[ROCm][CI] Select CPU platform for native no-GPU jobs (#49515)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -434,6 +434,12 @@ initialize_native_environment() {
   TIKTOKEN_RS_CACHE_DIR="${HF_HOME}/tiktoken-rs-cache"
   : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
   : "${HF_HUB_ETAG_TIMEOUT:=60}"
+  if [[ "${VLLM_CI_EXPECTED_GPU_COUNT:-1}" == "0" ]]; then
+    # CPU-only native jobs intentionally reuse the ROCm wheel. Make that target
+    # explicit so platform selection does not depend on wheel metadata.
+    VLLM_TARGET_DEVICE=cpu
+    export VLLM_TARGET_DEVICE
+  fi
   export TMPDIR VLLM_RPC_BASE_PATH
   export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
   export HF_HOME HF_DATASETS_CACHE HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT

tests/test_zen_cpu_platform_detection.py

@@ -2,7 +2,9 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 from unittest.mock import mock_open, patch
 
-from vllm.platforms import _is_amd_zen_cpu
+import pytest
+
+from vllm.platforms import _is_amd_zen_cpu, resolve_current_platform_cls_qualname
 
 
 def test_is_amd_zen_cpu_detects_amd_with_avx512():
@@ -35,3 +37,23 @@ def test_is_amd_zen_cpu_returns_false_for_intel_with_avx512():
 def test_is_amd_zen_cpu_returns_false_when_cpuinfo_missing():
     with patch("os.path.exists", return_value=False):
         assert not _is_amd_zen_cpu()
+
+
+def test_cpu_target_selects_cpu_platform_from_non_cpu_wheel(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.setenv("VLLM_TARGET_DEVICE", "cpu")
+
+    with (
+        patch("vllm.platforms.vllm_version_matches_substr") as version_matches,
+        patch("vllm.platforms._is_amd_zen_cpu", return_value=False),
+        patch("vllm.platforms.rocm_platform_plugin") as rocm_plugin,
+    ):
+        assert (
+            resolve_current_platform_cls_qualname() == "vllm.platforms.cpu.CpuPlatform"
+        )
+
+    # An explicit target does not depend on the installed wheel's version
+    # suffix or host accelerators (a native CI job can reuse a ROCm wheel).
+    version_matches.assert_not_called()
+    rocm_plugin.assert_not_called()

vllm/platforms/__init__.py

@@ -160,24 +160,32 @@ def _is_amd_zen_cpu() -> bool:
 
 
 def cpu_platform_plugin() -> str | None:
-    is_cpu = False
     logger.debug("Checking if CPU platform is available.")
-    try:
-        is_cpu = vllm_version_matches_substr("cpu")
-        if is_cpu:
-            logger.debug(
-                "Confirmed CPU platform is available because vLLM is built with CPU."
-            )
-        if not is_cpu:
-            import sys
-
-            is_cpu = sys.platform.startswith("darwin")
+    is_cpu = envs.VLLM_TARGET_DEVICE == "cpu"
+    if is_cpu:
+        logger.debug(
+            "Confirmed CPU platform is available because "
+            "VLLM_TARGET_DEVICE is set to CPU."
+        )
+    else:
+        try:
+            is_cpu = vllm_version_matches_substr("cpu")
             if is_cpu:
                 logger.debug(
-                    "Confirmed CPU platform is available because the machine is MacOS."
+                    "Confirmed CPU platform is available because vLLM is built "
+                    "with CPU."
                 )
-    except Exception as e:
-        logger.debug("CPU platform is not available because: %s", str(e))
+            if not is_cpu:
+                import sys
+
+                is_cpu = sys.platform.startswith("darwin")
+                if is_cpu:
+                    logger.debug(
+                        "Confirmed CPU platform is available because the machine "
+                        "is MacOS."
+                    )
+        except Exception as e:
+            logger.debug("CPU platform is not available because: %s", str(e))
 
     if not is_cpu:
         return None
@@ -209,6 +217,15 @@ builtin_platform_plugins = {
 
 
 def resolve_current_platform_cls_qualname() -> str:
+    # An explicit CPU target is authoritative. Native CPU-only CI jobs reuse
+    # an accelerator wheel and can run on accelerator hosts, so probing every
+    # plugin would otherwise activate both CPU and the host accelerator.
+    if envs.VLLM_TARGET_DEVICE == "cpu":
+        cpu_platform_cls_qualname = cpu_platform_plugin()
+        assert cpu_platform_cls_qualname is not None
+        logger.debug("Explicitly selected CPU platform.")
+        return cpu_platform_cls_qualname
+
     platform_plugins = load_plugins_by_group(PLATFORM_PLUGINS_GROUP)
 
     activated_plugins = []