Disable MIOpen GEMM solver to fix null workspace fallback

PyTorch passes ptr=0 size=0 workspace to MIOpen convolutions, causing
GemmFwdRest to warn and fall back to a slow path on every operation.
MIOPEN_DEBUG_CONV_GEMM=0 skips GEMM entirely and uses Direct/Winograd
solvers which have no workspace requirement.

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

docker-compose.yml

@@ -25,6 +25,11 @@ services:
       - miopen_cache:/app/miopen_cache
     environment:
       - HF_HUB_ENABLE_HF_TRANSFER=1
+      # Disable MIOpen's GEMM convolution solver. PyTorch passes a null workspace
+      # pointer (ptr=0, size=0) which causes GemmFwdRest to warn and fall back on
+      # every conv op. Disabling it makes MIOpen go straight to Direct/Winograd
+      # solvers that don't require workspace allocation.
+      - MIOPEN_DEBUG_CONV_GEMM=0
       # 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