diff --git a/Dockerfile b/Dockerfile
index 02835e7..4e224f0 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -90,7 +90,7 @@ RUN --mount=type=cache,id=repo-cache,target=/repo-cache \
         echo "Cache hit: Fetching flashinfer updates..." && \
         cd flashinfer && \
         git fetch origin && \
-        git fetch origin --tags && \
+        git fetch origin --tags --force && \
         (git checkout --detach origin/${FLASHINFER_REF} 2>/dev/null || git checkout ${FLASHINFER_REF}) && \
         git submodule update --init --recursive && \
         git clean -fdx && \
@@ -153,7 +153,7 @@ RUN --mount=type=cache,id=repo-cache,target=/repo-cache \
         echo "Cache hit: Fetching updates..." && \
         cd vllm && \
         git fetch origin && \
-        git fetch origin --tags && \
+        git fetch origin --tags --force && \
         (git checkout --detach origin/${VLLM_REF} 2>/dev/null || git checkout ${VLLM_REF}) && \
         git submodule update --init --recursive && \
         git clean -fdx && \
diff --git a/hf-download.sh b/hf-download.sh
index 4c4b4c2..c4c8d9f 100755
--- a/hf-download.sh
+++ b/hf-download.sh
@@ -1,7 +1,7 @@
 #!/bin/bash
 set -e
 
-HUB_PATH="$HOME/.cache/huggingface/hub"
+HUB_PATH="${HF_HOME:-$HOME/.cache/huggingface}/hub"
 
 # Default values
 COPY_HOSTS=()
@@ -42,7 +42,7 @@ copy_model_to_host() {
     local host_copy_start host_copy_end host_copy_time
     host_copy_start=$(date +%s)
     
-    if rsync -av --progress "$model_dir" "${SSH_USER}@${host}:$HUB_PATH/"; then
+    if rsync -av --mkpath --progress "$model_dir" "${SSH_USER}@${host}:$HUB_PATH/"; then
         host_copy_end=$(date +%s)
         host_copy_time=$((host_copy_end - host_copy_start))
         printf "Copy to %s completed in %02d:%02d:%02d\n" "$host" $((host_copy_time/3600)) $((host_copy_time%3600/60)) $((host_copy_time%60))
diff --git a/launch-cluster.sh b/launch-cluster.sh
index 2fe36dc..c5276f4 100755
--- a/launch-cluster.sh
+++ b/launch-cluster.sh
@@ -3,8 +3,9 @@
 # Default Configuration
 IMAGE_NAME="vllm-node"
 DEFAULT_CONTAINER_NAME="vllm_node"
+HF_CACHE_DIR="${HF_HOME:-$HOME/.cache/huggingface}"
 # Modify these if you want to pass additional docker args or set VLLM_SPARK_EXTRA_DOCKER_ARGS variable
-DOCKER_ARGS="-e NCCL_IGNORE_CPU_AFFINITY=1 -v $HOME/.cache/huggingface:/root/.cache/huggingface"
+DOCKER_ARGS="-e NCCL_IGNORE_CPU_AFFINITY=1 -v $HF_CACHE_DIR:/root/.cache/huggingface"
 
 # Append additional arguments from environment variable
 if [[ -n "$VLLM_SPARK_EXTRA_DOCKER_ARGS" ]]; then
diff --git a/mods/fix-qwen3-next-autoround/run.sh b/mods/fix-qwen3-next-autoround/run.sh
new file mode 100644
index 0000000..8e99055
--- /dev/null
+++ b/mods/fix-qwen3-next-autoround/run.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+set -e
+
+echo "Reverting PR #35156"
+if curl -L https://patch-diff.githubusercontent.com/raw/vllm-project/vllm/pull/35156.diff | patch -p1 -R -d /usr/local/lib/python3.12/dist-packages; then
+  echo "    OK"
+else
+  echo "    Patch can't be reversed, skipping"
+fi