diff --git a/docker/Dockerfile.ubuntu b/docker/Dockerfile.ubuntu
index fd51af0e..04ee893d 100644
--- a/docker/Dockerfile.ubuntu
+++ b/docker/Dockerfile.ubuntu
@@ -55,7 +55,6 @@ ARG CUR_TIME={cur_time}
 RUN echo $CUR_TIME
 
 RUN sh /tmp/install.sh {version_args} && \
-    pip install --no-cache-dir xformers==0.0.27 && \
     curl -fsSL https://ollama.com/install.sh | sh && \
     pip install --no-cache-dir -U funasr scikit-learn && \
     pip install --no-cache-dir -U qwen_vl_utils qwen_omni_utils pyav librosa timm transformers accelerate peft trl safetensors && \
diff --git a/docker/build_image.py b/docker/build_image.py
index 268ba658..c77418da 100644
--- a/docker/build_image.py
+++ b/docker/build_image.py
@@ -31,7 +31,7 @@ class Builder:
         if not args.cuda_version:
             args.cuda_version = '12.1.0'
         if not args.vllm_version:
-            args.vllm_version = '0.7.2'
+            args.vllm_version = '0.5.3'
         if not args.lmdeploy_version:
             args.lmdeploy_version = '0.6.2'
         if not args.autogptq_version:
@@ -267,15 +267,15 @@ class LLMImageBuilder(Builder):
             # A mirrored image of nvidia/cuda:12.4.0-devel-ubuntu22.04
             args.base_image = 'nvidia/cuda:12.4.0-devel-ubuntu22.04'
         if not args.torch_version:
-            args.torch_version = '2.5.1'
-            args.torchaudio_version = '2.5.1'
-            args.torchvision_version = '0.20.1'
+            args.torch_version = '2.6.0'
+            args.torchaudio_version = '2.6.0'
+            args.torchvision_version = '0.21.0'
         if not args.cuda_version:
             args.cuda_version = '12.4.0'
         if not args.vllm_version:
-            args.vllm_version = '0.7.2'
+            args.vllm_version = '0.8.3'
         if not args.lmdeploy_version:
-            args.lmdeploy_version = '0.7.0.post2'
+            args.lmdeploy_version = '0.7.2.post1'
         if not args.autogptq_version:
             args.autogptq_version = '0.7.1'
         if not args.flashattn_version:
@@ -342,8 +342,6 @@ class SwiftImageBuilder(LLMImageBuilder):
             args.torch_version = '2.5.1'
             args.torchaudio_version = '2.5.1'
             args.torchvision_version = '0.20.1'
-        if not args.cuda_version:
-            args.cuda_version = '12.4.0'
         if not args.vllm_version:
             args.vllm_version = '0.7.3'
         return super().init_args(args)