Migrate from RAFT to CUVS

.github/actions/build_cmake/action.yml

@@ -8,8 +8,8 @@ inputs:
     description: 'Enable GPU support.'
     required: false
     default: OFF
-  raft:
-    description: 'Enable RAFT support.'
+  cuvs:
+    description: 'Enable cuVS support.'
     required: false
     default: OFF
   rocm:
@@ -50,11 +50,11 @@ runs:
         if [ "${{ inputs.rocm }}" = "ON" ]; then
           :
         # regular CUDA for GPU builds
-        elif [ "${{ inputs.gpu }}" = "ON" ] && [ "${{ inputs.raft }}" = "OFF" ]; then
+        elif [ "${{ inputs.gpu }}" = "ON" ] && [ "${{ inputs.cuvs }}" = "OFF" ]; then
           conda install -y -q cuda-toolkit -c "nvidia/label/cuda-12.4.0"
-        # and CUDA from RAFT channel for RAFT builds
-        elif [ "${{ inputs.raft }}" = "ON" ]; then
-          conda install -y -q libraft=24.06 cuda-version=12.4 cuda-toolkit -c rapidsai -c "nvidia/label/cuda-12.4.0" -c conda-forge
+        # and CUDA from cuVS channel for cuVS builds
+        elif [ "${{ inputs.cuvs }}" = "ON" ]; then
+          conda install -y -q libcuvs=24.08 cuda-version=12.4 cuda-toolkit -c rapidsai -c conda-forge -c "nvidia/label/cuda-12.4.0"
         fi
 
         # install test packages
@@ -102,7 +102,7 @@ runs:
         sudo apt-get -qq clean >/dev/null
         sudo rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
     - name: Symblink system dependencies
-      if: inputs.raft == 'ON' || inputs.rocm == 'ON'
+      if: inputs.rocm == 'ON'
       shell: bash
       run: |
         # symblink system libraries for HIP compiler
@@ -119,7 +119,7 @@ runs:
               -DBUILD_TESTING=ON \
               -DBUILD_SHARED_LIBS=ON \
               -DFAISS_ENABLE_GPU=${{ inputs.gpu }} \
-              -DFAISS_ENABLE_RAFT=${{ inputs.raft }} \
+              -DFAISS_ENABLE_CUVS=${{ inputs.cuvs }} \
               -DFAISS_ENABLE_ROCM=${{ inputs.rocm }} \
               -DFAISS_OPT_LEVEL=${{ inputs.opt_level }} \
               -DFAISS_ENABLE_C_API=ON \
@@ -174,5 +174,5 @@ runs:
       if: always()
       uses: actions/upload-artifact@v4
       with:
-        name: test-results-arch=${{ runner.arch }}-opt=${{ inputs.opt_level }}-gpu=${{ inputs.gpu }}-raft=${{ inputs.raft }}-rocm=${{ inputs.rocm }}
+        name: test-results-arch=${{ runner.arch }}-opt=${{ inputs.opt_level }}-gpu=${{ inputs.gpu }}-cuvs=${{ inputs.cuvs }}-rocm=${{ inputs.rocm }}
         path: test-results

.github/actions/build_conda/action.yml

@@ -9,8 +9,8 @@ inputs:
     description: "CUDA toolkit version to use."
     default: ""
     required: false
-  raft:
-    description: "Enable RAFT support."
+  cuvs:
+    description: "Enable cuVS support."
     default: ""
     required: false
 runs:
@@ -59,34 +59,34 @@ runs:
       run: |
         conda build faiss --user pytorch --label ${{ inputs.label }} -c pytorch
     - name: Conda build (GPU)
-      if: inputs.label == '' && inputs.cuda != '' && inputs.raft == ''
+      if: inputs.label == '' && inputs.cuda != '' && inputs.cuvs == ''
       shell: ${{ steps.choose_shell.outputs.shell }}
       working-directory: conda
       run: |
         conda build faiss-gpu --variants '{ "cudatoolkit": "${{ inputs.cuda }}" }' \
             -c pytorch -c nvidia/label/cuda-${{ inputs.cuda }} -c nvidia
     - name: Conda build (GPU) w/ anaconda upload
-      if: inputs.label != '' && inputs.cuda != '' && inputs.raft == ''
+      if: inputs.label != '' && inputs.cuda != '' && inputs.cuvs == ''
       shell: ${{ steps.choose_shell.outputs.shell }}
       working-directory: conda
       env:
         PACKAGE_TYPE: ${{ inputs.label }}
       run: |
         conda build faiss-gpu --variants '{ "cudatoolkit": "${{ inputs.cuda }}" }' \
             --user pytorch --label ${{ inputs.label }} -c pytorch -c nvidia/label/cuda-${{ inputs.cuda }} -c nvidia
-    - name: Conda build (GPU w/ RAFT)
-      if: inputs.label == '' && inputs.cuda != '' && inputs.raft != ''
+    - name: Conda build (GPU w/ cuVS)
+      if: inputs.label == '' && inputs.cuda != '' && inputs.cuvs != ''
       shell: ${{ steps.choose_shell.outputs.shell }}
       working-directory: conda
       run: |
-        conda build faiss-gpu-raft --variants '{ "cudatoolkit": "${{ inputs.cuda }}" }' \
+        conda build faiss-gpu-cuvs --variants '{ "cudatoolkit": "${{ inputs.cuda }}" }' \
             -c pytorch -c nvidia/label/cuda-${{ inputs.cuda }} -c nvidia -c rapidsai -c rapidsai-nightly -c conda-forge
-    - name: Conda build (GPU w/ RAFT) w/ anaconda upload
-      if: inputs.label != '' && inputs.cuda != '' && inputs.raft != ''
+    - name: Conda build (GPU w/ cuVS) w/ anaconda upload
+      if: inputs.label != '' && inputs.cuda != '' && inputs.cuvs != ''
       shell: ${{ steps.choose_shell.outputs.shell }}
       working-directory: conda
       env:
         PACKAGE_TYPE: ${{ inputs.label }}
       run: |
-        conda build faiss-gpu-raft --variants '{ "cudatoolkit": "${{ inputs.cuda }}" }' \
+        conda build faiss-gpu-cuvs --variants '{ "cudatoolkit": "${{ inputs.cuda }}" }' \
             --user pytorch --label ${{ inputs.label }} -c pytorch -c nvidia/label/cuda-${{ inputs.cuda }} -c nvidia -c rapidsai -c rapidsai-nightly -c conda-forge

.github/workflows/build.yml

@@ -78,8 +78,8 @@ jobs:
         uses: ./.github/actions/build_cmake
         with:
           gpu: ON
-  linux-x86_64-GPU-w-RAFT-cmake:
-    name: Linux x86_64 GPU w/ RAFT (cmake)
+  linux-x86_64-GPU-w-CUVS-cmake:
+    name: Linux x86_64 GPU w/ cuVS (cmake)
     needs: linux-x86_64-cmake
     runs-on: 4-core-ubuntu-gpu-t4
     steps:
@@ -89,7 +89,7 @@ jobs:
         uses: ./.github/actions/build_cmake
         with:
           gpu: ON
-          raft: ON
+          cuvs: ON
   linux-x86_64-GPU-w-ROCm-cmake:
     name: Linux x86_64 GPU w/ ROCm (cmake)
     needs: linux-x86_64-cmake
@@ -199,8 +199,8 @@ jobs:
         with:
           label: main
           cuda: "11.4.4"
-  linux-x86_64-GPU-RAFT-packages-CUDA11-8-0:
-    name: Linux x86_64 GPU w/ RAFT packages (CUDA 11.8.0)
+  linux-x86_64-GPU-CUVS-packages-CUDA11-8-0:
+    name: Linux x86_64 GPU w/ cuVS packages (CUDA 11.8.0)
     if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v')
     runs-on: 4-core-ubuntu-gpu-t4
     env:
@@ -217,7 +217,7 @@ jobs:
           ANACONDA_API_TOKEN: ${{ secrets.ANACONDA_API_TOKEN }}
         with:
           label: main
-          raft: "ON"
+          cuvs: "ON"
           cuda: "11.8.0"
   linux-x86_64-GPU-packages-CUDA-12-1-1:
     name: Linux x86_64 GPU packages (CUDA 12.1.1)
@@ -238,8 +238,8 @@ jobs:
         with:
           label: main
           cuda: "12.1.1"
-  linux-x86_64-GPU-RAFT-packages-CUDA12-1-1:
-    name: Linux x86_64 GPU w/ RAFT packages (CUDA 12.1.1)
+  linux-x86_64-GPU-CUVS-packages-CUDA12-1-1:
+    name: Linux x86_64 GPU w/ cuVS packages (CUDA 12.1.1)
     if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v')
     runs-on: 4-core-ubuntu-gpu-t4
     env:
@@ -256,7 +256,7 @@ jobs:
           ANACONDA_API_TOKEN: ${{ secrets.ANACONDA_API_TOKEN }}
         with:
           label: main
-          raft: "ON"
+          cuvs: "ON"
           cuda: "12.1.1"
   windows-x86_64-packages:
     name: Windows x86_64 packages

.github/workflows/nightly.yml

@@ -38,8 +38,8 @@ jobs:
         with:
           label: nightly
           cuda: "11.4.4"
-  linux-x86_64-GPU-RAFT-CUDA11-8-0-nightly:
-    name: Linux x86_64 GPU w/ RAFT nightlies (CUDA 11.8.0)
+  linux-x86_64-GPU-CUVS-CUDA11-8-0-nightly:
+    name: Linux x86_64 GPU w/ cuVS nightlies (CUDA 11.8.0)
     runs-on: 4-core-ubuntu-gpu-t4
     env:
       CUDA_ARCHS: "70-real;72-real;75-real;80;86-real"
@@ -54,7 +54,7 @@ jobs:
           ANACONDA_API_TOKEN: ${{ secrets.ANACONDA_API_TOKEN }}
         with:
           label: nightly
-          raft: "ON"
+          cuvs: "ON"
           cuda: "11.8.0"
   linux-x86_64-GPU-CUDA-12-1-1-nightly:
     name: Linux x86_64 GPU nightlies (CUDA 12.1.1)
@@ -73,8 +73,8 @@ jobs:
         with:
           label: nightly
           cuda: "12.1.1"
-  linux-x86_64-GPU-RAFT-CUDA12-1-1-nightly:
-    name: Linux x86_64 GPU w/ RAFT nightlies (CUDA 12.1.1)
+  linux-x86_64-GPU-CUVS-CUDA12-1-1-nightly:
+    name: Linux x86_64 GPU w/ cuVS nightlies (CUDA 12.1.1)
     runs-on: 4-core-ubuntu-gpu-t4
     env:
       CUDA_ARCHS: "70-real;72-real;75-real;80;86-real"
@@ -89,7 +89,7 @@ jobs:
           ANACONDA_API_TOKEN: ${{ secrets.ANACONDA_API_TOKEN }}
         with:
           label: nightly
-          raft: "ON"
+          cuvs: "ON"
           cuda: "12.1.1"
   windows-x86_64-nightly:
     name: Windows x86_64 nightlies

CMakeLists.txt

@@ -33,7 +33,7 @@ if(FAISS_ENABLE_GPU)
   endif()
 endif()
 
-if(FAISS_ENABLE_RAFT)
+if(FAISS_ENABLE_CUVS)
 include(cmake/thirdparty/fetch_rapids.cmake)
 include(rapids-cmake)
 include(rapids-cpm)
@@ -60,7 +60,7 @@ list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
 # Valid values are "generic", "avx2", "avx512", "sve".
 option(FAISS_OPT_LEVEL "" "generic")
 option(FAISS_ENABLE_GPU "Enable support for GPU indexes." ON)
-option(FAISS_ENABLE_RAFT "Enable RAFT for GPU indexes." OFF)
+option(FAISS_ENABLE_CUVS "Enable cuVS for GPU indexes." OFF)
 option(FAISS_ENABLE_ROCM "Enable ROCm for GPU indexes." OFF)
 option(FAISS_ENABLE_PYTHON "Build Python extension." ON)
 option(FAISS_ENABLE_C_API "Build C API." OFF)
@@ -81,9 +81,9 @@ if(FAISS_ENABLE_GPU)
   endif()
 endif()
 
-if(FAISS_ENABLE_RAFT AND NOT TARGET raft::raft)
-  find_package(raft COMPONENTS compiled distributed)
-endif()
+if(FAISS_ENABLE_CUVS AND NOT TARGET cuvs::cuvs)
+   find_package(cuvs)
+ endif()
 
 add_subdirectory(faiss)
 

INSTALL.md

@@ -118,7 +118,7 @@ Several options can be passed to CMake, among which:
   values are `ON` and `OFF`),
   - `-DFAISS_ENABLE_PYTHON=OFF` in order to disable building python bindings
   (possible values are `ON` and `OFF`),
-  - `-DFAISS_ENABLE_RAFT=ON` in order to enable building the RAFT implementations
+  - `-DFAISS_ENABLE_CUVS=ON` in order to enable building the cuVS implementations
     of the IVF-Flat and IVF-PQ GPU-accelerated indices (default is `OFF`, possible
     values are `ON` and `OFF`)
   - `-DBUILD_TESTING=OFF` in order to disable building C++ tests,

benchs/bench_ivfflat_raft.py → benchs/bench_ivfflat_cuvs.py

@@ -4,7 +4,7 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 #
-# Copyright (c) 2023, NVIDIA CORPORATION.
+# Copyright (c) 2024, NVIDIA CORPORATION.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -44,8 +44,8 @@ def aa(*args, **kwargs):
    help='whether to benchmark add operation on GPU index')
 aa('--bm_search', default=True,
    help='whether to benchmark search operation on GPU index')
-aa('--raft_only', default=False, action='store_true',
-   help='whether to only produce RAFT enabled benchmarks')
+aa('--cuvs_only', default=False, action='store_true',
+   help='whether to only produce cuVS enabled benchmarks')
 
 
 group = parser.add_argument_group('IVF options')
@@ -70,9 +70,9 @@ def aa(*args, **kwargs):
 mr = rmm.mr.PoolMemoryResource(rmm.mr.CudaMemoryResource())
 rmm.mr.set_current_device_resource(mr)
 
-def bench_train_milliseconds(index, trainVecs, use_raft):
+def bench_train_milliseconds(index, trainVecs, use_cuvs):
     co = faiss.GpuMultipleClonerOptions()
-    co.use_raft = use_raft
+    co.use_cuvs = use_cuvs
     index_gpu = faiss.index_cpu_to_gpu(res, 0, index, co)
     t0 = time.time()
     index_gpu.train(trainVecs)
@@ -89,21 +89,21 @@ def bench_train_milliseconds(index, trainVecs, use_raft):
         for n_cols in dataset_dims:
             index = faiss.index_factory(n_cols, "IVF{},Flat".format(args.n_centroids))
             trainVecs = rs.rand(n_rows, n_cols).astype('float32')
-            raft_gpu_train_time = bench_train_milliseconds(
+            cuvs_gpu_train_time = bench_train_milliseconds(
                 index, trainVecs, True)
-            if args.raft_only:
-                print("Method: IVFFlat, Operation: TRAIN, dim: %d, n_centroids %d, numTrain: %d, RAFT enabled GPU train time: %.3f milliseconds" % (
-                    n_cols, args.n_centroids, n_rows, raft_gpu_train_time))
+            if args.cuvs_only:
+                print("Method: IVFFlat, Operation: TRAIN, dim: %d, n_centroids %d, numTrain: %d, cuVS enabled GPU train time: %.3f milliseconds" % (
+                    n_cols, args.n_centroids, n_rows, cuvs_gpu_train_time))
             else:
                 classical_gpu_train_time = bench_train_milliseconds(
                     index, trainVecs, False)
-                print("Method: IVFFlat, Operation: TRAIN, dim: %d, n_centroids %d, numTrain: %d, classical GPU train time: %.3f milliseconds, RAFT enabled GPU train time: %.3f milliseconds" % (
-                    n_cols, args.n_centroids, n_rows, classical_gpu_train_time, raft_gpu_train_time))
+                print("Method: IVFFlat, Operation: TRAIN, dim: %d, n_centroids %d, numTrain: %d, classical GPU train time: %.3f milliseconds, cuVS enabled GPU train time: %.3f milliseconds" % (
+                    n_cols, args.n_centroids, n_rows, classical_gpu_train_time, cuvs_gpu_train_time))
 
 
-def bench_add_milliseconds(index, addVecs, use_raft):
+def bench_add_milliseconds(index, addVecs, use_cuvs):
     co = faiss.GpuMultipleClonerOptions()
-    co.use_raft = use_raft
+    co.use_cuvs = use_cuvs
     index_gpu = faiss.index_cpu_to_gpu(res, 0, index, co)
     index_gpu.copyFrom(index)
     t0 = time.time()
@@ -125,20 +125,20 @@ def bench_add_milliseconds(index, addVecs, use_raft):
     for n_rows in addset_sizes:
         for n_cols in dataset_dims:
             addVecs = rs.rand(n_rows, n_cols).astype('float32')
-            raft_gpu_add_time = bench_add_milliseconds(index, addVecs, True)
-            if args.raft_only:
-                print("Method: IVFFlat, Operation: ADD, dim: %d, n_centroids %d, numAdd: %d, RAFT enabled GPU add time: %.3f milliseconds" % (
-                    n_train, n_rows, n_cols, args.n_centroids, raft_gpu_add_time))
+            cuvs_gpu_add_time = bench_add_milliseconds(index, addVecs, True)
+            if args.cuvs_only:
+                print("Method: IVFFlat, Operation: ADD, dim: %d, n_centroids %d, numAdd: %d, cuVS enabled GPU add time: %.3f milliseconds" % (
+                    n_train, n_rows, n_cols, args.n_centroids, cuvs_gpu_add_time))
             else:
                 classical_gpu_add_time = bench_add_milliseconds(
                     index, addVecs, False)
-                print("Method: IVFFlat, Operation: ADD, dim: %d, n_centroids %d, numAdd: %d, classical GPU add time: %.3f milliseconds, RAFT enabled GPU add time: %.3f milliseconds" % (
-                    n_train, n_rows, n_cols, args.n_centroids, classical_gpu_add_time, raft_gpu_add_time))
+                print("Method: IVFFlat, Operation: ADD, dim: %d, n_centroids %d, numAdd: %d, classical GPU add time: %.3f milliseconds, cuVS enabled GPU add time: %.3f milliseconds" % (
+                    n_train, n_rows, n_cols, args.n_centroids, classical_gpu_add_time, cuvs_gpu_add_time))
 
 
-def bench_search_milliseconds(index, addVecs, queryVecs, nprobe, k, use_raft):
+def bench_search_milliseconds(index, addVecs, queryVecs, nprobe, k, use_cuvs):
     co = faiss.GpuMultipleClonerOptions()
-    co.use_raft = use_raft
+    co.use_cuvs = use_cuvs
     index_gpu = faiss.index_cpu_to_gpu(res, 0, index, co)
     index_gpu.copyFrom(index)
     index_gpu.add(addVecs)
@@ -163,19 +163,19 @@ def bench_search_milliseconds(index, addVecs, queryVecs, nprobe, k, use_raft):
         addVecs = rs.rand(n_add, n_cols).astype('float32')
         for n_rows in queryset_sizes:
             queryVecs = rs.rand(n_rows, n_cols).astype('float32')
-            raft_gpu_search_time = bench_search_milliseconds(
+            cuvs_gpu_search_time = bench_search_milliseconds(
                 index, addVecs, queryVecs, args.nprobe, args.k, True)
-            if args.raft_only:
-                print("Method: IVFFlat, Operation: SEARCH, dim: %d, n_centroids: %d, numVecs: %d, numQuery: %d, nprobe: %d, k: %d, RAFT enabled GPU search time: %.3f milliseconds" % (
-                    n_cols, args.n_centroids, n_add, n_rows, args.nprobe, args.k, raft_gpu_search_time))
+            if args.cuvs_only:
+                print("Method: IVFFlat, Operation: SEARCH, dim: %d, n_centroids: %d, numVecs: %d, numQuery: %d, nprobe: %d, k: %d, cuVS enabled GPU search time: %.3f milliseconds" % (
+                    n_cols, args.n_centroids, n_add, n_rows, args.nprobe, args.k, cuvs_gpu_search_time))
             else:
                 classical_gpu_search_time = bench_search_milliseconds(
                     index, addVecs, queryVecs, args.nprobe, args.k, False)
-                print("Method: IVFFlat, Operation: SEARCH, dim: %d, n_centroids: %d, numVecs: %d, numQuery: %d, nprobe: %d, k: %d, classical GPU search time: %.3f milliseconds, RAFT enabled GPU search time: %.3f milliseconds" % (
-                    n_cols, args.n_centroids, n_add, n_rows, args.nprobe, args.k, classical_gpu_search_time, raft_gpu_search_time))
+                print("Method: IVFFlat, Operation: SEARCH, dim: %d, n_centroids: %d, numVecs: %d, numQuery: %d, nprobe: %d, k: %d, classical GPU search time: %.3f milliseconds, cuVS enabled GPU search time: %.3f milliseconds" % (
+                    n_cols, args.n_centroids, n_add, n_rows, args.nprobe, args.k, classical_gpu_search_time, cuvs_gpu_search_time))
 
     print("=" * 40)
-    print("Large RAFT Enabled Benchmarks")
+    print("Large cuVS Enabled Benchmarks")
     print("=" * 40)
     # Avoid classical GPU Benchmarks for large datasets because of OOM for more than 500000 queries and/or large dims as well as for large k
     queryset_sizes = [100000, 500000, 1000000]
@@ -188,7 +188,7 @@ def bench_search_milliseconds(index, addVecs, queryVecs, nprobe, k, use_raft):
         addVecs = rs.rand(n_add, n_cols).astype('float32')
         for n_rows in queryset_sizes:
             queryVecs = rs.rand(n_rows, n_cols).astype('float32')
-            raft_gpu_search_time = bench_search_milliseconds(
+            cuvs_gpu_search_time = bench_search_milliseconds(
                 index, addVecs, queryVecs, args.nprobe, args.k, True)
-            print("Method: IVFFlat, Operation: SEARCH, numTrain: %d, dim: %d, n_centroids: %d, numVecs: %d, numQuery: %d, nprobe: %d, k: %d, RAFT enabled GPU search time: %.3f milliseconds" % (
-                n_cols, args.n_centroids, n_add, n_rows, args.nprobe, args.k, raft_gpu_search_time))
+            print("Method: IVFFlat, Operation: SEARCH, numTrain: %d, dim: %d, n_centroids: %d, numVecs: %d, numQuery: %d, nprobe: %d, k: %d, cuVS enabled GPU search time: %.3f milliseconds" % (
+                n_cols, args.n_centroids, n_add, n_rows, args.nprobe, args.k, cuvs_gpu_search_time))