setup NVidia GPU Docker for llama.cpp and run perplexity test

setup.sh

# BTW: we are running in a nvidia/cuda:11.x.x-devel-ubuntu22.04

# install some extra Ubuntu packages
  apt install unzip libopenblas-dev nano git-lfs aria2c jq build-essential python3 python3-pip git -y
  pip install --upgrade pip setuptools wheel

# clone llama.cpp repo
	cd /workspace
	git clone https://github.com/ggerganov/llama.cpp.git

# setup & build llama.cpp
	cd /workspace/llama.cpp
	pip install -r requirements.txt
	pip install --upgrade pip
	make LLAMA_CUBLAS=1 -j



## get Open Llama 3B v1
#	mkdir -p /workspace/3b_open_llama_v1
#	wget -O /workspace/3b_open_llama_v1/pytorch_model.bin https://huggingface.co/openlm-research/open_llama_3b/resolve/main/pytorch_model.bin
#	wget -O /workspace/3b_open_llama_v1/tokenizer.model https://huggingface.co/openlm-research/open_llama_3b/resolve/main/tokenizer.model
#	ls /workspace/3b_open_llama_v1
#
## convert model to ggml F16 format & delete original models
#	cd /workspace/llama.cpp
#	python3 convert.py /workspace/3b_open_llama_v1
#	rm /workspace/3b_open_llama_v1/pytorch_model.bin
#	rm /workspace/3b_open_llama_v1/tokenizer.model
#	ls /workspace/3b_open_llama_v1
#
## quantize to q4_0
#	cd /workspace/llama.cpp
#	./quantize /workspace/3b_open_llama_v1/ggml-model-f16.bin /workspace/3b_open_llama_v1/ggml-model-q4_0.bin q4_0

# get quantized models
	mkdir -p /workspace/7b_open_llama_v1
	cd /workspace/7b_open_llama_v1
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q2_K.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q3_K_S.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q3_K.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q3_K_L.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q4_0.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q4_1.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q4_K_S.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q4_K.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q5_0.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q5_K_S.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q5_1.bin
	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q5_K.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q6_K.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-q8_0.bin
#	wget https://huggingface.co/SlyEcho/open_llama_7b_ggml/resolve/main/open-llama-7b-f16.bin
	ls /workspace/7b_open_llama_v1

# get wiki text
	cd /workspace
	wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
	unzip wikitext-2-raw-v1.zip
	rm wikitext-2-raw-v1.zip
	cd /workspace/wikitext-2-raw
	head -n406 wiki.test.raw > wiki.test.raw.406
	head -n103 wiki.test.raw > wiki.test.raw.103

# run perplexity
	cd /workspace/llama.cpp
	export model="/workspace/7b_open_llama_v1/open-llama-7b-q5_K.bin"
	export corpus="/workspace/wikitext-2-raw/wiki.test.raw.103"
	export context=512
	export batch=512
	export threads=8
	export gpu=24
	./perplexity -m $model -f $corpus -c $context -b $batch -t $threads -ngl $gpu