 13.77 GiB already allocated; 163.94 MiB free; 13.97 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

can I try to reduce the batch sizes to smaller values if yes where are they?

[bello7777]

@soulteary
i tried to access your blog and guidelines but i could not,
could you give me the steps and version of Docker container so i can deploy it on ec2 server as a still have a problem with memory

[KerfuffleV2]

@bello7777 You probably need to adjust the strategy. If you're using the pull request:

    model = RWKV(model=model_path, strategy='cuda fp16i8 *20 -> cuda fp16')

That's around line 28 in webui.py from that pull. You didn't say what you're actually doing, so there's no way to know if you're saying it failed when trying to load the model for inference, when converting, whatever.

But the most likely solution is to find whatever is running and how it's setting the strategy and reduce the number of layers it will send to the GPU. For example, in the line above you could try using cuda fp16i8 *10 -> cuda fp16 instead which should roughly half the required GPU memory.

After you get it going, you can use other tools to see how much GPU memory you have available and adjust the setting according.

[bello7777]

thanks solved and working

now moving to train the model