Docker: guard flash-attn install by Python ABI tag; skip on mismatch

Dockerfile

@@ -49,9 +49,16 @@ COPY --chown=user requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 
 # Optionally install flash-attn wheel (requires Python/torch/CUDA compatibility)
+# Will auto-skip if the wheel's Python tag does not match this image's Python.
 RUN if [ "$INSTALL_FLASH_ATTN" = "true" ]; then \
-      echo "Installing flash-attn from $FLASH_ATTN_WHEEL_URL" && \
-      pip install --no-cache-dir $FLASH_ATTN_WHEEL_URL; \
+      PYTAG=$(python -c "import sys; print(f'cp{sys.version_info.major}{sys.version_info.minor}')"); \
+      echo "Detected Python tag: $PYTAG"; \
+      if echo "$FLASH_ATTN_WHEEL_URL" | grep -q "$PYTAG"; then \
+        echo "Installing flash-attn from $FLASH_ATTN_WHEEL_URL" && \
+        pip install --no-cache-dir "$FLASH_ATTN_WHEEL_URL"; \
+      else \
+        echo "flash-attn wheel tag mismatch for $PYTAG. Skipping flash-attn install. Override FLASH_ATTN_WHEEL_URL if needed."; \
+      fi; \
     else \
       echo "Skipping flash-attn installation"; \
     fi