Merge pull request #197 from university-of-york/feature/llama_cpp

tidy up

docs/source/applications/llama_cpp.rst

@@ -3,9 +3,9 @@ Llama.cpp
 
 The main goal of `llama.cpp <https://github.com/ggerganov/llama.cpp>`_ is to enable LLM inference with minimal setup and state-of-the-art performance on a wide variety of hardware - locally and in the cloud.
 
-One of the extras included in ``llama.cpp`` is a fast, lightweight, pure C/C++ HTTP server front end. With the help of some extra ``ssh`` tunnels, this can allow you to easily interact with a LLM running on Viking through the web browser on your PC. Below is an example of this workflow.
+One of the extras included in ``llama.cpp`` is a fast, lightweight, pure C/C++ HTTP server to act as a front end. With the help of some extra ``ssh`` tunnels, this can allow you to easily interact with a LLM running on Viking through the web browser on your PC. Below is an example of this workflow.
 
-To begin with you'll need to download an LLM to Viking, for example `mistral-7b-instruct-v0.2.Q4_K_M.gguf <https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF/blob/main/mistral-7b-instruct-v0.2.Q4_K_M.gguf>`_. You can download this on Viking with the ``wget`` command for example:
+To begin with you'll need to download an LLM to Viking, for example `mistral-7b-instruct-v0.2.Q4_K_M.gguf <https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF>`_. You can download this on Viking with the ``curl`` command for example:
 
 .. code-block:: console
 
@@ -35,7 +35,7 @@ To run the server, you need to pass it the path to the LLM you downloaded, along
 
     $ server -m /path/to/mistral-7b-instruct-v0.2.Q4_K_M.gguf --n-gpu-layers 256 --ctx-size 2048
 
-Here I've also set the ``n-gpu-layers`` option which allows offloading some layers to the GPU for computation. Generally results in increased performance, and the ``ctx-size`` is the size of the prompt context (the default is 512).
+Here I've also set the ``n-gpu-layers`` option which allows offloading some layers to the GPU for computation. Generally results in increased performance, and the ``ctx-size`` is the size of the prompt context (the default is 512). More options are covered in the `README.md <https://github.com/ggerganov/llama.cpp/tree/master/examples/server>`_.
 
 .. note::
 
@@ -49,7 +49,7 @@ Set up the ssh tunnels
 
 The ``llama.cpp`` server is running on a GPU compute node on Viking, behind the login node. This means you can't immediately load up a web browser and connect to it. But with *two* ``ssh`` tunnels, you can.
 
-1. To forward you local connection to a login node
+1. To forward from your local PC to the login node
 2. To forward from the login node to the GPU compute node, where the server is running
 
 On your local PC
@@ -94,7 +94,7 @@ If everything is working, you should now be able to connect to the server from y
 
 .. Note::
 
-    The above two ssh tunnel commands can be done in one single command however, it will have the effect to leaving one of the ssh tunnels running on the login node after you have logged out which you should really kill when you're finished. If you're familiar with killing processes on Linux, an example command which you would only run in a terminal on your PC (not on Viking) would be:
+    The above two ``ssh`` tunnel commands can be done in one single command however, it will have the effect to leaving one of the ``ssh`` tunnels running on the login node after you have logged out which you should really kill when you're finished. If you're familiar with killing processes on Linux, an example command which you would only run in a terminal on your local PC (not on Viking) would be:
 
     .. code-block:: console