Revert to torch 2.5.1 + ROCm 6.1 (known working combination)

PyTorch 2.11.0 with ROCm 7.2 wheels against rocm/dev-ubuntu-22.04:latest
causes MIOpen version mismatches that force every convolution onto a slow
zero-workspace fallback path (41s synthesis). The existing working project
uses torch 2.5.1 + ROCm 6.1 successfully on the same base image.

Also remove MIOPEN_FIND_ENFORCE override - unnecessary with matched versions.

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

docker-compose.yml

@@ -25,15 +25,13 @@ services:
       - miopen_cache:/app/miopen_cache
     environment:
       - HF_HUB_ENABLE_HF_TRANSFER=1
-      # MIOpen: search for best solver once and cache (1=DB_UPDATE).
-      # Never use 3 (SEARCH_DB_UPDATE) in production - it benchmarks on every call.
-      - MIOPEN_FIND_ENFORCE=1
       # 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 ROCm 7.2. Older cards may still need an override:
+      # HSA_OVERRIDE_GFX_VERSION is only needed for GPUs not officially supported
+      # by the ROCm version in the base image. Common values:
       #   10.3.0 = RX 5000/6000 series
+      #   11.0.0 = RX 7000 series
       #   9.0.6  = Vega
       # - HSA_OVERRIDE_GFX_VERSION=10.3.0
       # - HF_TOKEN=your_token_here

requirements-rocm-init.txt

@@ -1,3 +1,3 @@
---index-url https://download.pytorch.org/whl/rocm7.2
-torch==2.11.0
-torchaudio==2.11.0
+--index-url https://download.pytorch.org/whl/rocm6.1
+torch==2.5.1
+torchaudio==2.5.1