diff --git a/scripts/uninstall.ps1 b/scripts/uninstall.ps1
index 020fb85781..34e60fbc93 100644
--- a/scripts/uninstall.ps1
+++ b/scripts/uninstall.ps1
@@ -474,12 +474,13 @@ function Uninstall-UnslothStudio {
             # Probe candidates by exit code ('' = default distro) since `wsl --list` emits UTF-16 PS
             # mis-parses. rm runs FIRST (the kills could SIGKILL this shell) and drops the dangling
             # /root/.local/bin/unsloth symlink. Scope STRICTLY to /root (the fallback's install dir);
-            # /home/*/.unsloth may be another user's. The 8888 kill is gated on an Unsloth install
-            # existing (checked BEFORE rm deletes the marker) so an unrelated listener survives. pkill
-            # matches argv containing /root/.unsloth/ (not bare names that would hit a user's own
-            # llama-server); the backslash + [h]-bracket in '/root/\.unslot[h]/' keep it from matching
-            # this command's own argv.
-            $_clean = '_had=0; if [ -d /root/.unsloth ] || [ -L /root/.local/bin/unsloth ]; then _had=1; fi; rm -rf /root/.unsloth /root/llama-cuda /root/provision_llama_cuda.sh /root/llama_cuda_build.log 2>/dev/null; rm -f /root/.local/bin/unsloth 2>/dev/null; if [ $_had -eq 1 ]; then fuser -k 8888/tcp 2>/dev/null; fi; pkill -9 -f ''/root/\.unslot[h]/'' 2>/dev/null; true'
+            # /home/*/.unsloth may be another user's. The 8888 kill only targets a listener whose
+            # process cmdline is under /root/.unsloth (Studio's bind), so an unrelated service on 8888
+            # -- Jupyter et al. default to it -- is NOT killed; it's also gated on an Unsloth install
+            # having existed (checked BEFORE rm deletes the marker). pkill matches argv containing
+            # /root/.unsloth/ (not bare names that would hit a user's own llama-server); the backslash
+            # + [h]-bracket in '/root/\.unslot[h]/' keep it from matching this command's own argv.
+            $_clean = '_had=0; if [ -d /root/.unsloth ] || [ -L /root/.local/bin/unsloth ]; then _had=1; fi; rm -rf /root/.unsloth /root/llama-cuda /root/provision_llama_cuda.sh /root/llama_cuda_build.log 2>/dev/null; rm -f /root/.local/bin/unsloth 2>/dev/null; if [ $_had -eq 1 ]; then for _p in $(fuser 8888/tcp 2>/dev/null); do grep -qa /root/\.unsloth/ /proc/$_p/cmdline 2>/dev/null && kill -9 $_p 2>/dev/null; done; fi; pkill -9 -f ''/root/\.unslot[h]/'' 2>/dev/null; true'
             # Clean only distros with evidence of a fallback install: the wsl-distro.txt marker or an
             # explicit UNSLOTH_WSL_DISTRO. The broad candidate probe is only for legacy marker-less
             # installs (ARM64 only); on x86 it would delete distros this installer never touched
diff --git a/studio/scripts/provision_llama_cuda.sh b/studio/scripts/provision_llama_cuda.sh
index 29ffc0052e..43779c0d10 100644
--- a/studio/scripts/provision_llama_cuda.sh
+++ b/studio/scripts/provision_llama_cuda.sh
@@ -172,6 +172,30 @@ if [ ! -d "$LLAMA_DIR/.git" ]; then
 fi
 cd "$LLAMA_DIR" || { _restore_prev; exit 0; }
 
+# When rebuilding in-place over an existing git checkout, the whole-dir backup above
+# was skipped (_LLAMA_BAK empty) -- but build/ may already hold a working (e.g. CPU)
+# llama-server from a prior setup.sh source build. The wipe-on-failure paths below
+# would destroy it with nothing to restore, leaving NO server despite the "keeps the
+# existing server" promise (a thermal shutdown mid-CUDA-build is a real failure mode
+# here). Back up the existing binaries so a failed rebuild can put them back. Only
+# bin/ (server + dlopen-ed backends) is needed; cheap since any pre-existing server
+# here is the non-CUDA fallback (a CUDA one would have exited at step 0).
+_BUILD_BAK=""
+if [ -z "$_LLAMA_BAK" ] && [ -x "$SERVER" ]; then
+    _BUILD_BAK="${LLAMA_DIR}.binbak.$$"
+    rm -rf "$_BUILD_BAK" 2>/dev/null
+    cp -a "$LLAMA_DIR/build/bin" "$_BUILD_BAK" 2>/dev/null || _BUILD_BAK=""
+fi
+_restore_build() {
+    if [ -n "$_BUILD_BAK" ] && [ -e "$_BUILD_BAK" ] && [ ! -x "$SERVER" ]; then
+        mkdir -p "$LLAMA_DIR/build" 2>/dev/null
+        rm -rf "$LLAMA_DIR/build/bin" 2>/dev/null
+        mv "$_BUILD_BAK" "$LLAMA_DIR/build/bin" 2>/dev/null && log "restored previous llama-server (rebuild failed)"
+    fi
+    [ -n "$_BUILD_BAK" ] && rm -rf "$_BUILD_BAK" 2>/dev/null
+    _BUILD_BAK=""
+}
+
 log "building CUDA llama.cpp (arch=$CUDA_ARCH, host=$HCXX) - this takes a few minutes..."
 _cmake_configure() {
     cmake -S . -B build -DCMAKE_BUILD_TYPE=Release \
@@ -185,7 +209,7 @@ _cmake_configure() {
 if ! _cmake_configure; then
     log "stale/incompatible CMake cache detected; wiping build dir for a clean CUDA configure"
     rm -rf build
-    _cmake_configure || { log "cmake configure failed"; cd /; _restore_prev; exit 0; }
+    _cmake_configure || { log "cmake configure failed"; cd /; _restore_build; _restore_prev; exit 0; }
 fi
 # Also builds the targets unsloth-zoo's GGUF exporter needs (llama-mtmd-cli,
 # llama-gguf-split). Jobs default to ~half the cores (full -j(nproc) CUDA builds
@@ -226,10 +250,13 @@ if ! _cmake_build; then
     # (undefined ggml_cuda_op_* refs); wipe and rebuild clean.
     log "build failed (likely interrupted/partial); wiping build dir and rebuilding clean"
     rm -rf build
-    _cmake_configure || { log "cmake configure failed"; cd /; _restore_prev; exit 0; }
-    _cmake_build || { log "cmake build failed"; cd /; _restore_prev; exit 0; }
+    _cmake_configure || { log "cmake configure failed"; cd /; _restore_build; _restore_prev; exit 0; }
+    _cmake_build || { log "cmake build failed"; cd /; _restore_build; _restore_prev; exit 0; }
 fi
 _cmake_build_extras
+# Drop the backup on a successful build, or restore the prior server if the rebuild
+# yielded none (idempotent; only restores when $SERVER is missing).
+_restore_build
 
 if is_cuda_server "$SERVER"; then
     log "CUDA llama-server ready: $SERVER"
diff --git a/studio/setup.sh b/studio/setup.sh
index 87e99acf53..bf22bea51e 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1253,11 +1253,15 @@ if [ "$_NEED_LLAMA_SOURCE_BUILD" = true ] && \
 fi
 
 # ── WSL2 aarch64 + NVIDIA, no nvcc yet: defer to the background CUDA build ──
-# install.ps1 builds the CUDA llama-server in the background; without nvcc,
-# section 9 could only make a slow CPU server that build discards. With nvcc we
-# fall through to section 9; opted out (UNSLOTH_NO_LLAMA_CUDA=1) the CPU build is
-# the only server.
+# install.ps1 builds the CUDA llama-server in the background AND signals that by
+# exporting UNSLOTH_WSL_LLAMA_DEFERRED=1 into this install. ONLY defer when that
+# flag is set: a direct in-WSL `unsloth studio update` has no background builder,
+# so deferring there would report "CUDA build running in background" while nothing
+# builds -- hiding a no-server state. Without the flag we fall through to section 9
+# (a slow CPU server, still better than a phantom background build). With nvcc we
+# fall through too; opted out (UNSLOTH_NO_LLAMA_CUDA=1) the CPU build is the only server.
 if [ "$_NEED_LLAMA_SOURCE_BUILD" = true ] \
+        && [ "${UNSLOTH_WSL_LLAMA_DEFERRED:-0}" = "1" ] \
         && [ "$_LLAMA_FORCE_COMPILE" != "1" ] \
         && [ -z "$_LLAMA_PR" ] \
         && [ "${UNSLOTH_NO_LLAMA_CUDA:-0}" != "1" ] \