audioqna: reduce microservice numbers

helm-charts/audioqna/Chart.yaml

@@ -6,24 +6,15 @@ name: audioqna
 description: The Helm chart to deploy AudioQnA
 type: application
 dependencies:
-  - name: asr
-    version: 0-latest
-    repository: "file://../common/asr"
   - name: whisper
     version: 0-latest
     repository: "file://../common/whisper"
-  - name: tts
-    version: 0-latest
-    repository: "file://../common/tts"
   - name: speecht5
     version: 0-latest
     repository: "file://../common/speecht5"
   - name: tgi
     version: 0-latest
     repository: "file://../common/tgi"
-  - name: llm-uservice
-    version: 0-latest
-    repository: "file://../common/llm-uservice"
   - name: ui
     version: 0-latest
     repository: "file://../common/ui"

helm-charts/audioqna/gaudi-values.yaml

@@ -12,6 +12,12 @@ tgi:
   MAX_INPUT_LENGTH: "1024"
   MAX_TOTAL_TOKENS: "2048"
   CUDA_GRAPHS: ""
+  HF_HUB_DISABLE_PROGRESS_BARS: 1
+  HF_HUB_ENABLE_HF_TRANSFER: 0
+  ENABLE_HPU_GRAPH: true
+  LIMIT_HPU_GRAPH: true
+  USE_FLASH_ATTENTION: true
+  FLASH_ATTENTION_RECOMPUTE: true
   livenessProbe:
     initialDelaySeconds: 5
     periodSeconds: 5

helm-charts/audioqna/templates/deployment.yaml

@@ -34,18 +34,18 @@ spec:
       containers:
         - name: {{ .Release.Name }}
           env:
-            - name: LLM_SERVICE_HOST_IP
-              value: {{ include "llm-uservice.fullname" (index .Subcharts "llm-uservice") }}
-            - name: LLM_SERVICE_PORT
-              value: {{ index .Values "llm-uservice" "service" "port" | quote }}
-            - name: ASR_SERVICE_HOST_IP
-              value: {{ include "asr.fullname" (index .Subcharts "asr") }}
-            - name: ASR_SERVICE_PORT
-              value: {{ index .Values "asr" "service" "port" | quote }}
-            - name: TTS_SERVICE_HOST_IP
-              value: {{ include "tts.fullname" (index .Subcharts "tts") }}
-            - name: TTS_SERVICE_PORT
-              value: {{ index .Values "tts" "service" "port" | quote }}
+            - name: LLM_SERVER_HOST_IP
+              value: {{ include "tgi.fullname" (index .Subcharts "tgi") }}
+            - name: LLM_SERVER_PORT
+              value: "80"
+            - name: WHISPER_SERVER_HOST_IP
+              value: {{ include "whisper.fullname" (index .Subcharts "whisper") }}
+            - name: WHISPER_SERVER_PORT
+              value: {{ index .Values "whisper" "service" "port" | quote }}
+            - name: SPEECHT5_SERVER_HOST_IP
+              value: {{ include "speecht5.fullname" (index .Subcharts "speecht5") }}
+            - name: SPEECHT5_SERVER_PORT
+              value: {{ index .Values "speecht5" "service" "port" | quote }}
           securityContext:
             {{- toYaml .Values.securityContext | nindent 12 }}
           image: "{{ .Values.image.repository }}:{{ .Values.image.tag | default .Chart.AppVersion }}"

helm-charts/common/speecht5/values.yaml

@@ -70,20 +70,20 @@ resources: {}
 
 livenessProbe:
   httpGet:
-    path: /v1/health
+    path: /health
     port: speecht5
   initialDelaySeconds: 5
   periodSeconds: 5
   failureThreshold: 24
 readinessProbe:
   httpGet:
-    path: /v1/health
+    path: /health
     port: speecht5
   initialDelaySeconds: 5
   periodSeconds: 5
 startupProbe:
   httpGet:
-    path: /v1/health
+    path: /health
     port: speecht5
   initialDelaySeconds: 5
   periodSeconds: 5

helm-charts/common/whisper/values.yaml

@@ -69,20 +69,20 @@ resources: {}
 
 livenessProbe:
   httpGet:
-    path: /v1/health
+    path: /health
     port: whisper
   initialDelaySeconds: 5
   periodSeconds: 5
   failureThreshold: 24
 readinessProbe:
   httpGet:
-    path: /v1/health
+    path: /health
     port: whisper
   initialDelaySeconds: 5
   periodSeconds: 5
 startupProbe:
   httpGet:
-    path: /v1/health
+    path: /health
     port: whisper
   initialDelaySeconds: 5
   periodSeconds: 5