diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 97962311fc..705ba708d2 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -7092,7 +7092,16 @@ def main() -> int:
         return EXIT_SUCCESS
 
     if args.smoke_test is not None:
-        host = detect_host()
+        # Honor the same host overrides as the install path so a caller that
+        # forwards --has-rocm / --rocm-gfx (setup.sh/setup.ps1 do, because the
+        # installer's own probe can miss amd-smi-only hosts) is held to the GPU
+        # offload check instead of silently resolving a CPU kind.
+        host = _apply_host_overrides(
+            detect_host(),
+            override_has_rocm = args.has_rocm,
+            override_rocm_gfx = args.rocm_gfx,
+            force_cpu = args.cpu_fallback,
+        )
         try:
             resolved_kind = smoke_test_server_binary(
                 args.smoke_test,
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index b52b02a8d2..3a4e9d4ca3 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -3051,7 +3051,10 @@ if (-not $NeedLlamaSourceBuild) {
         if (Test-Path -LiteralPath $builtServer) {
             Write-Host ""
             Write-Host "--- GPU smoke test ---" -ForegroundColor Cyan
-            & python "$PSScriptRoot\install_llama_prebuilt.py" --smoke-test "$builtServer" --install-dir "$LlamaCppDir" 2>&1 | Out-String | Write-Host
+            # $LlamaCudaBuild gates this block, so the build is CUDA; pass the
+            # explicit kind so the installer's own probe cannot resolve a CPU
+            # kind and skip the offload gate.
+            & python "$PSScriptRoot\install_llama_prebuilt.py" --smoke-test "$builtServer" --install-dir "$LlamaCppDir" --install-kind "windows-cuda" 2>&1 | Out-String | Write-Host
             $smokeExit = $LASTEXITCODE
             if ($smokeExit -eq 2) {
                 substep "GPU build runs on CPU only (GPU backend failed to initialize)" "Yellow"
diff --git a/studio/setup.sh b/studio/setup.sh
index b941fd6a82..10aa831b72 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -1362,11 +1362,23 @@ else
         # _gpu_fallback_label is empty for a pure CPU build (nothing to verify).
         if [ "$BUILD_OK" = true ]; then
             _SMOKE_LABEL="$(_gpu_fallback_label)"
-            if [ -n "$_SMOKE_LABEL" ] && [ -f "$_BUILD_TMP/build/bin/llama-server" ]; then
+            # Pass the GPU kind setup resolved, not the installer's own probe:
+            # on amd-smi-only / name-inferred ROCm hosts the child probe can miss
+            # the GPU and resolve a CPU kind, which would skip the offload gate.
+            _SMOKE_KIND=""
+            if [ "$_TRY_METAL_CPU_FALLBACK" = true ]; then
+                _SMOKE_KIND="macos-arm64"
+            elif [ "$GPU_BACKEND" = "cuda" ]; then
+                _SMOKE_KIND="linux-cuda"
+            elif [ "$GPU_BACKEND" = "rocm" ]; then
+                _SMOKE_KIND="linux-rocm"
+            fi
+            if [ -n "$_SMOKE_LABEL" ] && [ -n "$_SMOKE_KIND" ] && [ -f "$_BUILD_TMP/build/bin/llama-server" ]; then
                 # if/else keeps set -e from aborting before we read the code.
                 if python "$SCRIPT_DIR/install_llama_prebuilt.py" \
                         --smoke-test "$_BUILD_TMP/build/bin/llama-server" \
-                        --install-dir "$_BUILD_TMP" > "$_BUILD_TMP/gpu-smoke.log" 2>&1; then
+                        --install-dir "$_BUILD_TMP" \
+                        --install-kind "$_SMOKE_KIND" > "$_BUILD_TMP/gpu-smoke.log" 2>&1; then
                     _SMOKE_RC=0
                 else
                     _SMOKE_RC=$?