Add llama.cpp role

multipass/run.sh

@@ -23,7 +23,7 @@ fi
 if ! multipass list | grep $vm_name | grep Running > /dev/null
 then
     echo 'Creating and provisioning multipass VM...'
-    multipass launch --name $vm_name $ubuntu_version
+    multipass launch --name $vm_name $ubuntu_version --disk 15G
     multipass stop $vm_name
     multipass mount --type native $script_dir/../roles $vm_name:$target_dir/.ansible/roles
     multipass mount --type native $script_dir $vm_name:$target_dir/test

roles/llama.cpp/README.md

@@ -0,0 +1,8 @@
+# llama.cpp role
+This role installs [llama.cpp](https://github.com/ggml-org/llama.cpp), relying on a built version of llama.cpp created by 
+https://github.com/guardian/transcription-service/blob/main/.github/workflows/build-llama-cpp.yaml
+
+You can control which build of llama.cpp gets baked in via the llama_cpp_build_stage flag and the riffraff project 
+`investigations::transcription-service-llama-cpp`.
+
+This role expects the AWS Deep Learning AMI as the base image, which provides CUDA drivers and toolkit out of the box.

roles/llama.cpp/defaults/main.yml

@@ -0,0 +1,3 @@
+---
+target_dir: /opt/llama
+llama_cpp_build_stage: PROD

roles/llama.cpp/meta/main.yml

@@ -0,0 +1,3 @@
+---
+dependencies:
+  - aws-tools

roles/llama.cpp/tasks/main.yml

@@ -0,0 +1,34 @@
+---
+- name: Create target and models directory
+  file:
+    path: "{{ target_dir }}"
+    state: directory
+    owner: ubuntu
+    group: ubuntu
+    recurse: yes
+
+- name: Download compiled llama.cpp project
+  shell: |
+    aws s3 cp s3://amigo-data-{{ model_script_stage.lower() }}/deploy/{{ llama_cpp_build_stage }}/llama-cpp/llama-cpp-cuda.zip /tmp/llama-cpp-cuda.zip
+    unzip -o /tmp/llama-cpp-cuda.zip -d {{ target_dir }}
+
+- name: Set ownership of llama directory
+  file:
+    path: "{{ target_dir }}"
+    owner: ubuntu
+    group: ubuntu
+    recurse: yes
+
+- name: Symlink llama-cli into /usr/local/bin
+  file:
+    src: "{{ target_dir }}/llama.cpp/install/bin/llama-cli"
+    dest: /usr/local/bin/llama-cli
+    state: link
+
+- name: Symlink llama-server into /usr/local/bin
+  file:
+    src: "{{ target_dir }}/llama.cpp/install/bin/llama-server"
+    dest: /usr/local/bin/llama-server
+    state: link
+
+