Studio: gate ROCm hosts out of CPU prebuilt and require offload signal in smoke-test

- direct_upstream_release_plan: Linux x86_64 and arm64 CPU branches now also
  require not host.has_rocm, so a ROCm host whose HIP prebuilt is missing or
  rejected never silently takes a CPU upstream tarball as a 'success'.
- setup.ps1: a Windows ROCm host with no usable HIP prebuilt and no HIP source
  build path is now marked LlamaCppDegraded so the CPU last-resort installs a
  clearly labelled CPU build instead of a 'built' install that runs on CPU.
- validate_server gains require_gpu_signal; smoke_test_server_binary sets it for
  GPU install kinds so the '0 = offload confirmed' CLI contract treats a
  no-GPU-signal log as inconclusive (EXIT_ERROR) rather than a silent pass.
- run_smoke_spoof + unit test cover the new no-signal-is-inconclusive contract.

studio/install_llama_prebuilt.py

@@ -1611,7 +1611,12 @@ def direct_upstream_release_plan(
                     install_kind = "macos-x64",
                 )
             )
-    elif host.is_linux and host.is_x86_64 and not host.has_usable_nvidia:
+    elif (
+        host.is_linux
+        and host.is_x86_64
+        and not host.has_usable_nvidia
+        and not host.has_rocm
+    ):
         asset_name = f"llama-{release_tag}-bin-ubuntu-x64.tar.gz"
         asset_url = assets.get(asset_name)
         if asset_url:
@@ -1625,7 +1630,12 @@ def direct_upstream_release_plan(
                     install_kind = "linux-cpu",
                 )
             )
-    elif host.is_linux and host.is_arm64 and not host.has_usable_nvidia:
+    elif (
+        host.is_linux
+        and host.is_arm64
+        and not host.has_usable_nvidia
+        and not host.has_rocm
+    ):
         # Upstream ggml-org/llama.cpp ships llama-bNNNN-bin-ubuntu-arm64.tar.gz
         # (visible in the b9334 release manifest). Without this branch the
         # selector returned 0 attempts and the installer fell back to a
@@ -5766,6 +5776,7 @@ def validate_server(
     *,
     runtime_line: str | None = None,
     install_kind: str | None = None,
+    require_gpu_signal: bool = False,
 ) -> None:
     last_failure: PrebuiltFallback | None = None
     for port_attempt in range(1, SERVER_PORT_BIND_ATTEMPTS + 1):
@@ -5908,10 +5919,10 @@ def validate_server(
                     # the next bundle / source build instead of stopping here.
                     if _enable_gpu_layers:
                         log_handle.flush()
-                        if (
-                            server_log_shows_gpu_offload(read_full_log(log_path))
-                            is False
-                        ):
+                        offload = server_log_shows_gpu_offload(
+                            read_full_log(log_path)
+                        )
+                        if offload is False:
                             raise GpuOffloadFailure(
                                 "llama-server served a completion but loaded the "
                                 "model entirely on CPU despite GPU offload being "
@@ -5924,6 +5935,17 @@ def validate_server(
                                 "(unslothai/unsloth#5807, #5106):\n"
                                 + read_log_excerpt(log_path)
                             )
+                        # No GPU signal: install validation stays conservative
+                        # (never reject on no evidence), but --smoke-test sets
+                        # require_gpu_signal so its "0 = offload confirmed"
+                        # contract does not pass an unproven log.
+                        if offload is None and require_gpu_signal:
+                            raise PrebuiltFallback(
+                                "llama-server served a completion but its startup "
+                                "log carried no GPU-offload signal; smoke-test "
+                                "result is inconclusive:\n"
+                                + read_log_excerpt(log_path)
+                            )
                     return
         finally:
             if process is not None and process.poll() is None:
@@ -6925,6 +6947,9 @@ def smoke_test_server_binary(
         Path(install_dir).expanduser().resolve() if install_dir else server_path.parent
     )
     resolved_kind = install_kind or resolve_smoke_test_install_kind(host)
+    # For a GPU kind, the CLI contract is "0 = offload confirmed", so an
+    # inconclusive (no-signal) log must not pass -- require a positive signal.
+    require_signal = resolved_kind in _GPU_INSTALL_KINDS
     if probe:
         probe_path = Path(probe).expanduser().resolve()
         if not probe_path.exists():
@@ -6935,6 +6960,7 @@ def smoke_test_server_binary(
             host,
             resolved_install_dir,
             install_kind = resolved_kind,
+            require_gpu_signal = require_signal,
         )
     else:
         with tempfile.TemporaryDirectory(prefix = "unsloth-llama-smoke-") as tmp:
@@ -6948,6 +6974,7 @@ def smoke_test_server_binary(
                 host,
                 resolved_install_dir,
                 install_kind = resolved_kind,
+                require_gpu_signal = require_signal,
             )
     return resolved_kind
 

studio/setup.ps1

@@ -2667,6 +2667,14 @@ if (-not $NeedLlamaSourceBuild) {
     substep "GGUF inference and export will not be available." "Yellow"
     substep "Install CMake from https://cmake.org/download/ and re-run setup." "Yellow"
     $script:LlamaCppDegraded = $true
+} elseif ($HasROCm -and -not $HasNvidiaSmi) {
+    # Windows has no HIP source-build path, so a ROCm host whose HIP prebuilt was
+    # missing or rejected must not silently CPU-source-build. Mark degraded so
+    # the CPU-prebuilt last resort below installs a clearly labelled CPU build
+    # instead of a "built" install that runs on CPU (#5807).
+    Write-Host ""
+    step "llama.cpp" "no usable HIP prebuilt and no Windows HIP source build; using CPU prebuilt" "Yellow"
+    $script:LlamaCppDegraded = $true
 } else {
     # A source build is committed here. The CUDA toolkit is only needed now, so
     # resolve (and winget-install if needed) it lazily, failing fast if no

tests/studio/install/run_smoke_spoof.py

@@ -84,7 +84,7 @@ def main() -> int:
             ("cuda", GPU_KIND, 0, "GPU binary tagged GPU is accepted"),
             ("cuda_buffer", GPU_KIND, 0, "GPU buffer-format binary is accepted"),
             ("cpu", CPU_KIND, 0, "CPU binary tagged CPU is not gated"),
-            ("no_signal", GPU_KIND, 0, "no-signal log is not rejected"),
+            ("no_signal", GPU_KIND, 1, "no-signal GPU log is inconclusive (exit 1)"),
         ]
         for mode, kind, expected, label in cases:
             rc = run_smoke(wrapper, probe, kind, mode)

tests/studio/install/test_validate_server_gpu_offload.py

@@ -408,11 +408,26 @@ def test_cpu_kind_not_gpu_gated(patched_server, tmp_path):
 
 
 def test_gpu_intent_no_signal_accepted(patched_server, tmp_path):
-    # No buffer/device signal -> conservative: do not reject on no evidence.
+    # No buffer/device signal -> conservative: do not reject on no evidence
+    # (plain install validation; require_gpu_signal defaults False).
     patched_server(NO_SIGNAL_LOG)
     _run_validate(tmp_path, nvidia_host(), "linux-cuda")  # no raise
 
 
+def test_smoke_test_no_signal_gpu_is_inconclusive(patched_server, tmp_path):
+    # The smoke-test CLI sets require_gpu_signal, so a no-signal GPU log is
+    # inconclusive (PrebuiltFallback -> EXIT_ERROR), not a silent pass.
+    patched_server(NO_SIGNAL_LOG)
+    server = tmp_path / "llama-server"
+    server.write_text("#!/bin/sh\n")
+    probe = tmp_path / "probe.gguf"
+    probe.write_bytes(b"GGUF")
+    with pytest.raises(PrebuiltFallback):
+        smoke_test_server_binary(
+            str(server), nvidia_host(), install_dir = str(tmp_path), probe = str(probe)
+        )
+
+
 def test_rocm_cpu_only_rejected(patched_server, tmp_path):
     patched_server(CPU_ONLY_DEVICE_INFO_LOG)
     with pytest.raises(PrebuiltFallback):