diff --git a/managed_alloc.cu b/managed_alloc.cu
index 2ae0ef595..53e11858e 100644
--- a/managed_alloc.cu
+++ b/managed_alloc.cu
@@ -7,10 +7,11 @@
 //   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. NO prefetching — pages migrate on-demand via hardware page faults.
-//      Eager prefetching fills HBM+EGM and causes subsequent allocations
-//      to fail. On-demand migration is the correct behavior for unified
-//      memory with HBM + LPDDR EGM.
+//   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>
 
@@ -57,19 +58,25 @@ void* managed_malloc(size_t size, int device, cudaStream_t stream) {
   cpu_loc.id = cudaCpuDeviceId;
   cudaMemAdvise(ptr, size, cudaMemAdviseSetAccessedBy, cpu_loc);
 
-  // REMOVED: cudaMemPrefetchAsync — was causing allocation failures after
-  // model loading. Prefetching eagerly migrates ALL pages to GPU, filling
-  // up HBM+EGM. Once physical memory is consumed by prefetched pages, the
-  // next cudaMallocManaged call fails because the driver can't guarantee
-  // page-fault resolution for new allocations.
-  //
-  // On GH200 with EGM, the hardware handles page faults naturally via C2C
-  // NVLink. The cudaMemAdviseSetPreferredLocation(GPU) hint above tells
-  // the driver to prefer GPU placement, but allows fallback to LPDDR when
-  // HBM is full. That's exactly what we want — don't force it.
-  //
-  // Pages will migrate on-demand as they're accessed, which is the correct
-  // behavior for a unified memory system with 96 GiB HBM + 128+ GiB EGM.
+  // 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;
 }