fix: gate dynamo-memory numa module behind cfg(target_os = linux) (#6354)

Signed-off-by: PeaBrane <yanrpei@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

lib/memory/src/lib.rs

@@ -14,6 +14,7 @@
 pub mod actions;
 pub mod arena;
 pub mod nixl;
+#[cfg(target_os = "linux")]
 pub mod numa;
 
 /// Offset-based buffer views into underlying storage.
@@ -41,6 +42,7 @@ pub use device::DeviceStorage;
 #[cfg(target_os = "linux")]
 pub use disk::DiskStorage;
 pub use external::ExternalDeviceMemory;
+#[cfg(target_os = "linux")]
 pub use numa::{NumaNode, is_numa_enabled};
 pub use offset::OffsetBuffer;
 pub use pinned::PinnedStorage;

lib/memory/src/pinned.rs

@@ -68,8 +68,6 @@ impl PinnedStorage {
     /// - CUDA context creation fails
     /// - Memory allocation fails
     pub fn new_for_device(len: usize, device_id: Option<u32>) -> Result<Self> {
-        use super::numa;
-
         if len == 0 {
             return Err(StorageError::AllocationFailed(
                 "zero-sized allocations are not supported".into(),
@@ -80,24 +78,21 @@ impl PinnedStorage {
         let ctx = cuda_context(gpu_id)?;
 
         let ptr = match device_id {
-            Some(gpu_id) if numa::is_numa_enabled() => {
-                // NUMA-aware allocation via worker pool
+            #[cfg(target_os = "linux")]
+            Some(gpu_id) if super::numa::is_numa_enabled() => {
                 tracing::debug!(
                     "Using NUMA-aware allocation for {} bytes on GPU {}",
                     len,
                     gpu_id
                 );
-                numa::worker_pool::NumaWorkerPool::global()
+                super::numa::worker_pool::NumaWorkerPool::global()
                     .allocate_pinned_for_gpu(len, gpu_id)
                     .map_err(StorageError::AllocationFailed)? as usize
             }
-            _ => {
-                // Direct allocation (no NUMA or device_id not specified)
-                unsafe {
+            _ => unsafe {
                 ctx.bind_to_thread().map_err(StorageError::Cuda)?;
 
-                    let ptr =
-                        cudarc::driver::result::malloc_host(len, sys::CU_MEMHOSTALLOC_DEVICEMAP)
+                let ptr = cudarc::driver::result::malloc_host(len, sys::CU_MEMHOSTALLOC_DEVICEMAP)
                     .map_err(StorageError::Cuda)?;
 
                 let ptr = ptr as *mut u8;
@@ -106,8 +101,7 @@ impl PinnedStorage {
                 assert!(len < isize::MAX as usize);
 
                 ptr as usize
-                }
-            }
+            },
         };
 
         Ok(Self { ptr, len, ctx })