diff --git a/docker/Dockerfile.ubuntu b/docker/Dockerfile.ubuntu
index 73ab8713..53d2abcf 100644
--- a/docker/Dockerfile.ubuntu
+++ b/docker/Dockerfile.ubuntu
@@ -69,7 +69,6 @@ RUN if [ "$IMAGE_TYPE" = "gpu" ]; then \
     pip install --no-cache-dir mcore-bridge -i https://pypi.org/simple/ -U && \
     pip install --no-cache-dir pybind11 liger_kernel wandb swanlab nvitop pre-commit "transformers<5.9" "trl<1.0" "peft<0.20" huggingface-hub -U && \
     pip install git+https://github.com/NVIDIA/TransformerEngine.git@stable --no-build-isolation; \
-    pip install "flash-attn-4[cu13]==4.0.0b15"; \
     pip install git+https://github.com/deepseek-ai/DeepGEMM.git@v2.1.1.post3 --no-build-isolation; \
     pip install -U flash-linear-attention --no-build-isolation; \
     pip install -U git+https://github.com/Dao-AILab/causal-conv1d --no-build-isolation; \
diff --git a/docker/build_image.py b/docker/build_image.py
index fc525bac..5f1f6c55 100644
--- a/docker/build_image.py
+++ b/docker/build_image.py
@@ -405,7 +405,7 @@ class LatestGPUImageBuilder(StableGPUImageBuilder):
             args.torchaudio_version = '2.11.0'
             args.torchvision_version = '0.26.0'
         if not args.vllm_version:
-            args.vllm_version = '0.22.0'
+            args.vllm_version = '0.21.0'
         if not args.base_image:
             # A mirrored image of nvidia/cuda:12.4.0-devel-ubuntu22.04
             args.base_image = 'nvidia/cuda:13.0.3-cudnn-devel-ubuntu22.04'