[Platform] Fix RISC-V platform detection (lscpu parsing + non-NUMA meminfo) (#40427)

Signed-off-by: liuyudong <liuyudong@iscas.ac.cn>

setup.py

@@ -927,7 +927,9 @@ def get_vllm_version() -> str:
     elif _is_tpu():
         version += f"{sep}tpu"
     elif _is_cpu():
-        if envs.VLLM_TARGET_DEVICE == "cpu":
+        # Check the local VLLM_TARGET_DEVICE (may be set by auto-detect above),
+        # not envs.VLLM_TARGET_DEVICE, so CPU-only hosts still get `+cpu`.
+        if VLLM_TARGET_DEVICE == "cpu":
             version += f"{sep}cpu"
     elif _is_xpu():
         version += f"{sep}xpu"

vllm/platforms/__init__.py

@@ -177,7 +177,6 @@ def cpu_platform_plugin() -> str | None:
                 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))
 

vllm/utils/cpu_resource_utils.py

@@ -87,7 +87,13 @@ def get_memory_node_info(node_id: int = 0) -> MemoryNodeInfo:
 
     meminfo_path = f"/sys/devices/system/node/node{node_id}/meminfo"
     if not os.path.exists(meminfo_path):
-        raise RuntimeError(f"{meminfo_path} doesn't exit.")
+        # Non-NUMA systems (e.g. many RISC-V boards) don't expose per-node
+        # meminfo. Fall back to system-wide numbers from psutil.
+        vm = psutil.virtual_memory()
+        return MemoryNodeInfo(
+            total_memory=vm.total,
+            available_memory=vm.available,
+        )
 
     meminfo = {}
     with open(meminfo_path) as f:
@@ -147,19 +153,36 @@ def get_visible_memory_node() -> list[int]:
 
 
 @cache
+def _synthesize_cpu_list() -> list[LogicalCPUInfo]:
+    """Synthesize a flat CPU list: each logical CPU is its own core on
+    NUMA node 0.  Used when lscpu output is unavailable or unparsable
+    (e.g. macOS, RISC-V)."""
+    cpu_count = os.cpu_count()
+    assert cpu_count
+    return [LogicalCPUInfo(i, i, 0) for i in range(cpu_count)]
+
+
 def _get_cpu_list() -> list[LogicalCPUInfo]:
     if platform.system() == "Darwin":
         # For MacOS, no user-level CPU affinity and SMT, return all CPUs
-        cpu_count = os.cpu_count()
-        assert cpu_count
-        return [LogicalCPUInfo(i, i, 0) for i in range(cpu_count)]
+        return _synthesize_cpu_list()
 
     lscpu_output = subprocess.check_output(
         "lscpu --json --extended=CPU,CORE,NODE --online", shell=True, text=True
     )
 
-    # For platform without NUMA, replace '-' to '0'
-    lscpu_output = re.sub(r'"node":\s*-\s*(,|\n)', r'"node": 0\1', lscpu_output)
+    # For platforms without NUMA, map bare `-` node to 0 so non-NUMA
+    # systems keep the existing behavior from #39781.
+    lscpu_output = re.sub(r'"node":\s*-\s*(,|\n|\})', r'"node": 0\1', lscpu_output)
+
+    # On some architectures (notably RISC-V), lscpu also emits bare `-`
+    # for cpu/core.  Quote them so the JSON parses; they will decode to
+    # -1 and be filtered out below, triggering the synthesized fallback.
+    lscpu_output = re.sub(
+        r'("(?:cpu|core)":\s*)-\s*(,|\n|\})',
+        r'\1"-"\2',
+        lscpu_output,
+    )
 
     logical_cpu_list: list[LogicalCPUInfo] = json.loads(
         lscpu_output, object_hook=LogicalCPUInfo.json_decoder
@@ -170,4 +193,9 @@ def _get_cpu_list() -> list[LogicalCPUInfo]:
         x for x in logical_cpu_list if -1 not in (x.id, x.physical_core, x.numa_node)
     ]
 
+    # If lscpu returned no valid entries (e.g. RISC-V where all fields
+    # are bare `-`), fall back to synthesized topology.
+    if not logical_cpu_list:
+        logical_cpu_list = _synthesize_cpu_list()
+
     return logical_cpu_list