Remove tgi

helm/h2ogpt-chart/README.md

@@ -135,29 +135,6 @@ A Helm chart for h2oGPT
 | h2ogpt.updateStrategy.type | string | `"RollingUpdate"` |  |
 | nameOverride | string | `""` |  |
 | namespaceOverride | string | `""` |  |
-| tgi.containerArgs | string | `nil` |  |
-| tgi.enabled | bool | `false` | Enable tgi |
-| tgi.env | object | `{}` |  |
-| tgi.hfSecret | string | `nil` |  |
-| tgi.image.pullPolicy | string | `"IfNotPresent"` |  |
-| tgi.image.repository | string | `"ghcr.io/huggingface/text-generation-inference"` |  |
-| tgi.image.tag | string | `"0.9.3"` |  |
-| tgi.nodeSelector | string | `nil` |  |
-| tgi.overrideConfig | string | `nil` |  |
-| tgi.podAffinity | string | `nil` | Set hostname and zone to true for pod affinity rules based on hostname and zone. |
-| tgi.podAnnotations | object | `{}` |  |
-| tgi.podLabels | object | `{}` |  |
-| tgi.podSecurityContext | string | `nil` |  |
-| tgi.replicaCount | int | `1` |  |
-| tgi.resources | string | `nil` |  |
-| tgi.securityContext | string | `nil` |  |
-| tgi.service.port | int | `8080` |  |
-| tgi.service.type | string | `"ClusterIP"` |  |
-| tgi.storage.class | string | `nil` |  |
-| tgi.storage.size | string | `"512Gi"` |  |
-| tgi.storage.useEphemeral | bool | `true` |  |
-| tgi.tolerations | string | `nil` |  |
-| tgi.updateStrategy.type | string | `"RollingUpdate"` |  |
 | vllm.containerArgs[0] | string | `"--model"` |  |
 | vllm.containerArgs[1] | string | `"h2oai/h2ogpt-4096-llama2-7b-chat"` |  |
 | vllm.containerArgs[2] | string | `"--tokenizer"` |  |

helm/h2ogpt-chart/templates/h2ogpt-deployment.yaml

@@ -87,18 +87,7 @@ spec:
 
               python3 /workspace/generate.py
           {{- end }}
-          {{- if and .Values.tgi.enabled (not .Values.global.externalLLM.modelLock) }}
-          args:
-            - >
-              until wget -O- http://{{ include "h2ogpt.fullname" . }}-tgi-inference:{{ .Values.tgi.service.port }}/ >/dev/null 2>&1;
-                do
-                  echo "Waiting for inference service to become ready...";
-                  sleep 5;
-                done
-
-              python3 /workspace/generate.py
-          {{- end }}
-          {{- if and .Values.h2ogpt.enabled (not (or .Values.vllm.enabled .Values.tgi.enabled )) }}
+          {{- if and .Values.h2ogpt.enabled (not .Values.vllm.enabled ) }}
           args:
             - >
               python3 /workspace/generate.py
@@ -146,10 +135,6 @@ spec:
                 name: {{ include "h2ogpt.fullname" . }}-external-llm-secret
           {{- end }}
           env:
-          {{- if and .Values.tgi.enabled (not .Values.global.externalLLM.enabled) }}
-            - name: h2ogpt_inference_server
-              value: "http://{{ include "h2ogpt.fullname" . }}-tgi-inference:{{ .Values.tgi.service.port }}"
-          {{- end }}
           {{- if and .Values.vllm.enabled (not .Values.global.externalLLM.enabled) }}
             - name: h2ogpt_inference_server
               value: "vllm:{{ include "h2ogpt.fullname" . }}-vllm-inference:{{ .Values.vllm.service.port }}"

helm/h2ogpt-chart/templates/validations.yaml → helm/h2ogpt-chart/templates/validators.yaml

@@ -1,6 +1,3 @@
-{{- if and .Values.vllm.enabled .Values.tgi.enabled }}
-  {{- fail "Both TGI and vLLM cannot be enabled at the same time. Enable only one and try again" }}
-{{- end }}
 {{- if and ( and .Values.h2ogpt.enabled .Values.h2ogpt.agents.enabled) .Values.agents.enabled }}
   {{- fail " Both agents in both h2ogpt.agents cannot be enabled. Enably only one and try again" }}
 {{- end }}

helm/h2ogpt-chart/values.yaml

@@ -140,7 +140,7 @@ agents:
   # -- Enable agents, this must be `false` if `h2ogpt.agents.enabled` is `true`
   enabled: false
   autoscaling:
-    # Enable autoscaling for agents
+    # Enable autoscaling (HPA) for agents
     enabled: false
     minReplicas: 1
     maxReplicas: 2
@@ -239,48 +239,6 @@ agents:
   podAnnotations: {}
   podLabels: {}
 
-tgi:
-  # -- Enable tgi
-  enabled: false
-  replicaCount: 1
-
-  image:
-    repository: ghcr.io/huggingface/text-generation-inference
-    tag: 0.9.3
-    pullPolicy: IfNotPresent
-  # -- Set hostname and zone to true for pod affinity rules based on hostname and zone.
-  podAffinity:
-    # hostname:
-    # zone:
-
-  storage:
-    size: 512Gi
-    class: 
-    useEphemeral: true
-
-  overrideConfig:
-  hfSecret:
-  containerArgs:
-
-  service:
-    type: ClusterIP
-    port: 8080
-
-  updateStrategy:
-    type: RollingUpdate
-
-  podSecurityContext:
-  securityContext:
-
-  resources:
-  nodeSelector:
-  tolerations:
-
-  env: {}
-
-  podAnnotations: {}
-  podLabels: {}
-
 vllm:
   # -- Enable vllm
   enabled: false