chore(docker): pin flash-attn prebuilt wheel to cp311 torch2.7 cu126; remove pip fallback

Dockerfile

@@ -5,7 +5,7 @@ FROM ${BASE_IMAGE}
 # Enable by default for Hugging Face Spaces GPU builds; override locally with
 #   --build-arg INSTALL_FLASH_ATTN=false
 ARG INSTALL_FLASH_ATTN=true
-ARG FLASH_ATTN_WHEEL_URL=https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.0.8/flash_attn-2.7.4.post1+cu126torch2.7-cp310-cp310-linux_x86_64.whl
+ARG FLASH_ATTN_WHEEL_URL=https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.0.8/flash_attn-2.7.4.post1+cu126torch2.7-cp311-cp311-linux_x86_64.whl
 
 # Persist caches and model storage in Spaces, and enable fast transfers
 ENV HF_HUB_ENABLE_HF_TRANSFER=1 \
@@ -73,8 +73,7 @@ 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"; \
       else \
-        echo "flash-attn wheel tag mismatch for $PYTAG. Trying pip install flash-attn (no-build-isolation)." && \
-        (pip install --no-cache-dir --no-build-isolation flash-attn || echo "flash-attn pip install failed; continuing without flash-attn"); \
+        echo "flash-attn wheel tag mismatch for $PYTAG. Skipping flash-attn install."; \
       fi; \
     else \
       echo "Skipping flash-attn installation"; \