Fix MIOpen workspace warnings and add kernel cache persistence

MIOPEN_FIND_ENFORCE=3 tells MIOpen to only select solvers that fit in
available workspace, eliminating the GemmFwdRest fallback warnings and
the associated performance hit. Persisting the MIOpen cache via a named
volume avoids kernel recompilation on every container start.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

docker-compose.yml

@@ -22,11 +22,17 @@ services:
       - ./voices:/app/voices
       - ./reference_audio:/app/reference_audio
       - hf_cache:/app/hf_cache
+      - miopen_cache:/app/miopen_cache
     environment:
       - HF_HUB_ENABLE_HF_TRANSFER=1
+      # MIOpen: only select solvers that fit in available workspace (stops
+      # the GemmFwdRest workspace warnings and the slow fallback path).
+      - MIOPEN_FIND_ENFORCE=3
+      # Persist MIOpen's kernel cache so it doesn't recompile on every start.
+      - MIOPEN_USER_DB_PATH=/app/miopen_cache
+      - MIOPEN_CUSTOM_CACHE_DIR=/app/miopen_cache
       # HSA_OVERRIDE_GFX_VERSION is only needed for GPUs not natively supported
-      # by the installed ROCm version. ROCm 7.2 supports RX 7000 series and newer
-      # natively. Older cards may still need an override:
+      # by ROCm 7.2. Older cards may still need an override:
       #   10.3.0 = RX 5000/6000 series
       #   9.0.6  = Vega
       # - HSA_OVERRIDE_GFX_VERSION=10.3.0
@@ -34,3 +40,4 @@ services:
 
 volumes:
   hf_cache:
+  miopen_cache: