Switch transcription comparison to Voxtral

deploy/ai-server/docker-compose.audio.yml

@@ -103,15 +103,15 @@ services:
       - "--port"
       - "8000"
       - "--max-model-len"
-      - "32768"
+      - "16384"
       - "--gpu-memory-utilization"
-      - "0.62"
+      - "0.55"
       - "--api-key"
       - "${VLLM_API_KEY}"
       - "--max-num-seqs"
-      - "2"
+      - "1"
       - "--max-num-batched-tokens"
-      - "8192"
+      - "4096"
     healthcheck:
       test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health')"]
       interval: 30s

k8s/configmap.yaml

@@ -11,16 +11,16 @@ data:
   LLM_BASE_URL: "http://10.2.3.5:8002"
   LLM_MODEL: "qwen2.5-14b"
   LLM_TIMEOUT: "5m"
-  TRANSCRIPTION_PROVIDERS: "whisperx,qwen2-audio,voxtral-small"
+  TRANSCRIPTION_PROVIDERS: "voxtral-small"
   WHISPERX_URL: "http://10.2.3.5:8001"
   WHISPERX_TIMEOUT: "10m"
   WHISPERX_LEAD_SILENCE: "800ms"
   # Fill these after Qwen2-Audio and Voxtral are exposed as OpenAI-compatible
   # chat-completions endpoints on the AI server.
-  QWEN_AUDIO_BASE_URL: "http://10.2.3.5:8003"
+  QWEN_AUDIO_BASE_URL: ""
   QWEN_AUDIO_MODEL: "Qwen/Qwen2-Audio-7B-Instruct"
   QWEN_AUDIO_TIMEOUT: "10m"
-  VOXTRAL_BASE_URL: ""
+  VOXTRAL_BASE_URL: "http://10.2.3.5:8004"
   VOXTRAL_MODEL: "mistralai/Voxtral-Small-24B-2507"
   VOXTRAL_TIMEOUT: "10m"
   AUDIO_LLM_MAX_TOKENS: "4096"