switch kokoro to remsky/Kokoro-FastAPI

apps/llama/configs/config.yaml

@@ -235,10 +235,9 @@ models:
       --parallel 1
       ${common_args}
 
-  # STT via whisper.cpp (Vulkan GPU on RX 580, always loaded, ~600MB VRAM)
+  # STT via whisper.cpp (CPU-only, always loaded)
-  # Model auto-downloaded by init container, see deployment.yaml
+  # Model auto-downloaded from HuggingFace on first start
-  # Note: Vulkan whisper on AMD GPUs has known quality issues on some cards;
+  # whisper-small: ~240MB RAM, good accuracy/speed tradeoff on R5 3600
-  # if transcriptions come out as garbage/gibberish, add --no-gpu to fall back.
   "whisper-small":
     checkEndpoint: none
     cmd: |
@@ -249,6 +248,7 @@ models:
         --inference-path /transcriptions
         --convert
         --threads 6
+        --no-gpu
 
 
   # Image generation via stable-diffusion.cpp (sd-server)

apps/llama/kokoro.yaml

@@ -21,7 +21,7 @@ spec:
           # OpenAI-compatible Kokoro-FastAPI TTS server, CPU PyTorch backend.
           # Models baked into the image (no PVC needed).
           # v0.3.0 includes fix for per-request voice tensor memory leak (#459).
-          image: ghcr.io/remsky/kokoro-fastapi-cpu:v0.3.0
+          image: ghcr.io/remsky/kokoro-fastapi-cpu:v0.3.0@sha256:d7df384acb57929f88a696d892c1a2e0c6b60fa934802ebec9a88626003b8e9a
           ports:
             - containerPort: 8880
               name: http