Fix managed KV cache: use __cuda_array_interface__ instead of UntypedStorage.from_blob

UntypedStorage.from_blob was removed in PyTorch 2.11+. Use the
standard __cuda_array_interface__ protocol to wrap cudaMallocManaged
pointers into PyTorch tensors — this works across all PyTorch versions.

Also removed cudaMemAdvise calls — ctypes struct passing for
cudaMemLocation is broken on ARM64 (returns EINVAL). The advise hints
are optional; pages will page-fault to GPU on-demand regardless.

CPU memset (ctypes.memset) is still used instead of cudaMemset to
avoid forcing all pages into HBM during zeroing.

managed_alloc.cu

@@ -0,0 +1,96 @@
+// managed_alloc.cu - cudaMallocManaged allocator for PyTorch
+// Compile: nvcc -shared -o libmanaged_alloc.so managed_alloc.cu -Xcompiler -fPIC
+// Compatible with CUDA 13+ (uses cudaMemLocation API)
+//
+// Key design decisions for GH200 EGM:
+//   1. cudaMallocManaged → allocations can page-fault across HBM + EGM
+//   2. cudaMemAdviseSetPreferredLocation(GPU) → driver prefers keeping pages on GPU
+//   3. cudaMemAdviseSetAccessedBy(CPU) → CPU can access over C2C NVLink without
+//      triggering page migration back to system RAM (critical: prevents OOM)
+//   4. Selective prefetching — small allocations (model weights, <2 GiB)
+//      are prefetched to GPU so cuBLAS/cuDNN kernels can access them
+//      directly from HBM. Large allocations (KV cache blocks) stay in
+//      managed memory and page-fault on demand, since they're too large
+//      to fit in HBM and attention ops can tolerate page faults.
+#include <cuda_runtime.h>
+#include <stdio.h>
+
+extern "C" {
+
+// PyTorch pluggable allocator signature: void*(size_t, int, cudaStream_t)
+void* managed_malloc(size_t size, int device, cudaStream_t stream) {
+  void* ptr = nullptr;
+
+  // Set the device before allocating
+  cudaError_t err = cudaSetDevice(device);
+  if (err != cudaSuccess) {
+    fprintf(stderr, "[managed_alloc] cudaSetDevice(%d) failed: %s\n",
+            device, cudaGetErrorString(err));
+    return nullptr;
+  }
+
+  // Use cudaMallocManaged - this is the key: allocations can page-fault
+  // across HBM and LPDDR on GH200 with EGM enabled
+  err = cudaMallocManaged(&ptr, size, cudaMemAttachGlobal);
+  if (err != cudaSuccess) {
+    fprintf(stderr, "[managed_alloc] cudaMallocManaged failed: %s "
+                    "(size=%zu bytes / %.2f GiB)\n",
+            cudaGetErrorString(err), size, (double)size / (1024.0*1024.0*1024.0));
+    return nullptr;
+  }
+
+  // CUDA 13+ uses cudaMemLocation struct instead of int for device
+  cudaMemLocation gpu_loc;
+  gpu_loc.type = cudaMemLocationTypeDevice;
+  gpu_loc.id = device;
+
+  // Advise: prefer GPU placement. On GH200 with EGM, the hardware will
+  // migrate pages as needed, but the driver tries to keep them on GPU.
+  cudaMemAdvise(ptr, size, cudaMemAdviseSetPreferredLocation, gpu_loc);
+
+  // Advise: CPU will access this memory too. On GH200, this sets up
+  // remote mapping over C2C NVLink so CPU can read/write without
+  // triggering page migration back to system RAM. This is CRITICAL
+  // to prevent OOM on EGM systems where most system RAM was carved
+  // out for the GPU.
+  cudaMemLocation cpu_loc;
+  cpu_loc.type = cudaMemLocationTypeHost;
+  cpu_loc.id = cudaCpuDeviceId;
+  cudaMemAdvise(ptr, size, cudaMemAdviseSetAccessedBy, cpu_loc);
+
+  // Selective prefetch: migrate pages to GPU for small allocations only.
+  // Model weights (individual tensors) are typically <2 GiB and MUST be
+  // on GPU for cuBLAS GEMM operations — GPU compute kernels cannot
+  // page-fault into managed memory during execution.
+  // KV cache blocks are large and numerous; prefetching them all fills
+  // HBM and causes subsequent allocations to fail.
+  // The 2 GiB threshold separates "compute data" from "cache data".
+  const size_t PREFETCH_THRESHOLD = 2ULL * 1024 * 1024 * 1024; // 2 GiB
+
+  if (size > 0 && size < PREFETCH_THRESHOLD) {
+    err = cudaMemPrefetchAsync(ptr, size, gpu_loc, 0);
+    if (err != cudaSuccess) {
+      // Non-fatal: prefetch failure shouldn't prevent allocation.
+      // Pages will still be migrated on demand.
+      fprintf(stderr, "[managed_alloc] cudaMemPrefetchAsync warning: %s "
+                      "(size=%.2f GiB, will use on-demand migration)\n",
+              cudaGetErrorString(err), (double)size / (1024.0*1024.0*1024.0));
+    }
+  }
+
+  return ptr;
+}
+
+// PyTorch pluggable allocator signature: void(void*, size_t, int, cudaStream_t)
+void managed_free(void* ptr, size_t size, int device, cudaStream_t stream) {
+  if (ptr != nullptr) {
+    // Sync the stream before freeing to avoid use-after-free with
+    // managed memory (in-flight page faults can race with deallocation).
+    if (stream != nullptr) {
+      cudaStreamSynchronize(stream);
+    }
+    cudaFree(ptr);
+  }
+}
+
+} // extern "C"

vllm/envs.py

@@ -185,6 +185,7 @@ if TYPE_CHECKING:
     VLLM_MQ_MAX_CHUNK_BYTES_MB: int = 16
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
     VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
+    VLLM_KV_CACHE_USE_MANAGED_MEMORY: bool = False
     VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
     VLLM_USE_NVFP4_CT_EMULATIONS: bool = False
     VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
@@ -1378,6 +1379,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_KV_CACHE_LAYOUT": env_with_choices(
         "VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
     ),
+    # On GH200 with EGM, allocate KV cache via cudaMallocManaged so it
+    # can page-fault into LPDDR memory, enabling larger KV caches.
+    "VLLM_KV_CACHE_USE_MANAGED_MEMORY": lambda: bool(
+        int(os.getenv("VLLM_KV_CACHE_USE_MANAGED_MEMORY", "0"))
+    ),
     # Enable checking whether the generated logits contain NaNs,
     # indicating corrupted output. Useful for debugging low level bugs
     # or bad hardware but it may add compute overhead.

vllm/v1/worker/gpu_model_runner.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import ctypes
 import functools
 import gc
 import itertools
@@ -212,6 +213,81 @@ if TYPE_CHECKING:
 
 logger = init_logger(__name__)
 
+
+def _allocate_managed_kv_cache(size: int, device: torch.device) -> torch.Tensor:
+    """Allocate KV cache memory using cudaMallocManaged for GH200 EGM.
+
+    This allows the KV cache to transparently page-fault between HBM and
+    LPDDR, enabling much larger KV caches than HBM alone would allow.
+    Model weights and compute intermediates remain in HBM via the default
+    cudaMalloc — only the KV cache uses managed memory.
+
+    Key design decisions for KV cache (different from model weights):
+      - No cudaMemAdviseSetPreferredLocation(GPU). The KV cache is too
+        large for HBM (often 50-100+ GiB). Setting preferred location to
+        GPU would cause the driver to try migrating the entire allocation
+        to HBM, resulting in OOM. Pages page-fault to GPU on-demand during
+        attention operations and are evicted back to LPDDR when HBM is
+        needed.
+      - Zeroing is done via CPU memset, NOT cudaMemset. cudaMemset executes
+        on the device, which forces ALL pages to be migrated to GPU first —
+        exactly what we're trying to avoid. CPU memset leaves pages in LPDDR.
+      - Tensor wrapping uses __cuda_array_interface__ instead of the
+        deprecated UntypedStorage.from_blob (removed in PyTorch 2.11+).
+    """
+    cuda = ctypes.CDLL("libcudart.so")
+
+    # Allocate managed memory
+    ptr = ctypes.c_void_p()
+    err = cuda.cudaMallocManaged(ctypes.byref(ptr), ctypes.c_size_t(size),
+                                 ctypes.c_uint(1))  # cudaMemAttachGlobal
+    if err != 0:
+        cuda.cudaGetErrorString.restype = ctypes.c_char_p
+        err_str = cuda.cudaGetErrorString(err)
+        raise RuntimeError(
+            f"cudaMallocManaged failed for KV cache ({size} bytes / "
+            f"{size / 1024**3:.2f} GiB): {err_str.decode()}")
+
+    # Zero out the managed memory via CPU, NOT via cudaMemset.
+    # cudaMemset runs on the device, which forces ALL pages to be migrated
+    # to GPU before zeroing — defeating the entire purpose of keeping the
+    # KV cache in LPDDR. Using ctypes.memset (CPU) zeroes the pages while
+    # they remain in LPDDR/EGM. The pages will be lazily migrated to GPU
+    # only when the attention kernel actually reads them.
+    ctypes.memset(ptr, 0, size)
+
+    # Wrap the managed memory pointer as a PyTorch tensor using
+    # __cuda_array_interface__. This is the standard mechanism for creating
+    # tensors from external CUDA pointers, and works across PyTorch versions
+    # (UntypedStorage.from_blob was removed in PyTorch 2.11+).
+    # The tensor does NOT own the memory — we handle cudaFree ourselves
+    # when the KV cache is destroyed (process lifetime).
+    class _ManagedMemoryWrapper:
+        """Wrapper exposing __cuda_array_interface__ for a cudaMallocManaged pointer."""
+        def __init__(self, ptr_value, nbytes):
+            self._ptr = ptr_value
+            self._nbytes = nbytes
+
+        @property
+        def __cuda_array_interface__(self):
+            return {
+                "data": (self._ptr, False),  # (ptr, readonly)
+                "shape": (self._nbytes,),
+                "strides": None,
+                "typestr": "|i1",  # int8 bytes
+                "version": 3,
+            }
+
+    wrapper = _ManagedMemoryWrapper(ptr.value, size)
+    tensor = torch.as_tensor(wrapper, device=device)
+
+    logger.info("Allocated KV cache via cudaMallocManaged: %.2f GiB "
+                "(CPU memset, on-demand page-fault to GPU)",
+                size / 1024**3)
+
+    return tensor
+
+
 AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
 # list when ubatching is enabled
 PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
@@ -6526,9 +6602,17 @@ class GPUModelRunner(
         """
         kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
         for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
-            tensor = torch.zeros(
-                kv_cache_tensor.size, dtype=torch.int8, device=self.device
-            )
+            # On GH200 with EGM, allocate KV cache via cudaMallocManaged
+            # so it can spill into LPDDR memory. This is controlled by
+            # the VLLM_KV_CACHE_USE_MANAGED_MEMORY env var.
+            if envs.VLLM_KV_CACHE_USE_MANAGED_MEMORY:
+                tensor = _allocate_managed_kv_cache(
+                    kv_cache_tensor.size, self.device)
+            else:
+                tensor = torch.zeros(
+                    kv_cache_tensor.size, dtype=torch.int8,
+                    device=self.device
+                )
             for layer_name in kv_cache_tensor.shared_by:
                 kv_cache_raw_tensors[layer_name] = tensor