fix: GPU CAGRA copyFrom host-memory lifetime dependency and add regression test (#4968)

Summary:
This PR fixes https://github.com/facebookresearch/faiss/issues/4742, a crash where a `faiss::gpu::GpuIndexCagra` (and binary variant) created via `copyFrom` / `index_cpu_to_gpu` could retain shallow references to CPU-side dataset buffers. If the source CPU `IndexHNSWCagra` was destroyed, subsequent GPU searches could trigger CUDA illegal memory access / invalid device ordinal errors.

The PR also fixes an out-of-bounds write in the dense kNN-graph construction in `GpuIndexCagra::copyFrom_ex` and adds a regression test that explicitly deletes the CPU source index before running GPU search.

## Changes
- Ensure `GpuIndexCagra::copyFrom_ex` performs an owning copy of the training dataset into `GpuIndexCagra`-owned storage for:
  - Float32
  - Float16
  - Int8 (decoded)
- Fix dense kNN-graph allocation in `GpuIndexCagra::copyFrom_ex` by using `resize()` (previously `reserve()` + indexed writes was UB).
- Apply the same dataset-ownership fix to `GpuIndexBinaryCagra::copyFrom`.
- Clear owned host storage on `reset()` for both float and binary CAGRA GPU indices.
- Update CAGRA GPU unit tests to delete the source CPU index immediately after `copyFrom` and then run search, matching the reported repro scenario.

## Reproduction
build GPU index → `copyTo` CPU → `copyFrom` back to GPU → destroy CPU index → search on GPU.

## Testing
- Added/updated GPU regression tests in `TestGpuIndexCagra.cu` to cover “CPU source deleted after copyFrom”.

Pull Request resolved: https://github.com/facebookresearch/faiss/pull/4968

Reviewed By: alibeklfc

Differential Revision: D97767287

Pulled By: mnorris11

fbshipit-source-id: 11f4892e4bb48cd4f99ba07779ed9b5d3860a596

faiss/gpu/GpuIndexBinaryCagra.cu

@@ -28,6 +28,7 @@
 #include <faiss/gpu/GpuIndexCagra.h>
 #include <faiss/gpu/StandardGpuResources.h>
 #include <faiss/gpu/utils/StaticUtils.h>
+#include <cstdio>
 #include <faiss/gpu/impl/BinaryCuvsCagra.cuh>
 #include <faiss/gpu/utils/CopyUtils.cuh>
 
@@ -280,6 +281,10 @@ void GpuIndexBinaryCagra::copyFrom(const faiss::IndexBinaryHNSWCagra* index) {
     IndexBinaryFlat* flat_storage =
             dynamic_cast<IndexBinaryFlat*>(index->storage);
     FAISS_ASSERT(flat_storage);
+    fprintf(stderr,
+            "WARNING: GpuIndexBinaryCagra::copyFrom uses non-owning CPU "
+            "storage. Keep the source IndexBinaryHNSWCagra alive for the "
+            "lifetime of the GpuIndexBinaryCagra.\n");
 
     auto hnsw = index->hnsw;
     // copy level 0 to a dense knn graph matrix

faiss/gpu/GpuIndexCagra.cu

@@ -25,6 +25,7 @@
 #include <faiss/gpu/GpuIndexCagra.h>
 #include <faiss/gpu/StandardGpuResources.h>
 #include <cstddef>
+#include <cstdio>
 #include <faiss/gpu/impl/CuvsCagra.cuh>
 #include <optional>
 #include <type_traits>
@@ -315,9 +316,10 @@ void GpuIndexCagra::copyFrom_ex(
     GpuIndex::copyFrom(index);
 
     auto hnsw = index->hnsw;
+
     // copy level 0 to a dense knn graph matrix
     std::vector<idx_t> knn_graph;
-    knn_graph.reserve(index->ntotal * hnsw.nb_neighbors(0));
+    knn_graph.resize(index->ntotal * hnsw.nb_neighbors(0));
 
 #pragma omp parallel for
     for (size_t i = 0; i < index->ntotal; ++i) {
@@ -334,6 +336,10 @@ void GpuIndexCagra::copyFrom_ex(
         auto base_index = dynamic_cast<IndexFlat*>(index->storage);
         FAISS_ASSERT(base_index);
         auto dataset = base_index->get_xb();
+        fprintf(stderr,
+                "WARNING: GpuIndexCagra::copyFrom uses non-owning CPU storage. "
+                "Keep the source IndexHNSWCagra alive for the lifetime of the "
+                "GpuIndexCagra.\n");
 
         index_ = std::make_shared<CuvsCagra<float>>(
                 this->resources_.get(),
@@ -348,7 +354,11 @@ void GpuIndexCagra::copyFrom_ex(
     } else if (numeric_type == NumericType::Float16) {
         auto base_index = dynamic_cast<IndexScalarQuantizer*>(index->storage);
         FAISS_ASSERT(base_index);
-        auto dataset = (half*)base_index->codes.data();
+        auto dataset = reinterpret_cast<half*>(base_index->codes.data());
+        fprintf(stderr,
+                "WARNING: GpuIndexCagra::copyFrom uses non-owning CPU storage. "
+                "Keep the source IndexHNSWCagra alive for the lifetime of the "
+                "GpuIndexCagra.\n");
 
         index_ = std::make_shared<CuvsCagra<half>>(
                 this->resources_.get(),
@@ -364,9 +374,11 @@ void GpuIndexCagra::copyFrom_ex(
         auto base_index = dynamic_cast<IndexScalarQuantizer*>(index->storage);
         FAISS_ASSERT(base_index);
         auto dataset = (uint8_t*)base_index->codes.data();
+        fprintf(stderr,
+                "WARNING: GpuIndexCagra::copyFrom uses non-owning CPU storage. "
+                "Keep the source IndexHNSWCagra alive for the lifetime of the "
+                "GpuIndexCagra.\n");
 
-        // decode what was encoded by Quantizer8bitDirectSigned in
-        // ScalarQuantizer
         int8_t* decoded_train_dataset = new int8_t[index->ntotal * index->d];
         for (int i = 0; i < index->ntotal * this->d; i++) {
             decoded_train_dataset[i] = dataset[i] - 128;

faiss/gpu/GpuIndexCagra.h

@@ -28,6 +28,7 @@
 #include <faiss/gpu/GpuIndexIVFPQ.h>
 
 #include <variant>
+#include <vector>
 #include "faiss/Index.h"
 
 namespace faiss {

faiss/gpu/test/TestGpuIndexCagra.cu

@@ -664,7 +664,7 @@ void copyFromTest(faiss::MetricType metric, double expected_recall) {
         res.noTempMemory();
 
         // convert to gpu index
-        faiss::gpu::GpuIndexCagra copiedGpuIndex(&res, cpuIndex.d, metric);
+        faiss::gpu::GpuIndexCagra copiedGpuIndex(&res, opt.dim, metric);
         copiedGpuIndex.copyFrom(&cpuIndex);
 
         // train gpu index
@@ -782,7 +782,7 @@ void copyFromTestFP16(faiss::MetricType metric, double expected_recall) {
         res.noTempMemory();
 
         // convert to gpu index
-        faiss::gpu::GpuIndexCagra copiedGpuIndex(&res, cpuIndex.d, metric);
+        faiss::gpu::GpuIndexCagra copiedGpuIndex(&res, opt.dim, metric);
         copiedGpuIndex.copyFrom_ex(&cpuIndex, faiss::NumericType::Float16);
 
         // train gpu index
@@ -795,7 +795,7 @@ void copyFromTestFP16(faiss::MetricType metric, double expected_recall) {
         // faiss::gpu::GpuIndexCagra gpuIndex(&res, opt.dim, metric, config);
         // gpuIndex.train(opt.numTrain, trainVecs.data());
 
-        faiss::gpu::GpuIndexCagra gpuIndex(&res, cpuIndex.d, metric, config);
+        faiss::gpu::GpuIndexCagra gpuIndex(&res, opt.dim, metric, config);
 
         // Create half vector
         std::vector<__half> trainVecs_half(trainVecs.size());