ConfigLMM 0.4.0 → 0.5.0

data/Plugins/Apps/llama.cpp/llama.cpp.lmm.rb

@@ -0,0 +1,90 @@
+
+
+module ConfigLMM
+    module LMM
+        class Llamacpp < Framework::Plugin
+
+            USER = 'llama.cpp'
+            HOME_DIR = '/var/lib/llama.cpp'
+            PORT = 18900
+
+            def actionLlamacppDeploy(id, target, activeState, context, options)
+                raise Framework::PluginProcessError.new('Model field must be set!') unless target['Model']
+
+                self.withConnection(target['Location'], target) do |connection|
+                    Linux.withConnection(connection) do |linuxConnection|
+                        Podman.ensurePresent(linuxConnection, options)
+                        Podman.createUser(USER, HOME_DIR, 'llama.cpp', linuxConnection, options)
+                        linuxConnection.withUserShell(USER) do |shell|
+                            shell.createDirs(options, '~/models', '~/.cache')
+                        end
+
+                        path = Podman.containersPath(HOME_DIR)
+
+                        linuxConnection.fileWrite("#{path}/llama.cpp.env", '', options)
+                        if target['Model'].end_with?('.gguf')
+                            linuxConnection.fileAppend("#{path}/llama.cpp.env", 'LLAMA_ARG_MODEL=' + target['Model'], options)
+                        else
+                            linuxConnection.fileAppend("#{path}/llama.cpp.env", 'LLAMA_ARG_HF_REPO=' + target['Model'], options)
+                        end
+
+                        linuxConnection.setUserGroup("#{path}/llama.cpp.env", USER, USER, options)
+                        linuxConnection.setPrivate("#{path}/llama.cpp.env", options)
+
+                        linuxConnection.upload(__dir__ + '/llama.cpp.container', path, options)
+
+                        devices = linuxConnection.getGPUDevices(options)
+                        groups = linuxConnection.getGPUGroups(devices, options)
+                        linuxConnection.userAddGroups(USER, groups, options)
+
+                        image = nil
+                        if devices.include?('/dev/kfd')
+                            #image = 'full-rocm'
+                        elsif devices.include?('nvidia.com/gpu=all')
+                            image = 'full-cuda'
+                        end
+
+                        if image
+                            linuxConnection.fileReplace("#{path}/llama.cpp.container", ':full', ':' + image, options)
+                        end
+
+                        args = target['Args'].to_s
+                        linuxConnection.fileReplace("#{path}/llama.cpp.container", '\$ARGS', args, options)
+
+                        devicesString = ''
+                        if !devices.empty?
+                            devicesString = devices.map { |device| "AddDevice=#{device}" }.join('\n')
+                        end
+                        linuxConnection.fileReplace("#{path}/llama.cpp.container", '\$DEVICES', devicesString, { **options, escape: false })
+
+                        linuxConnection.reloadUserServices(USER, options)
+                        linuxConnection.restartUserService(USER, 'llama.cpp', options)
+                    end
+                end
+            end
+
+            def cleanup(configs, state, context, options)
+                cleanupType(:llamacpp, configs, state, context, options) do |item, id, state, context, options, connection|
+                    Linux.withConnection(connection) do |linuxConnection|
+
+                        linuxConnection.stopUserService(USER, 'llama.cpp', options)
+
+                        path = Podman.containersPath(HOME_DIR)
+                        linuxConnection.rm(path + '/llama.cpp.container', options[:dry])
+
+                        linuxConnection.reloadUserServices(USER, options)
+
+                        state.item(id)['Status'] = State::STATUS_DELETED unless options[:dry]
+
+                        if options[:destroy]
+                            linuxConnection.deleteUserAndGroup(USER, options)
+                            state.item(id)['Status'] = State::STATUS_DESTROYED unless options[:dry]
+                        end
+                    end
+                end
+            end
+
+        end
+    end
+end
+

data/Plugins/Apps/vLLM/vLLM.container

@@ -0,0 +1,32 @@
+
+[Unit]
+Description=vLLM container
+After=local-fs.target
+
+[Container]
+ContainerName=vLLM
+Image=$IMAGE
+Entrypoint=["python3", "-m", "vllm.entrypoints.openai.api_server"]
+Exec=$ARGS
+EnvironmentFile=/var/lib/vllm/.config/containers/systemd/vLLM.env
+PublishPort=127.0.0.1:18050:8000
+UserNS=keep-id:uid=1000,gid=1000
+# Need newer Podman to use this
+#GroupAdd=keep-groups
+PodmanArgs=--group-add keep-groups
+#PodmanArgs=--passwd-entry 'vllm:x:1000:1000:vLLM:/home/vllm:/bin/bash'
+Volume=/var/lib/vllm/.cache:/app/.cache
+Volume=/var/lib/vllm/.triton:/app/.triton
+Volume=/var/lib/vllm/.cache:/root/.cache
+Volume=/var/lib/vllm/.cache:/workspace/.cache
+ShmSize=1G
+$DEVICES
+LogDriver=journald
+AutoUpdate=registry
+
+[Service]
+TimeoutStartSec=12min
+Restart=on-failure
+
+[Install]
+WantedBy=multi-user.target default.target

data/Plugins/Apps/vLLM/vLLM.lmm.rb

@@ -0,0 +1,89 @@
+
+
+module ConfigLMM
+    module LMM
+        class VLLM < Framework::Plugin
+
+            USER = 'vllm'
+            HOME_DIR = '/var/lib/vllm'
+            PORT = 18050
+
+            CPU_IMAGE = 'public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v0.8.4'
+            CUDA_IMAGE = 'docker.io/vllm/vllm-openai:latest'
+            ROCM_IMAGE = 'docker.io/rocm/vllm:instinct_main'
+
+            def actionVLLMDeploy(id, target, activeState, context, options)
+                #raise Framework::PluginProcessError.new('Model field must be set!') unless target['Model']
+
+                self.withConnection(target['Location'], target) do |connection|
+                    Linux.withConnection(connection) do |linuxConnection|
+                        Podman.ensurePresent(linuxConnection, options)
+                        Podman.createUser(USER, HOME_DIR, 'vLLM', linuxConnection, options)
+                        linuxConnection.withUserShell(USER) do |shell|
+                            shell.createDirs(options, '~/.cache', '~/.triton')
+                        end
+
+                        path = Podman.containersPath(HOME_DIR)
+
+                        linuxConnection.fileWrite("#{path}/vLLM.env", '', options)
+
+                        linuxConnection.setUserGroup("#{path}/vLLM.env", USER, USER, options)
+                        linuxConnection.setPrivate("#{path}/vLLM.env", options)
+
+                        linuxConnection.upload(__dir__ + '/vLLM.container', path, options)
+
+                        devices = linuxConnection.getGPUDevices(options)
+                        groups = linuxConnection.getGPUGroups(devices, options)
+                        linuxConnection.userAddGroups(USER, groups, options)
+
+                        image = CPU_IMAGE
+                        if devices.include?('/dev/kfd')
+                            image = ROCM_IMAGE
+                        elsif devices.include?('nvidia.com/gpu=all')
+                            image = CUDA_IMAGE
+                        end
+                        linuxConnection.fileReplace("#{path}/vLLM.container", '\$IMAGE', image, options)
+
+                        args = target['Args'].to_s
+                        if target['Model']
+                            args += ' --model ' + target['Model'].to_s
+                        end
+                        linuxConnection.fileReplace("#{path}/vLLM.container", '\$ARGS', args, options)
+
+                        devicesString = ''
+                        if !devices.empty?
+                            devicesString = devices.map { |device| "AddDevice=#{device}" }.join('\n')
+                        end
+
+                        linuxConnection.fileReplace("#{path}/vLLM.container", '\$DEVICES', devicesString, { **options, escape: false })
+
+                        linuxConnection.reloadUserServices(USER, options)
+                        linuxConnection.restartUserService(USER, 'vLLM', options)
+                    end
+                end
+            end
+
+            def cleanup(configs, state, context, options)
+                cleanupType(:vLLM, configs, state, context, options) do |item, id, state, context, options, connection|
+                    Linux.withConnection(connection) do |linuxConnection|
+
+                        linuxConnection.stopUserService(USER, 'vLLM', options)
+
+                        path = Podman.containersPath(HOME_DIR)
+                        linuxConnection.rm(path + '/vLLM.container', options[:dry])
+
+                        linuxConnection.reloadUserServices(USER, options)
+
+                        state.item(id)['Status'] = State::STATUS_DELETED unless options[:dry]
+
+                        if options[:destroy]
+                            linuxConnection.deleteUserAndGroup(USER, options)
+                            state.item(id)['Status'] = State::STATUS_DESTROYED unless options[:dry]
+                        end
+                    end
+                end
+            end
+
+        end
+    end
+end

data/Plugins/OS/General/Utils.lmm.rb

@@ -0,0 +1,26 @@
+
+module ConfigLMM
+    module LMM
+        class Utils < Framework::Plugin
+
+            def actionFilesystemDeploy(id, target, activeState, context, options)
+                self.withConnection(target['Location'], target) do |connection|
+                    Linux.withConnection(connection) do |linuxConnection|
+                        if target['Directory']
+                            linuxConnection.createDirs(options, target['Directory'])
+                        end
+                        if target['Directories']
+                            linuxConnection.createDirs(options, *target['Directories'])
+                        end
+                        if target['Copy']
+                            target['Copy'].each do |source, target|
+                                linuxConnection.upload(source, target, options)
+                            end
+                        end
+                    end
+                end
+            end
+
+        end
+    end
+end