fix: harden numa cpuset lookup (#7060)

Signed-off-by: Ryan Olson <rolson@nvidia.com>

lib/kvbm-physical/src/layout/builder.rs

@@ -229,7 +229,7 @@ impl PhysicalLayoutBuilder<HasConfig, HasLayout, NoMemory> {
     ///
     /// # Arguments
     /// * `device_id` - If `Some(id)`, enables NUMA-aware allocation on the GPU's NUMA node
-    ///   (when `DYN_KVBM_ENABLE_NUMA=1` is set). If `None`, uses direct allocation.
+    ///   (disable with `DYN_MEMORY_DISABLE_NUMA=1`). If `None`, uses direct allocation.
     pub fn allocate_pinned(
         self,
         device_id: Option<u32>,

lib/llm/src/block_manager/numa_allocator.rs

@@ -4,12 +4,22 @@
 //! Re-export NUMA utilities from dynamo-memory.
 pub use dynamo_memory::numa::*;
 
+/// Check if NUMA optimization is explicitly opted-in for the block manager.
+///
+/// Set `DYN_KVBM_ENABLE_NUMA=1` to enable NUMA-aware allocation in the
+/// KV cache block manager. This is opt-in because the block manager
+/// manages its own pinned memory allocations separately from `PinnedStorage`.
+pub fn is_numa_enabled() -> bool {
+    matches!(
+        std::env::var("DYN_KVBM_ENABLE_NUMA").as_deref(),
+        Ok("1" | "true" | "yes")
+    )
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
 
-    // ── NumaNode tests ──────────────────────────────────────────────────
-
     #[test]
     fn test_numa_node_equality() {
         let node0a = NumaNode(0);
@@ -61,116 +71,101 @@ mod tests {
     #[test]
     fn test_numa_node_copy_clone() {
         let node1 = NumaNode(5);
-        let node2 = node1; // Copy
-        let node3 = node1; // Clone
+        let node2 = node1;
+        let node3 = node1;
 
         assert_eq!(node1, node2);
         assert_eq!(node1, node3);
         assert_eq!(node2, node3);
     }
 
-    // ── System detection tests ──────────────────────────────────────────
-
     #[test]
     fn test_get_current_cpu_numa_node() {
         let node = get_current_cpu_numa_node();
-
         if !node.is_unknown() {
             assert!(node.0 < 8, "NUMA node {} seems unreasonably high", node.0);
         }
     }
 
-    #[test]
-    fn test_get_device_numa_node_valid_gpu() {
-        let node = get_device_numa_node(0);
-        println!("GPU 0 detected on NUMA node: {}", node.0);
-    }
-
-    // ── Worker pool tests ───────────────────────────────────────────────
-    //
-    // NumaWorker and NumaWorkerPool::new() are private in dynamo-memory,
-    // so these tests go through the public NumaWorkerPool::global() API.
-
-    /// Check if CUDA is available for testing
-    fn is_cuda_available() -> bool {
-        if std::process::Command::new("nvidia-smi")
-            .arg("--query-gpu=count")
-            .arg("--format=csv,noheader")
-            .output()
-            .is_err()
-        {
-            return false;
-        }
-
-        crate::block_manager::storage::cuda::Cuda::device_or_create(0).is_ok()
-    }
-
     #[test]
     fn test_worker_pool_singleton() {
         let pool1 = worker_pool::NumaWorkerPool::global();
         let pool2 = worker_pool::NumaWorkerPool::global();
         assert!(std::ptr::eq(pool1, pool2));
     }
+}
+
+#[cfg(all(test, feature = "testing-cuda"))]
+mod cuda_tests {
+    use super::*;
 
     #[test]
-    fn test_worker_pool_allocate() {
-        if !is_cuda_available() {
-            eprintln!("Skipping test_worker_pool_allocate: CUDA not available");
-            return;
+    fn test_get_device_numa_node_valid_gpu() {
+        match get_device_numa_node(0) {
+            Some(node) => println!("GPU 0 detected on NUMA node: {}", node.0),
+            None => println!("GPU 0 has no determinable NUMA node"),
         }
+    }
 
+    #[test]
+    fn test_worker_pool_allocate() {
         let pool = worker_pool::NumaWorkerPool::global();
 
-        unsafe {
-            let ptr = pool.allocate_pinned_for_gpu(8192, 0).unwrap();
-            assert!(!ptr.is_null());
-
-            cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
+        match pool.allocate_pinned_for_gpu(8192, 0).unwrap() {
+            Some(ptr) => unsafe {
+                assert!(!ptr.is_null());
+                cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
+            },
+            None => {
+                println!("NUMA node unknown for GPU 0, allocation skipped");
+            }
         }
     }
 
     #[test]
     fn test_worker_pool_reuse() {
-        if !is_cuda_available() {
-            eprintln!("Skipping test_worker_pool_reuse: CUDA not available");
-            return;
-        }
-
         let pool = worker_pool::NumaWorkerPool::global();
 
-        unsafe {
-            let ptr1 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
-            let ptr2 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
-
-            assert!(!ptr1.is_null());
-            assert!(!ptr2.is_null());
-            assert_ne!(ptr1, ptr2);
-
-            cudarc::driver::result::free_host(ptr1 as *mut std::ffi::c_void).unwrap();
-            cudarc::driver::result::free_host(ptr2 as *mut std::ffi::c_void).unwrap();
+        let r1 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
+        let r2 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
+
+        match (r1, r2) {
+            (Some(ptr1), Some(ptr2)) => unsafe {
+                assert!(!ptr1.is_null());
+                assert!(!ptr2.is_null());
+                assert_ne!(ptr1, ptr2);
+                cudarc::driver::result::free_host(ptr1 as *mut std::ffi::c_void).unwrap();
+                cudarc::driver::result::free_host(ptr2 as *mut std::ffi::c_void).unwrap();
+            },
+            (None, None) => {
+                println!("NUMA node unknown, both allocations skipped");
+            }
+            _ => panic!("inconsistent NUMA detection between two calls for same GPU"),
         }
     }
 
     #[test]
     fn test_zero_size_allocation() {
-        if !is_cuda_available() {
-            eprintln!("Skipping test_zero_size_allocation: CUDA not available");
-            return;
-        }
-
         let pool = worker_pool::NumaWorkerPool::global();
         let result = pool.allocate_pinned_for_gpu(0, 0);
-        assert!(result.is_err());
-        assert!(result.unwrap_err().contains("zero"));
+        match result {
+            Ok(None) => {
+                println!("NUMA node unknown, zero-size check not reached");
+            }
+            Err(e) => {
+                assert!(e.contains("zero"));
+            }
+            Ok(Some(_)) => panic!("zero-size allocation should not succeed"),
+        }
     }
 
     #[test]
     fn test_pinned_allocation_api() {
         let pool = worker_pool::NumaWorkerPool::global();
 
-        unsafe {
-            if let Ok(ptr) = pool.allocate_pinned_for_gpu(1024, 0) {
-                assert!(!ptr.is_null());
+        if let Ok(Some(ptr)) = pool.allocate_pinned_for_gpu(1024, 0) {
+            assert!(!ptr.is_null());
+            unsafe {
                 cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
             }
         }

lib/llm/src/block_manager/storage/cuda.rs

@@ -216,7 +216,14 @@ impl PinnedStorage {
                 match numa_allocator::worker_pool::NumaWorkerPool::global()
                     .allocate_pinned_for_gpu(size, device_id)
                 {
-                    Ok(ptr) => ptr,
+                    Ok(Some(ptr)) => ptr,
+                    Ok(None) => {
+                        tracing::debug!(
+                            "NUMA node unknown for GPU {}, using direct allocation",
+                            device_id
+                        );
+                        malloc_host_prefer_writecombined(size)?
+                    }
                     Err(e) => {
                         tracing::warn!("NUMA allocation failed: {}, using direct allocation", e);
                         malloc_host_prefer_writecombined(size)?

lib/memory/src/device.rs

@@ -10,7 +10,7 @@ use std::collections::HashMap;
 use std::sync::{Arc, Mutex, OnceLock};
 
 /// Get or create a CUDA context for the given device.
-fn cuda_context(device_id: u32) -> Result<Arc<CudaContext>> {
+pub(crate) fn cuda_context(device_id: u32) -> Result<Arc<CudaContext>> {
     static CONTEXTS: OnceLock<Mutex<HashMap<u32, Arc<CudaContext>>>> = OnceLock::new();
     let mut map = CONTEXTS.get_or_init(Default::default).lock().unwrap();
 

lib/memory/src/lib.rs

@@ -43,7 +43,7 @@ pub use device::DeviceStorage;
 pub use disk::DiskStorage;
 pub use external::ExternalDeviceMemory;
 #[cfg(target_os = "linux")]
-pub use numa::{NumaNode, is_numa_enabled};
+pub use numa::{NumaNode, is_numa_disabled};
 pub use offset::OffsetBuffer;
 pub use pinned::PinnedStorage;
 pub use pool::{CudaMemPool, CudaMemPoolBuilder};

lib/memory/src/numa/mod.rs

@@ -15,30 +15,30 @@
 //!
 //! ## Usage
 //!
-//! NUMA optimization is opt-in via environment variable:
+//! NUMA optimization is enabled by default. To disable it:
 //! ```bash
-//! export DYN_KVBM_ENABLE_NUMA=1
+//! export DYN_MEMORY_DISABLE_NUMA=1
 //! ```
 //!
 //! When enabled, pinned memory allocations are routed through NUMA workers
 //! that are pinned to the target GPU's NUMA node, ensuring first-touch policy
-//! places pages on the correct node.
+//! places pages on the correct node. If the GPU's NUMA node cannot be
+//! determined, allocation falls back to the non-NUMA path transparently.
 
 pub mod topology;
 pub mod worker_pool;
 
+use cudarc::driver::sys::CUdevice_attribute_enum;
 use nix::libc;
 use serde::{Deserialize, Serialize};
-use std::{mem, process::Command};
+use std::{fs, mem, process::Command};
 
-/// Check if NUMA optimization is enabled via environment variable
+/// Check if NUMA optimization is disabled via environment variable.
 ///
-/// Set `DYN_KVBM_ENABLE_NUMA=1` to enable NUMA-aware allocation.
-/// Default: disabled (opt-in)
-pub fn is_numa_enabled() -> bool {
-    std::env::var("DYN_KVBM_ENABLE_NUMA")
-        .map(|v| v == "1" || v.to_lowercase() == "true")
-        .unwrap_or(false)
+/// NUMA-aware allocation is enabled by default. Set `DYN_MEMORY_DISABLE_NUMA=1`
+/// (or any truthy value) to disable it.
+pub fn is_numa_disabled() -> bool {
+    dynamo_config::env_is_truthy("DYN_MEMORY_DISABLE_NUMA")
 }
 
 /// Represents a NUMA node identifier.
@@ -92,88 +92,126 @@ pub fn get_current_cpu_numa_node() -> NumaNode {
     }
 }
 
-/// Resolve process-local CUDA device index to the physical identifier for nvidia-smi.
+/// Format a PCI bus address from domain, bus, and device IDs.
 ///
-/// When `CUDA_VISIBLE_DEVICES` is set, the process sees a remapped device space (e.g. only
-/// GPU 2 visible as device 0). nvidia-smi's `-i` flag expects the *physical* device index or
-/// UUID, not the process-local index. This function parses `CUDA_VISIBLE_DEVICES` to map
-/// process-local `device_id` to the correct physical identifier.
+/// Returns a string in the format `"DDDD:BB:DD.0"` suitable for sysfs lookups.
+fn format_pci_bus_address(domain: i32, bus: i32, device: i32) -> String {
+    format!("{:04x}:{:02x}:{:02x}.0", domain, bus, device)
+}
+
+/// Query the PCI bus address for a CUDA device from the CUDA driver API.
 ///
-/// Returns the identifier string to pass to `nvidia-smi -i` (physical index or UUID).
-fn cuda_device_id_to_nvidia_smi_id(device_id: u32) -> String {
-    let visible = match std::env::var("CUDA_VISIBLE_DEVICES") {
-        Ok(v) if !v.trim().is_empty() => v,
-        _ => return device_id.to_string(), // No remapping: identity
-    };
+/// Uses `CudaContext::attribute()` to read PCI domain, bus, and device IDs.
+/// This transparently handles `CUDA_VISIBLE_DEVICES` remapping since
+/// `CudaContext::new(ordinal)` operates on the process-local device index.
+fn get_pci_bus_address_from_cuda(device_id: u32) -> Option<String> {
+    let ctx = crate::device::cuda_context(device_id).ok()?;
+    let domain = ctx
+        .attribute(CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID)
+        .ok()?;
+    let bus = ctx
+        .attribute(CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_PCI_BUS_ID)
+        .ok()?;
+    let device = ctx
+        .attribute(CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID)
+        .ok()?;
+    Some(format_pci_bus_address(domain, bus, device))
+}
 
-    // Parse comma-separated list. Supports: "0,1,2", "2,3", "GPU-uuid", "2,GPU-uuid", etc.
-    let devices: Vec<&str> = visible
-        .split(',')
-        .map(|s| s.trim())
-        .filter(|s| !s.is_empty())
-        .collect();
-    if device_id as usize >= devices.len() {
-        tracing::warn!(
-            "device_id {} out of range for CUDA_VISIBLE_DEVICES ({} devices), using identity",
-            device_id,
-            devices.len()
-        );
-        return device_id.to_string();
+/// Read the NUMA node for a PCI device from sysfs.
+///
+/// Reads `/sys/bus/pci/devices/<pci_address>/numa_node`. Returns `None` if the
+/// file doesn't exist, can't be read, or contains `-1` (no NUMA affinity).
+fn read_numa_node_from_sysfs(pci_address: &str) -> Option<NumaNode> {
+    let path = format!("/sys/bus/pci/devices/{}/numa_node", pci_address);
+    let content = fs::read_to_string(&path).ok()?;
+    let node: i32 = content.trim().parse().ok()?;
+    if node < 0 {
+        // -1 means no NUMA affinity info available
+        None
+    } else {
+        Some(NumaNode(node as u32))
+    }
+}
+
+/// Fallback: query NUMA node from nvidia-smi using PCI bus address.
+///
+/// Uses the PCI BDF address (not env-var-based device index) so it is
+/// correct regardless of `CUDA_VISIBLE_DEVICES` remapping.
+fn get_numa_node_from_nvidia_smi(pci_address: &str) -> Option<NumaNode> {
+    let output = Command::new("nvidia-smi")
+        .args(["topo", "--get-numa-id-of-nearby-cpu", "-i", pci_address])
+        .output()
+        .ok()?;
+
+    if !output.status.success() {
+        return None;
     }
 
-    let id = devices[device_id as usize];
-    id.to_string()
+    let stdout = std::str::from_utf8(&output.stdout).ok()?;
+    let line = stdout.lines().next()?;
+    let numa_str = line.split(':').nth(1)?;
+    let node: u32 = numa_str.trim().parse().ok()?;
+    Some(NumaNode(node))
 }
 
 /// Get NUMA node for a GPU device.
 ///
-/// For GPU memory, the NUMA affinity depends on which PCIe bus the GPU is attached to.
-/// This is queried via nvidia-smi. Falls back to a heuristic (device_id % 2) if nvidia-smi
-/// is unavailable.
+/// Queries the PCI bus address from the CUDA driver API, then reads the NUMA
+/// node from sysfs. Falls back to nvidia-smi with the PCI address. Returns
+/// `None` if the NUMA node cannot be determined, signaling the caller to skip
+/// NUMA-aware allocation entirely rather than guessing wrong.
 ///
-/// When `CUDA_VISIBLE_DEVICES` is set, the process-local `device_id` is correctly mapped
-/// to the physical GPU identifier before querying nvidia-smi, so NUMA attribution is accurate.
+/// `CUDA_VISIBLE_DEVICES` is handled transparently because `CudaContext::new(ordinal)`
+/// operates on the process-local device index.
 ///
 /// # Arguments
 /// * `device_id` - CUDA device index (0, 1, 2, ...) as seen by the process
 ///
 /// # Returns
-/// The NUMA node closest to the specified GPU, or a heuristic fallback.
-pub fn get_device_numa_node(device_id: u32) -> NumaNode {
-    let nvidia_smi_id = cuda_device_id_to_nvidia_smi_id(device_id);
-
-    // Use nvidia-smi topo to get NUMA ID of nearest CPU
-    // -i must be physical device index or UUID, not process-local index
-    let output = match Command::new("nvidia-smi")
-        .args(["topo", "--get-numa-id-of-nearby-cpu", "-i", &nvidia_smi_id])
-        .output()
-    {
-        Ok(out) if out.status.success() => out,
-        _ => {
+/// The NUMA node closest to the specified GPU, or `None` if it cannot be determined.
+pub fn get_device_numa_node(device_id: u32) -> Option<NumaNode> {
+    // Step 1: Get PCI bus address from CUDA driver
+    let pci_address = match get_pci_bus_address_from_cuda(device_id) {
+        Some(addr) => addr,
+        None => {
             tracing::warn!(
-                "nvidia-smi failed for GPU {} (nvidia-smi -i {}), using heuristic",
-                device_id,
-                nvidia_smi_id
+                "Failed to get PCI address from CUDA for device {}, skipping NUMA optimization",
+                device_id
             );
-            return NumaNode(device_id % 2);
+            return None;
         }
     };
 
-    if let Ok(stdout) = std::str::from_utf8(&output.stdout)
-        && let Some(line) = stdout.lines().next()
-        && let Some(numa_str) = line.split(':').nth(1)
-        && let Ok(node) = numa_str.trim().parse::<u32>()
-    {
+    // Step 2: Read NUMA node from sysfs
+    if let Some(node) = read_numa_node_from_sysfs(&pci_address) {
+        tracing::trace!(
+            "GPU {} (PCI {}) on NUMA node {} (sysfs)",
+            device_id,
+            pci_address,
+            node.0
+        );
+        return Some(node);
+    }
+
+    // Step 3: Fallback to nvidia-smi with PCI address
+    if let Some(node) = get_numa_node_from_nvidia_smi(&pci_address) {
         tracing::trace!(
-            "GPU {} (physical {}) on NUMA node {}",
+            "GPU {} (PCI {}) on NUMA node {} (nvidia-smi)",
             device_id,
-            nvidia_smi_id,
-            node
+            pci_address,
+            node.0
         );
-        return NumaNode(node);
+        return Some(node);
     }
-    tracing::warn!("Failed to get NUMA node for GPU {}", device_id);
-    NumaNode::UNKNOWN
+
+    // No NUMA info available — caller should skip NUMA optimization
+    tracing::warn!(
+        "Could not determine NUMA node for GPU {} (PCI {}), skipping NUMA optimization",
+        device_id,
+        pci_address
+    );
+    None
 }
 
 /// Pin the current thread to a specific NUMA node's CPUs.
@@ -257,7 +295,6 @@ mod tests {
 
     #[test]
     fn test_numa_node_serialization() {
-        // Verify NumaNode can be serialized (important for benchmarking)
         let node = NumaNode(1);
         let json = serde_json::to_string(&node).unwrap();
         let deserialized: NumaNode = serde_json::from_str(&json).unwrap();
@@ -266,28 +303,14 @@ mod tests {
 
     #[test]
     fn test_get_current_cpu_numa_node() {
-        // Should either return a valid node or UNKNOWN
         let node = get_current_cpu_numa_node();
-
-        // If not unknown, should be a reasonable NUMA node number (< 8 on most systems)
         if !node.is_unknown() {
             assert!(node.0 < 8, "NUMA node {} seems unreasonably high", node.0);
         }
     }
 
-    #[test]
-    fn test_get_device_numa_node_valid_gpu() {
-        // Test GPU 0 detection
-        let node = get_device_numa_node(0);
-
-        // Should return either a valid node (0-7) or use heuristic (gpu_id % 2)
-        // On dual-socket systems, GPU 0 typically on node 0 or 1
-        println!("GPU 0 detected on NUMA node: {}", node.0);
-    }
-
     #[test]
     fn test_numa_node_hash() {
-        // Verify NumaNode can be used as a HashMap key
         use std::collections::HashMap;
 
         let mut map = HashMap::new();
@@ -301,13 +324,78 @@ mod tests {
 
     #[test]
     fn test_numa_node_copy_clone() {
-        // Verify NumaNode is Copy and Clone
         let node1 = NumaNode(5);
-        let node2 = node1; // Copy
-        let node3 = node1; // Clone
+        let node2 = node1;
+        let node3 = node1;
 
         assert_eq!(node1, node2);
         assert_eq!(node1, node3);
         assert_eq!(node2, node3);
     }
+
+    #[test]
+    fn test_format_pci_bus_address() {
+        assert_eq!(format_pci_bus_address(0, 0, 0), "0000:00:00.0");
+        assert_eq!(format_pci_bus_address(0, 0x3b, 0), "0000:3b:00.0");
+        assert_eq!(format_pci_bus_address(0, 0xaf, 0), "0000:af:00.0");
+        assert_eq!(format_pci_bus_address(0x10, 0x1a, 0x03), "0010:1a:03.0");
+    }
+
+    #[test]
+    fn test_read_numa_node_from_sysfs_nonexistent() {
+        assert!(read_numa_node_from_sysfs("ffff:ff:ff.0").is_none());
+    }
+}
+
+#[cfg(all(test, feature = "testing-cuda"))]
+mod cuda_tests {
+    use super::*;
+
+    #[test]
+    fn test_get_pci_bus_address_from_cuda() {
+        let addr = get_pci_bus_address_from_cuda(0).expect("should get PCI address for GPU 0");
+        // Validate BDF format: DDDD:BB:DD.0
+        let parts: Vec<&str> = addr.split(':').collect();
+        assert_eq!(
+            parts.len(),
+            3,
+            "PCI address should have 3 colon-separated parts: {}",
+            addr
+        );
+        assert_eq!(parts[0].len(), 4, "domain should be 4 hex chars: {}", addr);
+        assert!(parts[2].ends_with(".0"), "should end with .0: {}", addr);
+        println!("GPU 0 PCI address: {}", addr);
+    }
+
+    #[test]
+    fn test_read_numa_node_from_sysfs_real_gpu() {
+        let addr = get_pci_bus_address_from_cuda(0).expect("should get PCI address for GPU 0");
+        if let Some(node) = read_numa_node_from_sysfs(&addr) {
+            assert!(node.0 < 16, "NUMA node {} seems unreasonably high", node.0);
+            println!("GPU 0 (PCI {}) sysfs NUMA node: {}", addr, node.0);
+        } else {
+            println!(
+                "GPU 0 (PCI {}) has no sysfs NUMA info (single-socket?)",
+                addr
+            );
+        }
+    }
+
+    #[test]
+    fn test_get_device_numa_node_returns_some_or_none() {
+        let result = get_device_numa_node(0);
+        match result {
+            Some(node) => {
+                assert!(node.0 < 16, "NUMA node {} seems unreasonably high", node.0);
+                assert!(
+                    !node.is_unknown(),
+                    "should never return UNKNOWN inside Some"
+                );
+                println!("GPU 0 detected on NUMA node: {}", node.0);
+            }
+            None => {
+                println!("GPU 0 has no determinable NUMA node (single-socket or no sysfs info)");
+            }
+        }
+    }
 }

lib/memory/src/numa/worker_pool.rs

@@ -13,11 +13,9 @@
 //! - First-touch page allocation ensures correct NUMA placement
 
 use super::get_current_cpu_numa_node;
-use cudarc::driver::CudaContext;
 use cudarc::driver::result::malloc_host;
 use cudarc::driver::sys::CU_MEMHOSTALLOC_DEVICEMAP;
 use nix::libc;
-use std::collections::HashMap;
 use std::sync::mpsc::{Receiver, Sender, channel};
 use std::sync::{Arc, Mutex, OnceLock};
 use std::thread::{self, JoinHandle};
@@ -25,25 +23,6 @@ use std::time::Duration;
 
 use super::{NumaNode, get_device_numa_node};
 
-/// Get or create a CUDA context for the given device.
-fn cuda_context(device_id: u32) -> Result<Arc<CudaContext>, String> {
-    static CONTEXTS: OnceLock<Mutex<HashMap<u32, Arc<CudaContext>>>> = OnceLock::new();
-    let mut map = CONTEXTS.get_or_init(Default::default).lock().unwrap();
-
-    if let Some(existing) = map.get(&device_id) {
-        return Ok(existing.clone());
-    }
-
-    let ctx = CudaContext::new(device_id as usize).map_err(|e| {
-        format!(
-            "Failed to create CUDA context for device {}: {:?}",
-            device_id, e
-        )
-    })?;
-    map.insert(device_id, ctx.clone());
-    Ok(ctx)
-}
-
 /// Wrapper for raw pointer that can be sent between threads.
 ///
 /// # Safety
@@ -197,7 +176,8 @@ impl NumaWorker {
         }
 
         // Get or create CUDA context for this GPU
-        let ctx = cuda_context(gpu_id)?;
+        let ctx = crate::device::cuda_context(gpu_id)
+            .map_err(|e| format!("Failed to create CUDA context for device {}: {}", gpu_id, e))?;
 
         unsafe {
             // Bind CUDA context to this worker thread before allocation
@@ -370,19 +350,35 @@ impl NumaWorkerPool {
     /// Allocate CUDA pinned memory for a specific GPU (auto-detects NUMA node).
     ///
     /// This method:
-    /// 1. Determines the GPU's NUMA node via nvidia-smi
+    /// 1. Determines the GPU's NUMA node via CUDA driver PCI attributes + sysfs
     /// 2. Routes the allocation to a worker pinned to that node
     /// 3. The worker allocates and touches pages to ensure first-touch placement
     ///
+    /// Returns `None` if the GPU's NUMA node cannot be determined, signaling
+    /// the caller to fall back to non-NUMA allocation.
+    ///
     /// # Arguments
     /// * `size` - Number of bytes to allocate
     /// * `gpu_id` - CUDA device ID
     ///
     /// # Returns
-    /// Raw pointer to the allocated memory. Caller is responsible for freeing via
-    /// `cudarc::driver::result::free_host`.
-    pub fn allocate_pinned_for_gpu(&self, size: usize, gpu_id: u32) -> Result<*mut u8, String> {
-        let node = get_device_numa_node(gpu_id);
+    /// `Some(ptr)` on success, `None` if NUMA node is unknown (caller should
+    /// use non-NUMA allocation). Returns `Err` on allocation failure.
+    pub fn allocate_pinned_for_gpu(
+        &self,
+        size: usize,
+        gpu_id: u32,
+    ) -> Result<Option<*mut u8>, String> {
+        let node = match get_device_numa_node(gpu_id) {
+            Some(node) => node,
+            None => {
+                tracing::debug!(
+                    "NUMA node unknown for GPU {}, skipping NUMA-aware allocation",
+                    gpu_id
+                );
+                return Ok(None);
+            }
+        };
 
         tracing::debug!(
             "Allocating {} bytes pinned memory for GPU {} (NUMA node {})",
@@ -392,30 +388,16 @@ impl NumaWorkerPool {
         );
 
         let worker = self.get_or_spawn_worker(node)?;
-        worker.allocate(size, gpu_id).map(|send_ptr| send_ptr.0)
+        worker
+            .allocate(size, gpu_id)
+            .map(|send_ptr| Some(send_ptr.0))
     }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
-    use crate::numa::{get_current_cpu_numa_node, get_device_numa_node};
-
-    /// Check if CUDA is available for testing.
-    fn is_cuda_available() -> bool {
-        // Check if nvidia-smi is available
-        if std::process::Command::new("nvidia-smi")
-            .arg("--query-gpu=count")
-            .arg("--format=csv,noheader")
-            .output()
-            .is_err()
-        {
-            return false;
-        }
-
-        // Try to initialize CUDA context for device 0
-        cuda_context(0).is_ok()
-    }
+    use crate::numa::get_current_cpu_numa_node;
 
     #[test]
     fn test_worker_spawn() {
@@ -425,12 +407,48 @@ mod tests {
     }
 
     #[test]
-    fn test_worker_allocate_pinned() {
-        if !is_cuda_available() {
-            eprintln!("Skipping test_worker_allocate_pinned: CUDA not available");
-            return;
+    fn test_worker_pool_singleton() {
+        let pool1 = NumaWorkerPool::global();
+        let pool2 = NumaWorkerPool::global();
+        assert!(std::ptr::eq(pool1, pool2));
+    }
+
+    #[test]
+    fn test_get_current_cpu_numa_node() {
+        let node = get_current_cpu_numa_node();
+        if !node.is_unknown() {
+            println!("Current CPU on NUMA node: {}", node.0);
+        } else {
+            println!("NUMA node detection unavailable (single-node or fake NUMA)");
         }
+    }
+
+    #[test]
+    fn test_numa_node_display() {
+        let node = NumaNode(0);
+        assert_eq!(format!("{}", node), "NumaNode(0)");
+
+        let unknown = NumaNode::UNKNOWN;
+        assert_eq!(format!("{}", unknown), "UNKNOWN");
+    }
+
+    #[test]
+    fn test_numa_node_is_unknown() {
+        let valid = NumaNode(0);
+        assert!(!valid.is_unknown());
+
+        let unknown = NumaNode::UNKNOWN;
+        assert!(unknown.is_unknown());
+    }
+}
 
+#[cfg(all(test, feature = "testing-cuda"))]
+mod cuda_tests {
+    use super::*;
+    use crate::numa::get_device_numa_node;
+
+    #[test]
+    fn test_worker_allocate_pinned() {
         let node = NumaNode(0);
         let worker = NumaWorker::spawn(node).unwrap();
 
@@ -445,123 +463,83 @@ mod tests {
 
     #[test]
     fn test_worker_pool() {
-        if !is_cuda_available() {
-            eprintln!("Skipping test_worker_pool: CUDA not available");
-            return;
-        }
-
         let pool = NumaWorkerPool::new();
 
-        unsafe {
-            let ptr = pool.allocate_pinned_for_gpu(8192, 0).unwrap();
-            assert!(!ptr.is_null());
-
-            cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
+        match pool.allocate_pinned_for_gpu(8192, 0).unwrap() {
+            Some(ptr) => unsafe {
+                assert!(!ptr.is_null());
+                cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
+            },
+            None => {
+                println!(
+                    "NUMA node unknown for GPU 0, allocation skipped (expected on single-socket)"
+                );
+            }
         }
     }
 
-    #[test]
-    fn test_worker_pool_singleton() {
-        // Verify that global() returns the same instance
-        let pool1 = NumaWorkerPool::global();
-        let pool2 = NumaWorkerPool::global();
-
-        // They should be the same static reference
-        assert!(std::ptr::eq(pool1, pool2));
-    }
-
     #[test]
     fn test_worker_reuse() {
-        if !is_cuda_available() {
-            eprintln!("Skipping test_worker_reuse: CUDA not available");
-            return;
-        }
-
-        // Test that subsequent allocations for the same GPU reuse the same worker
         let pool = NumaWorkerPool::new();
 
-        unsafe {
-            // First allocation spawns worker for GPU 0
-            let ptr1 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
-
-            // Second allocation should reuse worker for GPU 0
-            let ptr2 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
-
-            assert!(!ptr1.is_null());
-            assert!(!ptr2.is_null());
-            assert_ne!(ptr1, ptr2);
-
-            cudarc::driver::result::free_host(ptr1 as *mut std::ffi::c_void).unwrap();
-            cudarc::driver::result::free_host(ptr2 as *mut std::ffi::c_void).unwrap();
+        // If NUMA node is unknown, both calls return None — that's fine
+        let r1 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
+        let r2 = pool.allocate_pinned_for_gpu(1024, 0).unwrap();
+
+        match (r1, r2) {
+            (Some(ptr1), Some(ptr2)) => unsafe {
+                assert!(!ptr1.is_null());
+                assert!(!ptr2.is_null());
+                assert_ne!(ptr1, ptr2);
+                cudarc::driver::result::free_host(ptr1 as *mut std::ffi::c_void).unwrap();
+                cudarc::driver::result::free_host(ptr2 as *mut std::ffi::c_void).unwrap();
+            },
+            (None, None) => {
+                println!("NUMA node unknown, both allocations skipped");
+            }
+            _ => panic!("inconsistent NUMA detection between two calls for same GPU"),
         }
     }
 
     #[test]
-    fn test_zero_size_allocation() {
-        // Test that zero-size allocations are rejected
+    fn test_zero_size_allocation_with_known_node() {
+        // Zero-size is rejected by the worker, but only if NUMA node is known.
+        // If NUMA node is unknown, allocate_pinned_for_gpu returns Ok(None) before
+        // reaching the worker.
         let pool = NumaWorkerPool::new();
         let result = pool.allocate_pinned_for_gpu(0, 0);
-        assert!(result.is_err());
-        assert!(result.unwrap_err().contains("zero"));
-    }
-
-    #[test]
-    fn test_get_current_cpu_numa_node() {
-        // Test that we can detect current CPU's NUMA node
-        let node = get_current_cpu_numa_node();
-
-        // On a real NUMA system, should return a valid node
-        // On fake NUMA or single-node, might return 0 or UNKNOWN
-        if !node.is_unknown() {
-            println!("Current CPU on NUMA node: {}", node.0);
-        } else {
-            println!("NUMA node detection unavailable (single-node or fake NUMA)");
+        match result {
+            Ok(None) => {
+                println!("NUMA node unknown, zero-size check not reached");
+            }
+            Err(e) => {
+                assert!(e.contains("zero"));
+            }
+            Ok(Some(_)) => panic!("zero-size allocation should not succeed"),
         }
     }
 
     #[test]
     fn test_get_device_numa_node() {
-        // Test GPU NUMA node detection
-        // This will only work if nvidia-smi is available
         let node = get_device_numa_node(0);
-
-        if !node.is_unknown() {
-            println!("GPU 0 on NUMA node: {}", node.0);
-            // Node should be 0 or 1 on typical dual-socket systems
-            assert!(node.0 <= 1 || node.0 == u32::MAX);
-        } else {
-            println!("GPU NUMA detection unavailable (no nvidia-smi or no GPU)");
+        match node {
+            Some(n) => {
+                assert!(n.0 < 16, "NUMA node {} seems unreasonably high", n.0);
+                println!("GPU 0 on NUMA node: {}", n.0);
+            }
+            None => {
+                println!("GPU 0 has no determinable NUMA node");
+            }
         }
     }
 
-    #[test]
-    fn test_numa_node_display() {
-        // Test Display implementation for NumaNode
-        let node = NumaNode(0);
-        assert_eq!(format!("{}", node), "NumaNode(0)");
-
-        let unknown = NumaNode::UNKNOWN;
-        assert_eq!(format!("{}", unknown), "UNKNOWN");
-    }
-
-    #[test]
-    fn test_numa_node_is_unknown() {
-        let valid = NumaNode(0);
-        assert!(!valid.is_unknown());
-
-        let unknown = NumaNode::UNKNOWN;
-        assert!(unknown.is_unknown());
-    }
-
     #[test]
     fn test_pinned_allocation_api() {
-        // Verify the public API works for pinned allocation
         let pool = NumaWorkerPool::new();
 
-        unsafe {
-            // Test that we can allocate pinned memory for a GPU
-            if let Ok(ptr) = pool.allocate_pinned_for_gpu(1024, 0) {
-                assert!(!ptr.is_null());
+        if let Some(ptr) = pool.allocate_pinned_for_gpu(1024, 0).unwrap() {
+            assert!(!ptr.is_null());
+            unsafe {
                 cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
             }
         }
@@ -569,22 +547,15 @@ mod tests {
 
     #[test]
     fn test_worker_channel_communication() {
-        // Test that worker receives and processes requests
         let node = NumaNode(0);
         let worker = NumaWorker::spawn(node).unwrap();
 
-        // Send allocation request
-        let result = worker.allocate(1024, 0);
-
-        // Should get a response (either success or timeout)
-        assert!(result.is_ok() || result.is_err());
+        let send_ptr = worker.allocate(1024, 0).unwrap();
+        let ptr = send_ptr.0;
+        assert!(!ptr.is_null());
 
-        if let Ok(send_ptr) = result {
-            unsafe {
-                let ptr = send_ptr.0;
-                assert!(!ptr.is_null());
-                cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
-            }
+        unsafe {
+            cudarc::driver::result::free_host(ptr as *mut std::ffi::c_void).unwrap();
         }
     }
 }

lib/memory/src/pinned.rs

@@ -7,22 +7,7 @@ use super::{MemoryDescriptor, Result, StorageError, StorageKind, actions, nixl::
 use cudarc::driver::CudaContext;
 use cudarc::driver::sys;
 use std::any::Any;
-use std::collections::HashMap;
-use std::sync::{Arc, Mutex, OnceLock};
-
-/// Get or create a CUDA context for the given device.
-fn cuda_context(device_id: u32) -> Result<Arc<CudaContext>> {
-    static CONTEXTS: OnceLock<Mutex<HashMap<u32, Arc<CudaContext>>>> = OnceLock::new();
-    let mut map = CONTEXTS.get_or_init(Default::default).lock().unwrap();
-
-    if let Some(existing) = map.get(&device_id) {
-        return Ok(existing.clone());
-    }
-
-    let ctx = CudaContext::new(device_id as usize)?;
-    map.insert(device_id, ctx.clone());
-    Ok(ctx)
-}
+use std::sync::Arc;
 
 /// CUDA pinned host memory allocated via cudaHostAlloc.
 #[derive(Debug)]
@@ -51,10 +36,11 @@ impl PinnedStorage {
 
     /// Allocate pinned memory, optionally NUMA-aware for a specific GPU.
     ///
-    /// When `device_id` is `Some`, the allocation is performed on a worker thread
-    /// pinned to the GPU's NUMA node, ensuring optimal memory placement via
-    /// first-touch policy, However, NUMA is only used if enabled via the
-    /// `DYN_KVBM_ENABLE_NUMA=1` environment variable.
+    /// When `device_id` is `Some`, NUMA-aware allocation is attempted by default:
+    /// a worker thread pinned to the GPU's NUMA node performs the allocation,
+    /// ensuring optimal memory placement via first-touch policy. If the GPU's
+    /// NUMA node cannot be determined, allocation falls back to the direct path.
+    /// Set `DYN_MEMORY_DISABLE_NUMA=1` to skip NUMA optimization entirely.
     ///
     /// When `device_id` is `None`, a direct allocation is performed on device 0.
     ///
@@ -75,21 +61,40 @@ impl PinnedStorage {
         }
 
         let gpu_id = device_id.unwrap_or(0);
-        let ctx = cuda_context(gpu_id)?;
-
-        let ptr = match device_id {
-            #[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
-                );
-                super::numa::worker_pool::NumaWorkerPool::global()
+        let ctx = crate::device::cuda_context(gpu_id)?;
+
+        // Try NUMA-aware allocation unless explicitly disabled
+        #[cfg(target_os = "linux")]
+        let numa_ptr = if let Some(gpu_id) = device_id {
+            if !super::numa::is_numa_disabled() {
+                match super::numa::worker_pool::NumaWorkerPool::global()
                     .allocate_pinned_for_gpu(len, gpu_id)
-                    .map_err(StorageError::AllocationFailed)? as usize
+                {
+                    Ok(Some(ptr)) => {
+                        tracing::debug!(
+                            "Using NUMA-aware allocation for {} bytes on GPU {}",
+                            len,
+                            gpu_id
+                        );
+                        Some(ptr as usize)
+                    }
+                    Ok(None) => None, // NUMA node unknown, fall through
+                    Err(e) => return Err(StorageError::AllocationFailed(e)),
+                }
+            } else {
+                None
             }
-            _ => unsafe {
+        } else {
+            None
+        };
+
+        #[cfg(not(target_os = "linux"))]
+        let numa_ptr: Option<usize> = None;
+
+        let ptr = if let Some(ptr) = numa_ptr {
+            ptr
+        } else {
+            unsafe {
                 ctx.bind_to_thread().map_err(StorageError::Cuda)?;
 
                 let ptr = cudarc::driver::result::malloc_host(len, sys::CU_MEMHOSTALLOC_DEVICEMAP)
@@ -101,7 +106,7 @@ impl PinnedStorage {
                 assert!(len < isize::MAX as usize);
 
                 ptr as usize
-            },
+            }
         };
 
         Ok(Self { ptr, len, ctx })