fix(gpu): latency multi-group

Bumps [docker/login-action](https://github.com/docker/login-action) from 4.0.0 to 4.1.0.
- [Release notes](https://github.com/docker/login-action/releases)
- [Commits](b45d80f862...4907a6ddec)

---
updated-dependencies:
- dependency-name: docker/login-action
  dependency-version: 4.1.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

.cargo/audit.toml

@@ -4,6 +4,9 @@ ignore = [
     "RUSTSEC-2024-0436",
     # Ignoring unmaintained 'bincode' crate. Getting rid of it would be too complex on the short term.
     "RUSTSEC-2025-0141",
+    # Ignoring unsoundness in 'rand' with custom logger. Rand update is currently blocked by
+    # arkworks and we do not use custom loggers.
+    "RUSTSEC-2026-0097",
 ]
 
 [output]

.github/workflows/aws_data_tests.yml

@@ -1,5 +1,5 @@
-# Run backward compatibility tests
-name: aws_tfhe_backward_compat_tests
+# Run data related tests
+name: aws_data_tests
 
 env:
   CARGO_TERM_COLOR: always
@@ -30,8 +30,8 @@ permissions:
 # zizmor: ignore[concurrency-limits] concurrency is managed after instance setup to ensure safe provisioning
 
 jobs:
-  backward-compat-tests:
-    name: aws_tfhe_backward_compat_tests/backward-compat-tests (bpr)
+  data-tests:
+    name: aws_data_tests/data-tests (bpr)
     if: (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') ||
       github.event_name != 'push'
     runs-on: "runs-on=${{ github.run_id }}/runner=cpu-small"
@@ -49,22 +49,24 @@ jobs:
       - name: Get LFS data sha
         id: hash-lfs-data
         run: |
-          SHA=$(git lfs ls-files -l -I utils/tfhe-backward-compat-data | sha256sum | cut -d' ' -f1)
+          SHA=$(git lfs ls-files -l -I utils/tfhe-backward-compat-data,tests/corrupted_inputs_deserialization | sha256sum | cut -d' ' -f1)
           echo "sha=${SHA}" >> "${GITHUB_OUTPUT}"
 
       - name: Retrieve data from cache
         id: retrieve-data-cache
-        uses: actions/cache/restore@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/restore@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: |
             utils/tfhe-backward-compat-data/**/*.cbor
             utils/tfhe-backward-compat-data/**/*.bcode
+            tests/corrupted_inputs_deserialization/**/*.bcode
           key: ${{ steps.hash-lfs-data.outputs.sha }}
 
       - name: Pull test data
         if: steps.retrieve-data-cache.outputs.cache-hit != 'true'
         run: |
           make pull_backward_compat_data
+          make pull_corrupted_inputs_data
 
       # Pull token was stored by action/checkout to be used by lfs, we don't need it anymore
       - name: Remove git credentials
@@ -80,14 +82,19 @@ jobs:
         run: |
           make test_backward_compatibility_ci
 
+      - name: Run corrupted inputs deserialization tests
+        run: |
+          make test_corrupted_inputs_ci
+
       - name: Store data in cache
         if: steps.retrieve-data-cache.outputs.cache-hit != 'true'
         continue-on-error: true
-        uses: actions/cache/save@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/save@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: |
             utils/tfhe-backward-compat-data/**/*.cbor
             utils/tfhe-backward-compat-data/**/*.bcode
+            tests/corrupted_inputs_deserialization/**/*.bcode
           key: ${{ steps.hash-lfs-data.outputs.sha }}
 
       - name: Set pull-request URL

.github/workflows/aws_tfhe_fast_tests.yml

@@ -16,7 +16,6 @@ env:
   PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
 
-
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
@@ -37,6 +36,7 @@ jobs:
       csprng_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.csprng_any_changed }}
       zk_pok_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.zk_pok_any_changed }}
       versionable_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.versionable_any_changed }}
+      safe_serialize_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.safe_serialize_any_changed }}
       core_crypto_test: ${{ env.IS_PULL_REQUEST == 'false' ||
         steps.changed-files.outputs.core_crypto_any_changed ||
         steps.changed-files.outputs.dependencies_any_changed }}
@@ -64,7 +64,7 @@ jobs:
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
         with:
           fetch-depth: 0
-          persist-credentials: 'false'
+          persist-credentials: "false"
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Check for file changes
@@ -79,6 +79,7 @@ jobs:
               - tfhe-zk-pok/**
               - utils/tfhe-versionable/**
               - utils/tfhe-versionable-derive/**
+              - utils/tfhe-safe-serialize/**
             csprng:
               - tfhe-csprng/**
             zk_pok:
@@ -86,6 +87,8 @@ jobs:
             versionable:
               - utils/tfhe-versionable/**
               - utils/tfhe-versionable-derive/**
+            safe_serialize:
+              - utils/tfhe-safe-serialize/**
             core_crypto:
               - tfhe/src/core_crypto/**
             boolean:
@@ -122,6 +125,7 @@ jobs:
           steps.changed-files.outputs.csprng_any_changed == 'true' ||
           steps.changed-files.outputs.zk_pok_any_changed == 'true' ||
           steps.changed-files.outputs.versionable_any_changed == 'true' ||
+          steps.changed-files.outputs.safe_serialize_any_changed == 'true' ||
           steps.changed-files.outputs.core_crypto_any_changed == 'true' ||
           steps.changed-files.outputs.boolean_any_changed == 'true' ||
           steps.changed-files.outputs.shortint_any_changed == 'true' ||
@@ -145,7 +149,7 @@ jobs:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
         with:
-          persist-credentials: 'false'
+          persist-credentials: "false"
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
@@ -170,6 +174,11 @@ jobs:
         run: |
           make test_versionable
 
+      - name: Run tfhe-safe-serialize tests
+        if: needs.should-run.outputs.safe_serialize_test == 'true'
+        run: |
+          make test_safe_serialize
+
       - name: Run core tests
         if: needs.should-run.outputs.core_crypto_test == 'true'
         run: |
@@ -191,7 +200,7 @@ jobs:
 
       - name: Node cache restoration
         id: node-cache
-        uses: actions/cache/restore@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/restore@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: |
             ~/.nvm
@@ -204,7 +213,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/save@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         if: steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |

.github/workflows/aws_tfhe_noise_checks.yml

@@ -34,7 +34,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -99,7 +99,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/aws_tfhe_wasm_tests.yml

@@ -14,12 +14,11 @@ env:
   PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
 
-
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
-    types: [ labeled ]
+    types: [labeled]
 
 permissions:
   contents: read
@@ -32,16 +31,16 @@ jobs:
     if: github.event_name == 'workflow_dispatch' || contains(github.event.label.name, 'approved')
     runs-on: ubuntu-latest
     permissions:
-      pull-requests: read  # Needed to check for file change
+      pull-requests: read # Needed to check for file change
     outputs:
       wasm_test: ${{ github.event_name == 'workflow_dispatch' ||
-          steps.changed-files.outputs.wasm_any_changed }}
+        steps.changed-files.outputs.wasm_any_changed }}
     steps:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
         with:
           fetch-depth: 0
-          persist-credentials: 'false'
+          persist-credentials: "false"
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Check for file changes
@@ -63,6 +62,7 @@ jobs:
                 - tfhe/js_on_wasm_tests/**
                 - tfhe/web_wasm_parallel_tests/**
                 - utils/tfhe-versionable/**
+                - utils/tfhe-safe-serialize/**
                 - .github/workflows/aws_tfhe_wasm_tests.yml
 
   wasm-tests:
@@ -78,7 +78,7 @@ jobs:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
         with:
-          persist-credentials: 'false'
+          persist-credentials: "false"
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
@@ -92,7 +92,7 @@ jobs:
 
       - name: Node cache restoration
         id: node-cache
-        uses: actions/cache/restore@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/restore@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: |
             ~/.nvm
@@ -105,7 +105,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/save@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         if: steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |
@@ -128,15 +128,21 @@ jobs:
         run: |
           make test_nodejs_wasm_api_ci
 
-      - name: Run parallel wasm tests
-        run: |
-          make test_web_js_api_parallel_chrome_ci
-
       - name: Run wasm_par_mq tests
         run: |
           make test_wasm_par_mq_chrome_ci
           make test_wasm_par_mq_firefox_ci
 
+      - name: Run parallel wasm tests
+        run: |
+          make test_web_js_api_parallel_chrome_ci
+          make test_web_js_api_parallel_firefox_ci
+
+      - name: Run cross origin wasm tests
+        run: |
+          make test_web_js_api_cross_origin_chrome_ci
+          make test_web_js_api_cross_origin_firefox_ci
+
       - name: Run x86_64/wasm zk compatibility tests
         run: |
           make test_zk_wasm_x86_compat_ci

.github/workflows/backward_compat_pr_change_report.yml

@@ -5,8 +5,9 @@ name: backward_compat_pr_change_report
 
 on:
   pull_request:
-    paths:
-      - 'utils/tfhe-lints/snapshots/lint_enum_snapshots_*.json'
+
+env:
+  CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
 
 permissions:
   contents: read
@@ -16,9 +17,35 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  should-run:
+    name: backward_compat_pr_change_report/should-run
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: read  # Needed to check for file change
+    outputs:
+      backward_report: ${{ steps.changed-files.outputs.backward_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
+        with:
+          fetch-depth: 0
+          persist-credentials: 'false'
+          token: ${{ env.CHECKOUT_TOKEN }}
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@22103cc46bda19c2b464ffe86db46df6922fd323 # v47.0.5
+        with:
+          files_yaml: |
+            backward:
+              - utils/tfhe-lints/snapshots/*.json
+
   change-report:
     name: backward_compat_pr_change_report/change-report (bpr)
     runs-on: ubuntu-latest
+    needs: should-run
+    if:
+      needs.should-run.outputs.backward_report == 'true'
     permissions:
       pull-requests: write # To send and modify message in the PR
     steps:
@@ -53,8 +80,9 @@ jobs:
           fi
 
       - name: Find existing comment
+        if: steps.report.outputs.has_report == 'true'
         id: find-comment
-        uses: peter-evans/find-comment@3eae4d37986fb5a8592848f6a574fdf654e61f9e # v3.1.0
+        uses: peter-evans/find-comment@b30e6a3c0ed37e7c023ccd3f1db5c6c0b0c23aad # v4.0.0
         with:
           issue-number: ${{ github.event.pull_request.number }}
           body-includes: '**Backward-compat snapshot:'

.github/workflows/benchmark_cpu.yml

@@ -14,11 +14,12 @@ on:
           - signed_integer
           - integer_compression
           - integer_zk
+          - msm_zk
           - shortint
           - shortint_oprf
           - hlapi_unsigned
           - hlapi_signed
-          - hlapi_erc20
+          - hlapi_erc7984
           - hlapi_dex
           - hlapi_noise_squash
           - hlapi_kvstore
@@ -92,8 +93,8 @@ jobs:
 
           if inputs_command == "integer_zk":
             files_to_parse.append("pke_zk_crs_sizes.csv")
-          elif inputs_command == "hlapi_erc20":
-            files_to_parse.append("erc20_pbs_count.csv")
+          elif inputs_command == "hlapi_erc7984":
+            files_to_parse.append("erc7984_pbs_count.csv")
           elif inputs_command == "hlapi_dex":
             files_to_parse.extend(
               [

.github/workflows/benchmark_cpu_common.yml

@@ -107,7 +107,7 @@ jobs:
             ]:
               f.write(f"""{env_name}=["{'", "'.join(values_to_join)}"]\n""")
 
-      - name: Set martix arguments outputs
+      - name: Set matrix arguments outputs
         id: set_matrix_args
         run: | # zizmor: ignore[template-injection] these env variable are safe
           {
@@ -126,7 +126,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -261,7 +261,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_cpu_weekly.yml

@@ -108,14 +108,14 @@ jobs:
       SLAB_URL: ${{ secrets.SLAB_URL }}
       SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
 
-  run-benchmarks-hlapi-erc20:
-    name: benchmark_cpu_weekly/run-benchmarks-hlapi-erc20
+  run-benchmarks-hlapi-erc7984:
+    name: benchmark_cpu_weekly/run-benchmarks-hlapi-erc7984
     if: needs.prepare-inputs.outputs.is_weekly_bench_group_2 == 'true'
     needs: prepare-inputs
     uses: ./.github/workflows/benchmark_cpu_common.yml
     with:
-      command: hlapi_erc20
-      additional_file_to_parse: erc20_pbs_count.csv
+      command: hlapi_erc7984
+      additional_file_to_parse: erc7984_pbs_count.csv
     secrets:
       BOT_USERNAME: ${{ secrets.BOT_USERNAME }}
       SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}

.github/workflows/benchmark_ct_key_sizes.yml

@@ -33,7 +33,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -137,7 +137,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_documentation.yml

@@ -17,6 +17,10 @@ on:
         description: "Run GPU core-crypto benchmarks"
         type: boolean
         default: true
+      run-gpu-zk-benchmarks:
+        description: "Run GPU ZK benchmarks"
+        type: boolean
+        default: true
       run-hpu-benchmarks:
         description: "Run HPU benchmarks"
         type: boolean
@@ -36,7 +40,7 @@ jobs:
     uses: ./.github/workflows/benchmark_cpu_common.yml
     if: inputs.run-cpu-benchmarks
     with:
-      command: integer,hlapi_erc20
+      command: integer,hlapi_erc7984
       op_flavor: fast_default
       bench_type: both
       precisions_set: documentation
@@ -91,7 +95,7 @@ jobs:
     with:
       profile: multi-h100-sxm5
       hardware_name: n3-H100-SXM5x8
-      command: integer_multi_bit,hlapi_erc20
+      command: integer_multi_bit,hlapi_erc7984
       op_flavor: fast_default
       bench_type: both
       precisions_set: documentation
@@ -110,7 +114,7 @@ jobs:
     uses: ./.github/workflows/benchmark_hpu_common.yml
     if: inputs.run-hpu-benchmarks
     with:
-      command: integer,hlapi_erc20
+      command: integer,hlapi_erc7984
       op_flavor: default
       bench_type: both
       precisions_set: documentation
@@ -165,21 +169,42 @@ jobs:
       SLAB_URL: ${{ secrets.SLAB_URL }}
       SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
 
+  run-benchmarks-gpu-zk-server:
+    name: benchmark_documentation/run-benchmarks-gpu-zk-server
+    uses: ./.github/workflows/benchmark_gpu_common.yml
+    if: inputs.run-gpu-zk-benchmarks
+    with:
+      profile: multi-h100-sxm5
+      hardware_name: n3-H100-SXM5x8
+      command: integer_zk
+      op_flavor: default
+      bench_type: both
+    secrets:
+      BOT_USERNAME: ${{ secrets.BOT_USERNAME }}
+      SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+      SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+      REPO_CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN }}
+      JOB_SECRET: ${{ secrets.JOB_SECRET }}
+      SLAB_ACTION_TOKEN: ${{ secrets.SLAB_ACTION_TOKEN }}
+      SLAB_URL: ${{ secrets.SLAB_URL }}
+      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
+
   generate-svgs-with-benchmarks-run:
     name: benchmark-documentation/generate-svgs-with-benchmarks-run
     if: ${{ always() &&
-      (inputs.run-cpu-benchmarks || inputs.run-gpu-integer-benchmarks || inputs.run-gpu-core-crypto-benchmarks ||inputs.run-hpu-benchmarks) &&
+      (inputs.run-cpu-benchmarks || inputs.run-gpu-integer-benchmarks || inputs.run-gpu-core-crypto-benchmarks || inputs.run-gpu-zk-benchmarks || inputs.run-hpu-benchmarks) &&
       inputs.generate-svgs }}
     needs: [
       run-benchmarks-cpu-integer, run-benchmarks-gpu-integer, run-benchmarks-hpu-integer,
       run-benchmarks-cpu-zk-server, run-benchmarks-cpu-zk-client,
-      run-benchmarks-cpu-core-crypto, run-benchmarks-gpu-core-crypto
+      run-benchmarks-cpu-core-crypto, run-benchmarks-gpu-core-crypto,
+      run-benchmarks-gpu-zk-server
     ]
     uses: ./.github/workflows/generate_svgs.yml
     with:
       time_span_days: 5
       generate-cpu-svgs: ${{ inputs.run-cpu-benchmarks }}
-      generate-gpu-svgs: ${{ inputs.run-gpu-integer-benchmarks || inputs.run-gpu-core-crypto-benchmarks }}
+      generate-gpu-svgs: ${{ inputs.run-gpu-integer-benchmarks || inputs.run-gpu-core-crypto-benchmarks || inputs.run-gpu-zk-benchmarks }}
       generate-hpu-svgs: ${{ inputs.run-hpu-benchmarks }}
     secrets:
       DATA_EXTRACTOR_DATABASE_USER: ${{ secrets.DATA_EXTRACTOR_DATABASE_USER }}
@@ -188,7 +213,7 @@ jobs:
 
   generate-svgs-without-benchmarks-run:
     name: benchmark-documentation/generate-svgs-without-benchmarks-run
-    if: ${{ !(inputs.run-cpu-benchmarks || inputs.run-gpu-integer-benchmarks || inputs.run-gpu-core-crypto-benchmarks || inputs.run-hpu-benchmarks) &&
+    if: ${{ !(inputs.run-cpu-benchmarks || inputs.run-gpu-integer-benchmarks || inputs.run-gpu-core-crypto-benchmarks || inputs.run-gpu-zk-benchmarks || inputs.run-hpu-benchmarks) &&
       inputs.generate-svgs }}
     uses: ./.github/workflows/generate_svgs.yml
     with:

.github/workflows/benchmark_gpu.yml

@@ -31,10 +31,14 @@ on:
           - pbs128
           - ks
           - ks_pbs
+          - tfhe_zk_pok
+          - msm_zk
           - integer_zk
+          - integer_zk_experimental
           - integer_aes
           - integer_aes256
-          - hlapi_erc20
+          - hlapi_erc7984
+          - hlapi_erc7984_multi_group
           - hlapi_dex
           - hlapi_noise_squash
       op_flavor:
@@ -120,8 +124,8 @@ jobs:
 
           if inputs_command == "integer_zk":
             files_to_parse.append("pke_zk_crs_sizes.csv")
-          elif inputs_command == "hlapi_erc20":
-            files_to_parse.append("erc20_pbs_count.csv")
+          elif inputs_command == "hlapi_erc7984":
+            files_to_parse.append("erc7984_pbs_count.csv")
           elif inputs_command == "hlapi_dex":
             files_to_parse.extend(
               [

.github/workflows/benchmark_gpu_common.yml

@@ -111,7 +111,7 @@ jobs:
             ]:
               f.write(f"""{env_name}=["{'", "'.join(values_to_join)}"]\n""")
 
-      - name: Set martix arguments outputs
+      - name: Set matrix arguments outputs
         id: set_matrix_args
         run: | # zizmor: ignore[template-injection] these env variable are safe
           {
@@ -126,17 +126,11 @@ jobs:
     needs: prepare-matrix
     runs-on: ubuntu-latest
     outputs:
-      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
-      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
-      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
-      # otherwise we'll try to run the next job on a non-existing on-demand instance.
-      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
-      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
+      runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
-      - name: Start remote instance
-        id: start-remote-instance
-        continue-on-error: true
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -145,25 +139,6 @@ jobs:
           backend: ${{ inputs.backend }}
           profile: ${{ inputs.profile }}
 
-      - name: Acknowledge remote instance failure
-        if: steps.start-remote-instance.outcome == 'failure' &&
-          inputs.profile != 'single-h100'
-        run: |
-          echo "Remote instance instance has failed to start (profile provided: '${INPUTS_PROFILE}')"
-          echo "Permanent instance instance cannot be used as a substitute (profile needed: 'single-h100')"
-          exit 1
-        env:
-          INPUTS_PROFILE: ${{ inputs.profile }}
-
-      # This will allow to fallback on permanent instances running on Hyperstack.
-      - name: Use permanent remote instance
-        id: use-permanent-instance
-        if: env.SECRETS_AVAILABLE == 'true' &&
-          steps.start-remote-instance.outcome == 'failure' &&
-          inputs.profile == 'single-h100'
-        run: |
-          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
-
   # Install dependencies only once since cuda-benchmarks uses a matrix strategy, thus running multiple times.
   install-dependencies:
     name: benchmark_gpu_common/install-dependencies
@@ -184,7 +159,6 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
-        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -263,8 +237,10 @@ jobs:
           BENCH_PARAMS_TYPE: ${{ matrix.params_type }}
           BENCH_COMMAND: ${{ matrix.command }}
           PRECISIONS_SET: ${{ inputs.precisions_set }}
+          __TFHE_RS_BENCH_MULTI_PROC_GROUPS: 2
 
       - name: Parse results
+        if: ${{ inputs.command != 'hlapi_erc7984_multi_group' }}
         run: |
           python3 ./ci/benchmark_parser.py target/criterion "${RESULTS_FILENAME}" \
           --database tfhe_rs \
@@ -282,6 +258,39 @@ jobs:
           REF_NAME: ${{ github.ref_name }}
           BENCH_TYPE: ${{ matrix.bench_type }}
 
+      - name: Parse and merge erc7984_multi_group results
+        if: ${{ inputs.command == 'hlapi_erc7984_multi_group' }}
+        run: |
+          python3 ./ci/benchmark_parser.py tfhe-benchmark/target_p0/criterion "${RESULTS_FILENAME_P0}" \
+          --database tfhe_rs \
+          --hardware "${INPUTS_HARDWARE_NAME}" \
+          --backend gpu \
+          --project-version "${COMMIT_HASH}" \
+          --branch "${REF_NAME}" \
+          --commit-date "${COMMIT_DATE}" \
+          --bench-date "${BENCH_DATE}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --bench-type "${BENCH_TYPE}"
+          python3 ./ci/benchmark_parser.py tfhe-benchmark/target_p1/criterion "${RESULTS_FILENAME_P1}" \
+          --database tfhe_rs \
+          --hardware "${INPUTS_HARDWARE_NAME}" \
+          --backend gpu \
+          --project-version "${COMMIT_HASH}" \
+          --branch "${REF_NAME}" \
+          --commit-date "${COMMIT_DATE}" \
+          --bench-date "${BENCH_DATE}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --bench-type "${BENCH_TYPE}"
+          python3 ./ci/merge_multi_group_results.py --bench-type "${BENCH_TYPE}" --output "${RESULTS_FILENAME}" "${RESULTS_FILENAME_P0}" "${RESULTS_FILENAME_P1}"
+        env:
+          INPUTS_HARDWARE_NAME: ${{ inputs.hardware_name }}
+          REF_NAME: ${{ github.ref_name }}
+          BENCH_TYPE: ${{ matrix.bench_type }}
+          RESULTS_FILENAME_P0: parsed_benchmark_results_p0_${{ github.sha }}.json
+          RESULTS_FILENAME_P1: parsed_benchmark_results_p1_${{ github.sha }}.json
+
       - name: Parse additional benchmarks results files
         if: ${{ inputs.additional_file_to_parse }}
         run: |
@@ -333,13 +342,13 @@ jobs:
 
   teardown-instance:
     name: benchmark_gpu_common/teardown-instance
-    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
+    if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [ setup-instance, cuda-benchmarks, slack-notify ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_gpu_coprocessor.yml

@@ -42,7 +42,7 @@ env:
   OPTIMIZATION_TARGET: "throughput"
   BATCH_SIZE: "5000"
   SCHEDULING_POLICY: "MAX_PARALLELISM"
-  BENCHMARKS: "erc20"
+  BENCHMARKS: "erc7984"
   BRANCH_NAME: ${{ github.ref_name }}
   COMMIT_SHA: ${{ github.sha }}
   SLAB_SECRET: ${{ secrets.JOB_SECRET }}
@@ -77,7 +77,7 @@ jobs:
           if [[ ${IS_MANUAL_RUN} == true ]]; then
             PROFILE_RAW="${PROFILE_MANUAL_RUN}"
           else
-            PROFILE_RAW="${PROFILE}"
+            PROFILE_RAW="${PROFILE_SCHEDULED_RUN}"
           fi
           # shellcheck disable=SC2001
           PROFILE_VAL=$(echo "${PROFILE_RAW}" | sed 's|.*[[:space:]](\(.*\))|\1|')
@@ -94,7 +94,7 @@ jobs:
     steps:
       - name: Start remote instance
         id: start-remote-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -204,7 +204,7 @@ jobs:
         uses: foundry-rs/foundry-toolchain@8789b3e21e6c11b2697f5eb56eddae542f746c10
 
       - name: Cache cargo
-        uses: actions/cache@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 # v5.0.3
+        uses: actions/cache@668228422ae6a00e4ad889ee87cd7109ec5666a7 # v5.0.4
         with:
           path: |
             ~/.cargo/registry
@@ -214,14 +214,14 @@ jobs:
           restore-keys: ${{ runner.os }}-cargo-
 
       - name: Login to GitHub Container Registry
-        uses: docker/login-action@b45d80f862d83dbcd57f89517bcf500b2ab88fb2 # v4.0.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           registry: ghcr.io
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Login to Chainguard Registry
-        uses: docker/login-action@b45d80f862d83dbcd57f89517bcf500b2ab88fb2 # v4.0.0
+        uses: docker/login-action@4907a6ddec9925e35a0a9e82d7399ccc52663121 # v4.1.0
         with:
           registry: cgr.dev
           username: ${{ secrets.CGR_USERNAME }}
@@ -248,13 +248,13 @@ jobs:
           npm install && npm run deploy:emptyProxies && npx hardhat compile
         working-directory: fhevm/
 
-      - name: Profile erc20 no-cmux benchmark on GPU
+      - name: Profile erc7984 no-cmux benchmark on GPU
         run: |
           BENCHMARK_BATCH_SIZE="${BATCH_SIZE}" \
           FHEVM_DF_SCHEDULE="${SCHEDULING_POLICY}" \
           BENCHMARK_TYPE="THROUGHPUT_200" \
           OPTIMIZATION_TARGET="${OPTIMIZATION_TARGET}" \
-          make -e "profile_erc20_gpu"
+          make -e "profile_erc7984_gpu"
         working-directory: fhevm/coprocessor/fhevm-engine/tfhe-worker
 
       - name: Get nsys profile name
@@ -333,7 +333,7 @@ jobs:
     steps:
       - name: Stop remote instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_hpu.yml

@@ -14,7 +14,7 @@ on:
           - integer
           - hlapi_unsigned
           - hlapi_signed
-          - hlapi_erc20
+          - hlapi_erc7984
       op_flavor:
         description: "Operations set to run"
         type: choice

.github/workflows/benchmark_hpu_common.yml

@@ -95,7 +95,7 @@ jobs:
             ]:
               f.write(f"""{env_name}=["{'", "'.join(values_to_join)}"]\n""")
 
-      - name: Set martix arguments outputs
+      - name: Set matrix arguments outputs
         id: set_matrix_args
         run: | # zizmor: ignore[template-injection] these env variable are safe
           {
@@ -121,7 +121,7 @@ jobs:
     steps:
       # Needed as long as hw_regmap repository is private
       - name: Configure SSH
-        uses: webfactory/ssh-agent@a6f90b1f127823b31d4d4a8d96047790581349bd # v0.9.1
+        uses: webfactory/ssh-agent@e83874834305fe9a4a2997156cb26c5de65a8555 # v0.10.0
         with:
           ssh-private-key: ${{ secrets.SSH_PRIVATE_KEY }}
 

.github/workflows/benchmark_perf_regression.yml

@@ -143,7 +143,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -387,7 +387,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_summary.yml

@@ -114,6 +114,27 @@ jobs:
       SLAB_URL: ${{ secrets.SLAB_URL }}
       SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
 
+  run-benchmarks-gpu-erc7984-multi-group:
+    name: benchmark_documentation/run-benchmarks-gpu-erc7984-multi-group
+    uses: ./.github/workflows/benchmark_gpu_common.yml
+    if: inputs.run-gpu-benchmarks
+    needs: parse-gpu-inputs
+    with:
+      profile: ${{ needs.parse-gpu-inputs.outputs.profile }}
+      hardware_name: ${{ needs.parse-gpu-inputs.outputs.hardware_name }}
+      command: hlapi_erc7984_multi_group
+      bench_type: throughput
+      params_type: multi_bit
+    secrets:
+      BOT_USERNAME: ${{ secrets.BOT_USERNAME }}
+      SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+      SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+      REPO_CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN }}
+      JOB_SECRET: ${{ secrets.JOB_SECRET }}
+      SLAB_ACTION_TOKEN: ${{ secrets.SLAB_ACTION_TOKEN }}
+      SLAB_URL: ${{ secrets.SLAB_URL }}
+      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
+
 # TODO add make recipe for HPU benchmarks
 #  run-benchmarks-hpu:
 #    name: benchmark_documentation/run-benchmarks-hpu

.github/workflows/benchmark_tfhe_fft.yml

@@ -40,7 +40,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -137,7 +137,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_tfhe_ntt.yml

@@ -40,7 +40,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -137,7 +137,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/benchmark_wasm_client_common.yml

@@ -63,7 +63,7 @@ jobs:
           with open(env_file, "a") as f:
             f.write(f"""BROWSER=["{'", "'.join(split_browser)}"]\n""")
 
-      - name: Set martix arguments output
+      - name: Set matrix arguments output
         id: set_matrix_arg
         run: | # zizmor: ignore[template-injection] this env variable is safe
           echo "browser=${{ toJSON(env.BROWSER) }}" >> "${GITHUB_OUTPUT}"
@@ -77,7 +77,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -124,7 +124,7 @@ jobs:
 
       - name: Node cache restoration
         id: node-cache
-        uses: actions/cache/restore@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/restore@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: |
             ~/.nvm
@@ -137,7 +137,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/save@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         if: steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |
@@ -158,9 +158,9 @@ jobs:
         env:
           BROWSER: ${{ matrix.browser }}
 
-      - name: Run benchmarks (unsafe coop)
+      - name: Run benchmarks (cross origin)
         run: |
-          make bench_web_js_api_unsafe_coop_"${BROWSER}"_ci
+          make bench_web_js_api_cross_origin_"${BROWSER}"_ci
         env:
           BROWSER: ${{ matrix.browser }}
 
@@ -218,7 +218,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/cargo_build_common.yml

@@ -94,7 +94,7 @@ jobs:
           with open(env_file, "a") as f:
             f.write(f"""RUNNERS=["{'", "'.join(runners)}"]\n""")
 
-      - name: Set martix runners outputs
+      - name: Set matrix runners outputs
         id: set_matrix_runners
         run: | # zizmor: ignore[template-injection] these env variable are safe
           echo "runners=${{ toJSON(env.RUNNERS) }}" >> "${GITHUB_OUTPUT}"
@@ -138,7 +138,7 @@ jobs:
       - name: Node cache restoration
         if: inputs.run-pcc-cpu-batch == 'pcc_batch_2'
         id: node-cache
-        uses: actions/cache/restore@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/restore@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: |
             ~/.nvm
@@ -151,7 +151,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/save@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         if: inputs.run-pcc-cpu-batch == 'pcc_batch_2' && steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |

.github/workflows/cargo_test_ntt.yml

@@ -63,7 +63,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -146,7 +146,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/ci_lint.yml

@@ -43,14 +43,14 @@ jobs:
           echo "version=$(make zizmor_version)" >> "${GITHUB_OUTPUT}"
 
       - name: Check workflows security
-        uses: zizmorcore/zizmor-action@0dce2577a4760a2749d8cfb7a84b7d5585ebcb7d # v0.5.0
+        uses: zizmorcore/zizmor-action@71321a20a9ded102f6e9ce5718a2fcec2c4f70d8 # v0.5.2
         with:
           advanced-security: 'false' # Print results directly in logs
           persona: pedantic
           version: ${{ steps.get_zizmor.outputs.version }}
 
       - name: Ensure SHA pinned actions
-        uses: zgosalvez/github-actions-ensure-sha-pinned-actions@70c4af2ed5282c51ba40566d026d6647852ffa3e # v5.0.1
+        uses: zgosalvez/github-actions-ensure-sha-pinned-actions@ca46236c6ce584ae24bc6283ba8dcf4b3ec8a066 # v5.0.4
         with:
           allowlist: |
             slsa-framework/slsa-github-generator

.github/workflows/code_coverage.yml

@@ -74,7 +74,7 @@ jobs:
           make test_shortint_cov
 
       - name: Upload tfhe coverage to Codecov
-        uses: codecov/codecov-action@671740ac38dd9b0130fbe1cec585b89eea48d3de
+        uses: codecov/codecov-action@57e3a136b779b570ffcdbf80b3bdc90e7fab3de2
         if: steps.changed-files.outputs.tfhe_any_changed == 'true'
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
@@ -88,7 +88,7 @@ jobs:
           make test_integer_cov
 
       - name: Upload tfhe coverage to Codecov
-        uses: codecov/codecov-action@671740ac38dd9b0130fbe1cec585b89eea48d3de
+        uses: codecov/codecov-action@57e3a136b779b570ffcdbf80b3bdc90e7fab3de2
         if: steps.changed-files.outputs.tfhe_any_changed == 'true'
         with:
           token: ${{ secrets.CODECOV_TOKEN }}

.github/workflows/generate_svgs.yml

@@ -209,60 +209,98 @@ jobs:
       DATA_EXTRACTOR_DATABASE_HOST: ${{ secrets.DATA_EXTRACTOR_DATABASE_HOST }}
       DATA_EXTRACTOR_DATABASE_PASSWORD: ${{ secrets.DATA_EXTRACTOR_DATABASE_PASSWORD }}
 
+  gpu-zk-server-latency-table:
+    name: generate_documentation_svgs/gpu-zk-server-latency-table
+    uses: ./.github/workflows/generate_svg_common.yml
+    if: inputs.generate-gpu-svgs
+    with:
+      backend: gpu
+      hardware_name: n3-H100-SXM5x8
+      layer: integer
+      bench_subset: zk
+      pbs_kind: multi_bit
+      grouping_factor: 4
+      bench_type: latency
+      time_span_days: ${{ inputs.time_span_days }}
+      output_filename: gpu-zk-benchmark-latency
+    secrets:
+      DATA_EXTRACTOR_DATABASE_USER: ${{ secrets.DATA_EXTRACTOR_DATABASE_USER }}
+      DATA_EXTRACTOR_DATABASE_HOST: ${{ secrets.DATA_EXTRACTOR_DATABASE_HOST }}
+      DATA_EXTRACTOR_DATABASE_PASSWORD: ${{ secrets.DATA_EXTRACTOR_DATABASE_PASSWORD }}
+
+  gpu-zk-server-throughput-table:
+    name: generate_documentation_svgs/gpu-zk-server-throughput-table
+    uses: ./.github/workflows/generate_svg_common.yml
+    if: inputs.generate-gpu-svgs
+    with:
+      backend: gpu
+      hardware_name: n3-H100-SXM5x8
+      layer: integer
+      bench_subset: zk
+      pbs_kind: multi_bit
+      grouping_factor: 4
+      bench_type: throughput
+      time_span_days: ${{ inputs.time_span_days }}
+      output_filename: gpu-zk-benchmark-throughput
+    secrets:
+      DATA_EXTRACTOR_DATABASE_USER: ${{ secrets.DATA_EXTRACTOR_DATABASE_USER }}
+      DATA_EXTRACTOR_DATABASE_HOST: ${{ secrets.DATA_EXTRACTOR_DATABASE_HOST }}
+      DATA_EXTRACTOR_DATABASE_PASSWORD: ${{ secrets.DATA_EXTRACTOR_DATABASE_PASSWORD }}
+
   # -----------------------------------------------------------
-  # ERC20 benchmarks tables
+  # ERC7984 benchmarks tables
   # -----------------------------------------------------------
 
-  cpu-erc20-latency-throughput-table:
-    name: generate_documentation_svgs/cpu-erc20-latency-throughput-table
+  cpu-erc7984-latency-throughput-table:
+    name: generate_documentation_svgs/cpu-erc7984-latency-throughput-table
     uses: ./.github/workflows/generate_svg_common.yml
     if: inputs.generate-cpu-svgs
     with:
       backend: cpu
       hardware_name: hpc7a.96xlarge
       layer: hlapi
-      bench_subset: erc20
+      bench_subset: erc7984
       pbs_kind: classical
       bench_type: both
       time_span_days: ${{ inputs.time_span_days }}
-      output_filename: cpu-hlapi-erc20-benchmark-latency-throughput
+      output_filename: cpu-hlapi-erc7984-benchmark-latency-throughput
     secrets:
       DATA_EXTRACTOR_DATABASE_USER: ${{ secrets.DATA_EXTRACTOR_DATABASE_USER }}
       DATA_EXTRACTOR_DATABASE_HOST: ${{ secrets.DATA_EXTRACTOR_DATABASE_HOST }}
       DATA_EXTRACTOR_DATABASE_PASSWORD: ${{ secrets.DATA_EXTRACTOR_DATABASE_PASSWORD }}
 
-  gpu-erc20-latency-throughput-table:
-    name: generate_documentation_svgs/gpu-erc20-latency-throughput-table
+  gpu-erc7984-latency-throughput-table:
+    name: generate_documentation_svgs/gpu-erc7984-latency-throughput-table
     uses: ./.github/workflows/generate_svg_common.yml
     if: inputs.generate-gpu-svgs
     with:
       backend: gpu
       hardware_name: n3-H100-SXM5x8
       layer: hlapi
-      bench_subset: erc20
+      bench_subset: erc7984
       pbs_kind: multi_bit
       grouping_factor: 4
       bench_type: both
       time_span_days: ${{ inputs.time_span_days }}
-      output_filename: gpu-hlapi-erc20-benchmark-h100x8-sxm5-latency-throughput
+      output_filename: gpu-hlapi-erc7984-benchmark-h100x8-sxm5-latency-throughput
     secrets:
       DATA_EXTRACTOR_DATABASE_USER: ${{ secrets.DATA_EXTRACTOR_DATABASE_USER }}
       DATA_EXTRACTOR_DATABASE_HOST: ${{ secrets.DATA_EXTRACTOR_DATABASE_HOST }}
       DATA_EXTRACTOR_DATABASE_PASSWORD: ${{ secrets.DATA_EXTRACTOR_DATABASE_PASSWORD }}
 
-  hpu-erc20-latency-throughput-table:
-    name: generate_documentation_svgs/hpu-erc20-latency-throughput-table
+  hpu-erc7984-latency-throughput-table:
+    name: generate_documentation_svgs/hpu-erc7984-latency-throughput-table
     uses: ./.github/workflows/generate_svg_common.yml
     if: inputs.generate-hpu-svgs
     with:
       backend: hpu
       hardware_name: hpu_x1
       layer: hlapi
-      bench_subset: erc20
+      bench_subset: erc7984
       pbs_kind: classical
       bench_type: both
       time_span_days: ${{ inputs.time_span_days }}
-      output_filename: hpu-hlapi-erc20-benchmark-hpux1-latency-throughput.svg
+      output_filename: hpu-hlapi-erc7984-benchmark-hpux1-latency-throughput.svg
     secrets:
       DATA_EXTRACTOR_DATABASE_USER: ${{ secrets.DATA_EXTRACTOR_DATABASE_USER }}
       DATA_EXTRACTOR_DATABASE_HOST: ${{ secrets.DATA_EXTRACTOR_DATABASE_HOST }}

.github/workflows/gpu_code_validation_tests.yml

@@ -43,7 +43,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -149,7 +149,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_core_h100_tests.yml

@@ -23,7 +23,7 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
-    types: [ labeled ]
+    types: [ labeled, opened, synchronize ]
 
 permissions:
   contents: read
@@ -38,6 +38,7 @@ jobs:
       pull-requests: read  # Needed to check for file change
     outputs:
       gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+      core_crypto_changed: ${{ steps.changed-files.outputs.core_crypto_any_changed }}
     steps:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
@@ -62,29 +63,24 @@ jobs:
               - tfhe/src/integer/server_key/radix_parallel/tests_cases_unsigned.rs
               - tfhe/src/shortint/parameters/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_core_h100_tests.yml'
+            core_crypto:
+              - tfhe/src/core_crypto/gpu/**
 
   setup-instance:
     name: gpu_core_h100_tests/setup-instance
     needs: should-run
     if: github.event_name != 'pull_request' ||
-      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
-      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.core_crypto_changed == 'true')
     runs-on: ubuntu-latest
     outputs:
-      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
-      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
-      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
-      # otherwise we'll try to run the next job on a non-existing on-demand instance.
-      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
-      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
+      runner-name: ${{ steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        continue-on-error: true
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -93,13 +89,6 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
-      # This will allow to fallback on permanent instances running on Hyperstack.
-      - name: Use permanent remote instance
-        id: use-permanent-instance
-        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
-        run: |
-          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
-
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -132,7 +121,6 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
-        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -176,14 +164,14 @@ jobs:
 
   teardown-instance:
     name: gpu_core_h100_tests/teardown-instance
-    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
+    if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_fast_tests.yml

@@ -77,7 +77,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -182,7 +182,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_full_h100_tests.yml

@@ -25,17 +25,11 @@ jobs:
     name: gpu_full_h100_tests/setup-instance
     runs-on: ubuntu-latest
     outputs:
-      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
-      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
-      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
-      # otherwise we'll try to run the next job on a non-existing on-demand instance.
-      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
-      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
+      runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
-      - name: Start remote instance
-        id: start-remote-instance
-        continue-on-error: true
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -44,13 +38,6 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
-      # This will allow to fallback on permanent instances running on Hyperstack.
-      - name: Use permanent remote instance
-        id: use-permanent-instance
-        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
-        run: |
-          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
-
   cuda-tests-linux:
     name: gpu_full_h100_tests/cuda-tests-linux
     needs: [ setup-instance ]
@@ -74,7 +61,6 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
-        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -118,13 +104,13 @@ jobs:
 
   teardown-instance:
     name: gpu_full_h100_tests/teardown-instance
-    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
+    if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_full_multi_gpu_tests.yml

@@ -80,7 +80,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -186,7 +186,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_hlapi_h100_tests.yml

@@ -23,7 +23,7 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
-    types: [ labeled ]
+    types: [ labeled, opened, synchronize ]
 
 permissions:
   contents: read
@@ -38,6 +38,7 @@ jobs:
       pull-requests: read  # Needed to check for file change
     outputs:
       gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+      core_crypto_changed: ${{ steps.changed-files.outputs.core_crypto_any_changed }}
     steps:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
@@ -65,27 +66,23 @@ jobs:
               - tfhe/src/c_api/**
               - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_hlapi_h100_tests.yml'
+            core_crypto:
+              - tfhe/src/core_crypto/gpu/**
 
   setup-instance:
     name: gpu_hlapi_h100_tests/setup-instance
     needs: should-run
     if: github.event_name != 'pull_request' ||
-      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
-      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.core_crypto_changed == 'true')
     runs-on: ubuntu-latest
     outputs:
-      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
-      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
-      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
-      # otherwise we'll try to run the next job on a non-existing on-demand instance.
-      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
-      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
+      runner-name: ${{ steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        continue-on-error: true
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -94,13 +91,6 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
-      # This will allow to fallback on permanent instances running on Hyperstack.
-      - name: Use permanent remote instance
-        id: use-permanent-instance
-        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
-        run: |
-          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
-
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -133,7 +123,6 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
-        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -184,14 +173,14 @@ jobs:
 
   teardown-instance:
     name: gpu_hlapi_h100_tests/teardown-instance
-    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
+    if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_integer_long_run_tests.yml

@@ -17,8 +17,8 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   schedule:
-    # Nightly tests will be triggered each evening 8p.m.
-    - cron: "0 20 * * *"
+    # Weekly tests will be triggered every Monday at 8p.m.
+    - cron: "0 20 * * 1"
   pull_request:
 
 
@@ -28,17 +28,48 @@ permissions:
 # zizmor: ignore[concurrency-limits] concurrency is managed after instance setup to ensure safe provisioning
 
 jobs:
+  should-run:
+    name: gpu_integer_long_run_tests/should-run
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: read  # Needed to check for file change
+    outputs:
+      is_needed_in_gpu_ci: ${{ env.IS_PR == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
+        with:
+          fetch-depth: 0
+          persist-credentials: 'false'
+          token: ${{ env.CHECKOUT_TOKEN }}
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@22103cc46bda19c2b464ffe86db46df6922fd323 # v47.0.5
+        with:
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/src/shortint/parameters/**
+              - '.github/workflows/gpu_integer_long_run_tests.yml'
+
   setup-instance:
     name: gpu_integer_long_run_tests/setup-instance
-    if: github.event_name != 'schedule' ||
-      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
+    needs: [should-run]
+    if: github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      needs.should-run.outputs.is_needed_in_gpu_ci == 'true'
     runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -112,7 +143,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_memory_sanitizer.yml

@@ -74,7 +74,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -166,7 +166,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_memory_sanitizer_h100.yml

@@ -74,7 +74,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -166,7 +166,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_pcc.yml

@@ -38,7 +38,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -131,6 +131,10 @@ jobs:
         env:
           GCC_VERSION: ${{ matrix.gcc }}
 
+      - name: Run semgrep and lint checks on CUDA code
+        run: |
+          make semgrep_and_lint_gpu_code
+
       - name: Run fmt checks
         run: |
           make check_fmt_gpu
@@ -139,10 +143,6 @@ jobs:
         run: |
           make pcc_gpu
 
-      - name: Run semgrep and lint checks on CUDA code
-        run: |
-          make semgrep_and_lint_gpu_code
-
       - name: Run semver checks on tfhe-cuda-backend
         run: |
           make semver_check_cuda_backend
@@ -176,7 +176,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_signed_integer_classic_tests.yml

@@ -63,7 +63,6 @@ jobs:
               - tfhe/src/shortint/parameters/**
               - tfhe/src/high_level_api/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_signed_integer_classic_tests.yml'
               - scripts/integer-tests.sh
 
@@ -80,7 +79,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -169,7 +168,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_signed_integer_h100_tests.yml

@@ -23,7 +23,7 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
-    types: [ labeled ]
+    types: [ labeled, opened, synchronize ]
 
 permissions:
   contents: read
@@ -38,6 +38,7 @@ jobs:
       pull-requests: read  # Needed to check for file change
     outputs:
       gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+      core_crypto_changed: ${{ steps.changed-files.outputs.core_crypto_any_changed }}
     steps:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
@@ -63,30 +64,25 @@ jobs:
               - tfhe/src/shortint/parameters/**
               - tfhe/src/high_level_api/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_signed_integer_h100_tests.yml'
               - scripts/integer-tests.sh
+            core_crypto:
+              - tfhe/src/core_crypto/gpu/**
 
   setup-instance:
     name: gpu_signed_integer_h100_tests/setup-instance
     needs: should-run
     if: github.event_name != 'pull_request' ||
-      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
-      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.core_crypto_changed == 'true')
     runs-on: ubuntu-latest
     outputs:
-      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
-      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
-      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
-      # otherwise we'll try to run the next job on a non-existing on-demand instance.
-      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
-      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
+      runner-name: ${{ steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        continue-on-error: true
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -95,13 +91,6 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
-      # This will allow to fallback on permanent instances running on Hyperstack.
-      - name: Use permanent remote instance
-        id: use-permanent-instance
-        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
-        run: |
-          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
-
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -134,7 +123,6 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
-        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -176,14 +164,14 @@ jobs:
 
   teardown-instance:
     name: gpu_signed_integer_h100_tests/teardown-instance
-    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
+    if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_signed_integer_tests.yml

@@ -64,7 +64,6 @@ jobs:
               - tfhe/src/shortint/parameters/**
               - tfhe/src/high_level_api/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_signed_integer_tests.yml'
               - scripts/integer-tests.sh
 
@@ -81,7 +80,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -178,7 +177,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_unsigned_integer_classic_tests.yml

@@ -63,7 +63,6 @@ jobs:
               - tfhe/src/shortint/parameters/**
               - tfhe/src/high_level_api/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_unsigned_integer_classic_tests.yml'
               - scripts/integer-tests.sh
 
@@ -80,7 +79,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -169,7 +168,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_unsigned_integer_h100_tests.yml

@@ -23,7 +23,7 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
-    types: [ labeled ]
+    types: [ labeled, opened, synchronize ]
 
 permissions:
   contents: read
@@ -38,6 +38,7 @@ jobs:
       pull-requests: read  # Needed to check for file change
     outputs:
       gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+      core_crypto_changed: ${{ steps.changed-files.outputs.core_crypto_any_changed }}
     steps:
       - name: Checkout tfhe-rs
         uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
@@ -63,30 +64,25 @@ jobs:
               - tfhe/src/shortint/parameters/**
               - tfhe/src/high_level_api/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_unsigned_integer_h100_tests.yml'
               - scripts/integer-tests.sh
+            core_crypto:
+              - tfhe/src/core_crypto/gpu/**
 
   setup-instance:
     name: gpu_unsigned_integer_h100_tests/setup-instance
     needs: should-run
     if: github.event_name == 'workflow_dispatch' ||
-      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
-      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.core_crypto_changed == 'true')
     runs-on: ubuntu-latest
     outputs:
-      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
-      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
-      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
-      # otherwise we'll try to run the next job on a non-existing on-demand instance.
-      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
-      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
+      runner-name: ${{ steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        continue-on-error: true
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -95,13 +91,6 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
-      # This will allow to fallback on permanent instances running on Hyperstack.
-      - name: Use permanent remote instance
-        id: use-permanent-instance
-        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
-        run: |
-          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
-
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -134,7 +123,6 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
-        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -176,14 +164,14 @@ jobs:
 
   teardown-instance:
     name: gpu_unsigned_integer_h100_tests/teardown-instance
-    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
+    if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_unsigned_integer_tests.yml

@@ -64,7 +64,6 @@ jobs:
               - tfhe/src/shortint/parameters/**
               - tfhe/src/high_level_api/**
               - tfhe/src/c_api/**
-              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_unsigned_integer_tests.yml'
               - scripts/integer-tests.sh
 
@@ -81,7 +80,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -178,7 +177,7 @@ jobs:
       - name: Stop instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_zk_tests.yml

@@ -51,7 +51,13 @@ jobs:
         with:
           files_yaml: |
             gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
               - backends/zk-cuda-backend/**
+              - tfhe/src/shortint/parameters/**
+              - tfhe/src/zk/**
+              - tfhe-zk-pok/**
               - '.github/workflows/gpu_zk_tests.yml'
               - ci/slab.toml
 
@@ -67,7 +73,7 @@ jobs:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -126,6 +132,9 @@ jobs:
       - name: Run zk-cuda-backend integration tests
         run: |
           make test_zk_cuda_backend
+          make test_zk_pok_experimental_gpu
+          make test_integer_zk_gpu
+          make test_integer_zk_experimental_gpu
 
   slack-notify:
     name: gpu_zk_tests/slack-notify
@@ -158,7 +167,7 @@ jobs:
       - name: Stop remote instance
         id: stop-instance
         if: env.SECRETS_AVAILABLE == 'true'
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/make_release_common.yml

@@ -101,13 +101,13 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Download artifact
-        uses: actions/download-artifact@70fc10c6e5e1ce46ad2ea6f2b72d43f7d47b13c3 # v8.0.0
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
         with:
           name: crate-${{ inputs.package-name }}
           path: target/package
 
       - name: Authenticate on registry
-        uses: rust-lang/crates-io-auth-action@b7e9a28eded4986ec6b1fa40eeee8f8f165559ec # v1.0.3
+        uses: rust-lang/crates-io-auth-action@bbd81622f20ce9e2dd9622e3218b975523e45bbe # v1.0.4
         id: auth
 
       - name: Publish crate.io package

.github/workflows/make_release_common_cuda.yml

@@ -1,12 +1,36 @@
-name: make_release_cuda
+# Common workflow to make crate release for CUDA backend
+name: make_release_common_cuda
 
 on:
-  workflow_dispatch:
+  workflow_call:
     inputs:
-      dry_run:
-        description: "Dry-run"
+      package-name:
+        type: string
+        required: true
+      dry-run:
         type: boolean
         default: true
+    secrets:
+      REPO_CHECKOUT_TOKEN:
+        required: true
+      SLAB_ACTION_TOKEN:
+        required: true
+      SLAB_BASE_URL:
+        required: true
+      SLAB_URL:
+        required: true
+      JOB_SECRET:
+        required: true
+      SLACK_CHANNEL:
+        required: true
+      BOT_USERNAME:
+        required: true
+      SLACK_WEBHOOK:
+        required: true
+      ALLOWED_TEAM:
+        required: true
+      READ_ORG_TOKEN:
+        required: true
 
 env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
@@ -21,15 +45,15 @@ permissions: {}
 
 jobs:
   verify-triggering-actor:
-    name: make_release_cuda/verify-triggering-actor
+    name: make_release_common_cuda/verify-triggering-actor
     if: startsWith(github.ref, 'refs/tags/')
     uses: ./.github/workflows/verify_triggering_actor.yml
     secrets:
-      ALLOWED_TEAM: ${{ secrets.RELEASE_TEAM }}
+      ALLOWED_TEAM: ${{ secrets.ALLOWED_TEAM }}
       READ_ORG_TOKEN: ${{ secrets.READ_ORG_TOKEN }}
 
   setup-instance:
-    name: make_release_cuda/setup-instance
+    name: make_release_common_cuda/setup-instance
     needs: verify-triggering-actor
     runs-on: ubuntu-latest
     outputs:
@@ -37,7 +61,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -47,7 +71,7 @@ jobs:
           profile: gpu-build
 
   package:
-    name: make_release_cuda/package
+    name: make_release_common_cuda/package
     needs: setup-instance
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     outputs:
@@ -76,7 +100,6 @@ jobs:
           toolchain: stable
 
       - name: Export CUDA variables
-        if: ${{ !cancelled() }}
         run: |
           echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
           {
@@ -89,7 +112,6 @@ jobs:
 
       # Specify the correct host compilers
       - name: Export gcc and g++ variables
-        if: ${{ !cancelled() }}
         run: |
           {
             echo "CC=/usr/bin/gcc-${GCC_VERSION}";
@@ -101,12 +123,14 @@ jobs:
           GCC_VERSION: ${{ matrix.gcc }}
 
       - name: Prepare package
+        env:
+          PACKAGE: ${{ inputs.package-name }}
         run: |
-          cargo package -p tfhe-cuda-backend
+          cargo package -p "${PACKAGE}"
 
       - uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7.0.0
         with:
-          name: crate-tfhe-cuda-backend
+          name: crate-${{ inputs.package-name }}
           path: target/package/*.crate
 
       - name: generate hash
@@ -114,8 +138,8 @@ jobs:
         run: cd target/package && echo "hash=$(sha256sum ./*.crate | base64 -w0)" >> "${GITHUB_OUTPUT}"
 
   provenance:
-    name: make_release_cuda/provenance
-    if: ${{ !inputs.dry_run  }}
+    name: make_release_common_cuda/provenance
+    if: ${{ !inputs.dry-run  }}
     needs: [package]
     # This action cannot be pinned to a specific commit (see https://github.com/slsa-framework/slsa-github-generator/blob/main/README.md#referencing-slsa-builders-and-generators)
     uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0 # zizmor: ignore[unpinned-uses] as said above SLSA cannot be pinned by tag today
@@ -128,7 +152,7 @@ jobs:
       base64-subjects: ${{ needs.package.outputs.hash }}
 
   publish-cuda-release:
-    name: make_release_cuda/publish-cuda-release
+    name: make_release_common_cuda/publish-cuda-release
     needs: [setup-instance, package] # for comparing hashes
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     permissions:
@@ -150,7 +174,6 @@ jobs:
           toolchain: stable
 
       - name: Export CUDA variables
-        if: ${{ !cancelled() }}
         run: |
           echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
           {
@@ -163,7 +186,6 @@ jobs:
 
       # Specify the correct host compilers
       - name: Export gcc and g++ variables
-        if: ${{ !cancelled() }}
         run: |
           {
             echo "CC=/usr/bin/gcc-${GCC_VERSION}";
@@ -175,24 +197,25 @@ jobs:
           GCC_VERSION: ${{ matrix.gcc }}
 
       - name: Download artifact
-        uses: actions/download-artifact@70fc10c6e5e1ce46ad2ea6f2b72d43f7d47b13c3 # v8.0.0
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
         with:
-          name: crate-tfhe-cuda-backend
+          name: crate-${{ inputs.package-name }}
           path: target/package
 
       - name: Authenticate on registry
-        uses: rust-lang/crates-io-auth-action@b7e9a28eded4986ec6b1fa40eeee8f8f165559ec # v1.0.3
+        uses: rust-lang/crates-io-auth-action@bbd81622f20ce9e2dd9622e3218b975523e45bbe # v1.0.4
         id: auth
 
       - name: Publish crate.io package
         env:
           CARGO_REGISTRY_TOKEN: ${{ steps.auth.outputs.token }}
-          DRY_RUN: ${{ inputs.dry_run && '--dry-run' || '' }}
+          PACKAGE: ${{ inputs.package-name }}
+          DRY-RUN: ${{ inputs.dry-run && '--dry-run' || '' }}
         run: |
-          # DRY_RUN expansion cannot be double quoted when variable contains empty string otherwise cargo publish 
-          # would fail. This is safe since DRY_RUN is handled in the env section above.
+          # dry-run expansion cannot be double quoted when variable contains empty string otherwise cargo publish
+          # would fail. This is safe since dry-run is handled in the env section above.
           # shellcheck disable=SC2086
-          cargo publish -p tfhe-cuda-backend ${DRY_RUN}
+          cargo publish -p "${PACKAGE}" ${DRY-RUN}
 
       - name: Generate hash
         id: published_hash
@@ -204,7 +227,7 @@ jobs:
         uses: rtCamp/action-slack-notify@e31e87e03dd19038e411e38ae27cbad084a90661 # v2.3.3
         env:
           SLACK_COLOR: failure
-          SLACK_MESSAGE: "SLSA tfhe-cuda-backend crate - hash comparison failure: (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "SLSA ${{ inputs.package-name }} crate - hash comparison failure: (${{ env.ACTION_RUN_URL }})"
 
       - name: Slack Notification
         if: ${{ failure() || (cancelled() && github.event_name != 'pull_request') }}
@@ -212,17 +235,17 @@ jobs:
         uses: rtCamp/action-slack-notify@e31e87e03dd19038e411e38ae27cbad084a90661 # v2.3.3
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "tfhe-cuda-backend release finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "${{ inputs.package-name }} release finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
 
   teardown-instance:
-    name: make_release_cuda/teardown-instance
+    name: make_release_common_cuda/teardown-instance
     if: ${{ always() && needs.setup-instance.result == 'success' }}
     needs: [setup-instance, publish-cuda-release]
     runs-on: ubuntu-latest
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@0a812986560d3f10dc65728b1ccb9ae4c48a8a16 # v1.5.1
+        uses: zama-ai/slab-github-runner@5aee5d157f4a0201e5eaefc9cc648e5f9f5472a5 # v1.6.0
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -235,4 +258,4 @@ jobs:
         uses: rtCamp/action-slack-notify@e31e87e03dd19038e411e38ae27cbad084a90661
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (publish-cuda-release) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (${{ inputs.package-name }} release) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/make_release_tfhe.yml

@@ -16,6 +16,10 @@ on:
         description: "Push web js package"
         type: boolean
         default: true
+      push_web_compat_package:
+        description: "Push web compat (cross-origin) js package"
+        type: boolean
+        default: true
       push_node_package:
         description: "Push node js package"
         type: boolean
@@ -99,6 +103,23 @@ jobs:
           tag: ${{ env.NPM_TAG }}
           provenance: true
 
+      - name: Build web compat (cross-origin) package
+        if: ${{ inputs.push_web_compat_package }}
+        run: |
+          rm -rf tfhe/pkg
+
+          make build_web_js_api
+          sed -i 's/"tfhe"/"tfhe-compat"/g' tfhe/pkg/package.json
+
+      - name: Publish web compat (cross-origin) package
+        if: ${{ inputs.push_web_compat_package }}
+        uses: JS-DevTools/npm-publish@0fd2f4369c5d6bcfcde6091a7c527d810b9b5c3f
+        with:
+          package: tfhe/pkg/package.json
+          dry-run: ${{ inputs.dry_run }}
+          tag: ${{ env.NPM_TAG }}
+          provenance: true
+
       - name: Build Node package
         if: ${{ inputs.push_node_package }}
         run: |

.github/workflows/make_release_tfhe_cuda.yml

@@ -0,0 +1,44 @@
+# Publish new release of tfhe-rs CUDA backend on crates.io.
+name: make_release_tfhe_cuda
+
+on:
+  workflow_dispatch:
+    inputs:
+      dry_run:
+        description: "Dry-run"
+        type: boolean
+        default: true
+
+env:
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+permissions: {}
+
+# zizmor: ignore[concurrency-limits] only Zama organization members can trigger this workflow
+
+jobs:
+  make-release:
+    name: make_release_tfhe_cuda/make-release
+    uses: ./.github/workflows/make_release_common_cuda.yml
+    with:
+      package-name: "tfhe-cuda-backend"
+      dry-run: ${{ inputs.dry_run }}
+    permissions:
+      actions: read # Needed to detect the GitHub Actions environment
+      id-token: write # Needed to create the provenance via GitHub OIDC
+      contents: write # Needed to upload assets/artifacts
+    secrets:
+      BOT_USERNAME: ${{ secrets.BOT_USERNAME }}
+      SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+      SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+      REPO_CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN }}
+      ALLOWED_TEAM: ${{ secrets.RELEASE_TEAM }}
+      READ_ORG_TOKEN: ${{ secrets.READ_ORG_TOKEN }}
+      SLAB_ACTION_TOKEN: ${{ secrets.SLAB_ACTION_TOKEN }}
+      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
+      SLAB_URL: ${{ secrets.SLAB_URL }}
+      JOB_SECRET: ${{ secrets.JOB_SECRET }}

.github/workflows/make_release_tfhe_safe_serialize.yml

@@ -0,0 +1,32 @@
+name: make_release_tfhe_safe_serialize
+
+on:
+  workflow_dispatch:
+    inputs:
+      dry_run:
+        description: "Dry-run"
+        type: boolean
+        default: true
+
+permissions: {}
+
+# zizmor: ignore[concurrency-limits] only Zama organization members can trigger this workflow
+
+jobs:
+  make-release:
+    name: make_release_tfhe_safe_serialize/make-release
+    uses: ./.github/workflows/make_release_common.yml
+    with:
+      package-name: "tfhe-safe-serialize"
+      dry-run: ${{ inputs.dry_run }}
+    permissions:
+      actions: read # Needed to detect the GitHub Actions environment
+      id-token: write # Needed to create the provenance via GitHub OIDC
+      contents: write # Needed to upload assets/artifacts
+    secrets:
+      BOT_USERNAME: ${{ secrets.BOT_USERNAME }}
+      SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+      SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+      REPO_CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN }}
+      ALLOWED_TEAM: ${{ secrets.RELEASE_TEAM }}
+      READ_ORG_TOKEN: ${{ secrets.READ_ORG_TOKEN }}

.github/workflows/make_release_zk_cuda.yml

@@ -0,0 +1,44 @@
+# Publish new release of CUDA Zero-Knowledge primitives on crates.io.
+name: make_release_zk_cuda
+
+on:
+  workflow_dispatch:
+    inputs:
+      dry_run:
+        description: "Dry-run"
+        type: boolean
+        default: true
+
+env:
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+permissions: {}
+
+# zizmor: ignore[concurrency-limits] only Zama organization members can trigger this workflow
+
+jobs:
+  make-release:
+    name: make_release_zk_cuda/make-release
+    uses: ./.github/workflows/make_release_common_cuda.yml
+    with:
+      package-name: "zk-cuda-backend"
+      dry-run: ${{ inputs.dry_run }}
+    permissions:
+      actions: read # Needed to detect the GitHub Actions environment
+      id-token: write # Needed to create the provenance via GitHub OIDC
+      contents: write # Needed to upload assets/artifacts
+    secrets:
+      BOT_USERNAME: ${{ secrets.BOT_USERNAME }}
+      SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+      SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+      REPO_CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN }}
+      ALLOWED_TEAM: ${{ secrets.RELEASE_TEAM }}
+      READ_ORG_TOKEN: ${{ secrets.READ_ORG_TOKEN }}
+      SLAB_ACTION_TOKEN: ${{ secrets.SLAB_ACTION_TOKEN }}
+      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
+      SLAB_URL: ${{ secrets.SLAB_URL }}
+      JOB_SECRET: ${{ secrets.JOB_SECRET }}

.github/workflows/parameters_check.yml

@@ -53,7 +53,7 @@ jobs:
 
       - name: Restore Sagemath image from cache
         id: docker-cache
-        uses: actions/cache/restore@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/restore@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: /tmp/sagemath_image
           key: sagemath-image-${{ env.SAGEMATH_VERSION }}-${{ github.sha }}
@@ -76,7 +76,7 @@ jobs:
       - name: Store Sagemath image in cache
         if: steps.docker-cache.outputs.cache-hit != 'true'
         continue-on-error: true
-        uses: actions/cache/save@cdf6c1fa76f9f475f3d7449005a359c84ca0f306 #v5.0.3
+        uses: actions/cache/save@668228422ae6a00e4ad889ee87cd7109ec5666a7 #v5.0.4
         with:
           path: /tmp/sagemath_image
           key: sagemath-image-${{ env.SAGEMATH_VERSION }}-${{ github.sha }}

.github/workflows/sync_on_push.yml

@@ -24,6 +24,8 @@ jobs:
           SOURCE_REPO: "zama-ai/tfhe-rs"
           SOURCE_BRANCH: "main"
           DESTINATION_BRANCH: "main"
+          SOURCE_TAGS: "refs/tags/*"
+          DESTINATION_TAGS: "refs/tags/*"
           USERNAME: ${{ secrets.BOT_USERNAME }}
           TOKEN: ${{ secrets.SYNC_REPO_TOKEN }}
           DEST_REPO: ${{ secrets.SYNC_DEST_REPO }}
@@ -33,49 +35,16 @@ jobs:
           git clone --quiet "https://${USERNAME}:${TOKEN}@github.com/${SOURCE_REPO}.git" ./tfhe-rs --origin source && cd ./tfhe-rs
           git remote add destination "https://${USERNAME}:${TOKEN}@github.com/${DEST_REPO}.git"
 
-          # The LFS config disables pulling files by default, so remove it
-          # TODO: see if we need to more precisely fetch LFS files or if git is smart
-          rm .lfsconfig
-
           echo ">>> Fetching all branches references down locally so subsequent commands can see them..."
-          git fetch source '+refs/heads/*:refs/heads/*' --update-head-ok
+          git fetch --all --tags --update-head-ok --quiet
 
-          echo ">>> Print out all branches"
-          git --no-pager branch -a -vv
+          echo ">>> Sync LFS items from source..."
+          ./scripts/lfs_sync.sh source destination "${SOURCE_BRANCH}"
 
-          echo ">>> Pull LFS items from source..."
-          git lfs pull source "${SOURCE_BRANCH}"
-
-          echo ">>> Pushing git changes and LFS content..."
+          echo ">>> Pushing git changes for ${SOURCE_BRANCH}..."
           git push destination "${SOURCE_BRANCH}:${DESTINATION_BRANCH}" -f
 
-          shred --remove .git/config
-
-      - name: git-sync-tags
-        env:
-          SOURCE_REPO: "zama-ai/tfhe-rs"
-          SOURCE_BRANCH: "refs/tags/*"
-          DESTINATION_BRANCH: "refs/tags/*"
-          USERNAME: ${{ secrets.BOT_USERNAME }}
-          TOKEN: ${{ secrets.SYNC_REPO_TOKEN }}
-          DEST_REPO: ${{ secrets.SYNC_DEST_REPO }}
-        run: |
-          echo ">>> Cloning source repo..."
-          git lfs install
-          git clone --quiet "https://${USERNAME}:${TOKEN}@github.com/${SOURCE_REPO}.git" ./tfhe-rs-tag --origin source && cd ./tfhe-rs-tag
-          git remote add destination "https://${USERNAME}:${TOKEN}@github.com/${DEST_REPO}.git"
-
-          # The LFS config disables pulling files by default, so remove it
-          # TODO: see if we need to more precisely fetch LFS files for new tags or if git is smart
-          rm .lfsconfig
-
-          echo ">>> Fetching all branches references down locally so subsequent commands can see them..."
-          git fetch source '+refs/heads/*:refs/heads/*' --update-head-ok
-
-          echo ">>> Print out all branches"
-          git --no-pager branch -a -vv
-
-          echo ">>> Pushing git changes and LFS content..."
-          git push destination "${SOURCE_BRANCH}:${DESTINATION_BRANCH}" -f
+          echo ">>> Pushing git tags..."
+          git push destination "${SOURCE_TAGS}:${DESTINATION_TAGS}" -f
 
           shred --remove .git/config

.gitignore

@@ -25,6 +25,7 @@ dieharder_run.log
 
 # Cuda local build
 backends/tfhe-cuda-backend/cuda/cmake-build-debug/
+backends/tfhe-cuda-backend/cuda/build/
 
 # WASM tests
 tfhe/web_wasm_parallel_tests/server.PID
@@ -34,6 +35,9 @@ node_modules/
 package-lock.json
 utils/wasm-par-mq/examples/*/pkg/
 
+# Commit lock files of backward data generation crates
+!utils/tfhe-backward-compat-data/crates/generate_*/Cargo.lock
+
 # Python .env
 .env
 __pycache__

.linelint.yml

@@ -12,6 +12,7 @@ ignore:
   - utils/tfhe-lints/**/main.stderr
   - utils/tfhe-lints/**/*.json
   - utils/tfhe-backward-compat-data/**/*.ron # ron files are autogenerated
+  - tests/corrupted_inputs_deserialization/data/proven_compact_list/**/metadata.txt
 
 rules:
   # checks if file ends in a newline character

Cargo.toml

@@ -14,10 +14,12 @@ members = [
     "tfhe-fft",
     "tfhe-ntt",
     "tfhe-zk-pok",
+    "utils/benchmark_spec",
     "utils/param_dedup",
     "utils/tfhe-backward-compat-checker",
     "utils/tfhe-backward-compat-data",
     "utils/tfhe-backward-compat-data/crates/add_new_version",
+    "utils/tfhe-safe-serialize",
     "utils/tfhe-versionable",
     "utils/tfhe-versionable-derive",
     "utils/wasm-par-mq",
@@ -43,6 +45,7 @@ rand = "0.8"
 rayon = "1.11"
 serde = { version = "1.0", default-features = false }
 wasm-bindgen = { version = "0.2.114" }
+wasm-bindgen-futures = { version = "0.4.56" }
 # js-sys (at this point in time) automatically enables the unsafe-eval feature which we do not want
 # this does not prevent other deps from enabling it, but it at least conveys our need to not have it
 # we still enable std, which was part of default before

LICENSE

@@ -1,6 +1,6 @@
 BSD 3-Clause Clear License
 
-Copyright © 2025 ZAMA.
+Copyright © 2026 ZAMA.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without modification,

Makefile

@@ -26,6 +26,7 @@ BENCH_CUSTOM_COMMAND:=
 NODE_VERSION=24.12
 BACKWARD_COMPAT_DATA_DIR=utils/tfhe-backward-compat-data
 BACKWARD_COMPAT_DATA_GEN_VERSION:=$(TFHE_VERSION)
+CORRUPTED_INPUTS_TEST=tests/corrupted_inputs_deserialization
 TEST_VECTORS_DIR=apps/test-vectors
 CURRENT_TFHE_VERSION:=$(shell grep '^version[[:space:]]*=' tfhe/Cargo.toml | cut -d '=' -f 2 | xargs)
 WASM_PACK_VERSION="0.13.1"
@@ -121,6 +122,12 @@ install_build_wasm32_target:
 	( echo "Unable to install wasm32-unknown-unknown target toolchain, check your rustup installation. \
 	Rustup can be downloaded at https://rustup.rs/" && exit 1 )
 
+.PHONY: install_check_wasm32_target # Install the wasm32 toolchain used for checks
+install_check_wasm32_target:
+	rustup target add wasm32-unknown-unknown --toolchain "$(RS_CHECK_TOOLCHAIN)" || \
+	( echo "Unable to install wasm32-unknown-unknown target toolchain, check your rustup installation. \
+	Rustup can be downloaded at https://rustup.rs/" && exit 1 )
+
 .PHONY: install_cargo_nextest # Install cargo nextest used for shortint tests
 install_cargo_nextest:
 	@cargo nextest --version > /dev/null 2>&1 || \
@@ -305,8 +312,10 @@ semgrep_and_lint_gpu_code: semgrep_lint_setup_venv
 	find "$(TFHECUDA_SRC)" -name '*.h' -o -name '*.cuh' -o -name '*.cu' \
 		| grep -v '/cmake-build-debug/' \
 		| grep -v '/build/' \
-		| xargs venv/bin/semgrep --config "$(TFHECUDA_SRC)/.semgrep/release-ordering.yaml" --scan-unknown-extensions
+		| xargs venv/bin/semgrep --error --config "$(TFHECUDA_SRC)/.semgrep/release-ordering.yaml" --scan-unknown-extensions
 	venv/bin/python3 "scripts/check_scratch_cleanup.py"
+	@# Split the search string using shell string concatenation so the Makefile line doesn't match itself
+	! git ls-files | xargs grep -n 'TODO: ADD COMM''ENT'
 
 .PHONY: semver_check_cuda_backend # Run semver checks on tfhe-cuda-backend
 semver_check_cuda_backend:
@@ -349,23 +358,23 @@ check_fmt_js: check_nvm_installed
 .PHONY: check_fmt_toml # Check TOML files format
 check_fmt_toml: install_taplo
 	@RUST_LOG=warn taplo fmt --check || \
-	echo "TOML files format check failed. Please run 'make fmt_toml'"
+	{ echo "TOML files format check failed. Please run 'make fmt_toml'"; exit 1; }
 
 .PHONY: check_typos # Check for typos in codebase
 check_typos: install_typos_checker
-	@typos && echo "No typos found"
+	@git ls-files ":!*.png" ":!*.cbor" ":!*.bcode" ":!*.ico" ":!*/twiddles.cu" | typos --file-list - && echo "No typos found"
 
 .PHONY: clippy_gpu # Run clippy lints on tfhe with "gpu" enabled
 clippy_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
-		--features=boolean,shortint,integer,internal-keycache,gpu,pbs-stats,extended-types,zk-pok \
+		--features=boolean,shortint,integer,internal-keycache,gpu,gpu-experimental-zk,pbs-stats,extended-types,zk-pok \
 		--all-targets \
 		-p tfhe -- --no-deps -D warnings
 
 .PHONY: check_gpu # Run check on tfhe with "gpu" enabled
 check_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" check \
-		--features=boolean,shortint,integer,internal-keycache,gpu,pbs-stats \
+		--features=boolean,shortint,integer,internal-keycache,gpu,gpu-experimental-zk,pbs-stats \
 		--all-targets \
 		-p tfhe
 
@@ -379,7 +388,7 @@ clippy_hpu: install_rs_check_toolchain
 .PHONY: clippy_gpu_hpu # Run clippy lints on tfhe with "gpu" and "hpu" enabled
 clippy_gpu_hpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
-		--features=boolean,shortint,integer,internal-keycache,gpu,hpu,pbs-stats,extended-types,zk-pok \
+		--features=boolean,shortint,integer,internal-keycache,gpu,gpu-experimental-zk,hpu,pbs-stats,extended-types,zk-pok \
 		--all-targets \
 		-p tfhe -- --no-deps -D warnings
 
@@ -472,7 +481,7 @@ clippy_rustdoc_gpu: install_rs_check_toolchain
 	fi && \
 	CARGO_TERM_QUIET=true CLIPPYFLAGS="-D warnings" RUSTDOCFLAGS="--no-run --test-builder ./scripts/clippy_driver.sh -Z unstable-options" \
 		cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" test --doc \
-		--features=boolean,shortint,integer,zk-pok,pbs-stats,strings,experimental,gpu \
+		--features=boolean,shortint,integer,zk-pok,pbs-stats,strings,experimental,gpu,gpu-experimental-zk \
 		-p tfhe -- --nocapture
 
 .PHONY: clippy_c_api # Run clippy lints enabling the boolean, shortint and the C API
@@ -483,11 +492,17 @@ clippy_c_api: install_rs_check_toolchain
 
 .PHONY: clippy_js_wasm_api # Run clippy lints enabling the boolean, shortint, integer and the js wasm API
 clippy_js_wasm_api: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
+		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api,extended-types \
+		-p tfhe -- --no-deps -D warnings
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
 		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api,zk-pok,extended-types \
 		-p tfhe -- --no-deps -D warnings
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
-		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api,extended-types \
+		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api,zk-pok,extended-types,parallel-wasm-api \
+		-p tfhe -- --no-deps -D warnings
+	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
+		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api,zk-pok,extended-types,cross-origin-wasm-api \
 		-p tfhe -- --no-deps -D warnings
 
 .PHONY: clippy_tasks # Run clippy lints on helper tasks crate.
@@ -503,7 +518,7 @@ clippy_trivium: install_rs_check_toolchain
 .PHONY: clippy_ws_tests # Run clippy on the workspace level tests
 clippy_ws_tests: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --tests \
-		-p tests --features=shortint,integer,zk-pok -- --no-deps -D warnings
+		-p tests --features=shortint,integer,zk-pok,strings -- --no-deps -D warnings
 
 .PHONY: clippy_all_targets # Run clippy lints on all targets (benches, examples, etc.)
 clippy_all_targets: install_rs_check_toolchain
@@ -528,6 +543,15 @@ clippy_zk_pok: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
 		-p tfhe-zk-pok --features=experimental -- --no-deps -D warnings
 
+.PHONY: clippy_zk_pok_wasm # Run clippy lints on tfhe-zk-pok for wasm32 target
+clippy_zk_pok_wasm: install_rs_check_toolchain install_check_wasm32_target
+	RUSTFLAGS="$(WASM_RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
+		--target wasm32-unknown-unknown \
+		-p tfhe-zk-pok -- --no-deps -D warnings
+	RUSTFLAGS="$(WASM_RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
+		--target wasm32-unknown-unknown \
+		-p tfhe-zk-pok --features cross-origin-wasm -- --no-deps -D warnings
+
 .PHONY: clippy_versionable # Run clippy lints on tfhe-versionable
 clippy_versionable: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
@@ -535,6 +559,11 @@ clippy_versionable: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
 		-p tfhe-versionable -- --no-deps -D warnings
 
+.PHONY: clippy_safe_serialize # Run clippy lints on tfhe-safe-serialize
+clippy_safe_serialize: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
+		-p tfhe-safe-serialize -- --no-deps -D warnings
+
 .PHONY: clippy_param_dedup # Run clippy lints on param_dedup tool
 clippy_param_dedup: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
@@ -565,10 +594,12 @@ clippy_test_vectors: install_rs_check_toolchain
 	cd apps/test-vectors; RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
 		-p tfhe-test-vectors -- --no-deps -D warnings
 
+# WARNING: This target is not directly run in CI. When adding a subtarget here,
+# MAKE SURE TO ALSO ADD IT TO A PCC BATCH BELOW
 .PHONY: clippy_all # Run all clippy targets
 clippy_all: clippy_rustdoc clippy clippy_boolean clippy_shortint clippy_integer clippy_all_targets \
-clippy_c_api clippy_js_wasm_api clippy_tasks clippy_core clippy_tfhe_csprng clippy_zk_pok clippy_trivium \
-clippy_versionable clippy_tfhe_lints clippy_ws_tests clippy_bench clippy_param_dedup \
+clippy_c_api clippy_js_wasm_api clippy_tasks clippy_core clippy_tfhe_csprng clippy_zk_pok clippy_zk_pok_wasm clippy_trivium \
+clippy_versionable clippy_safe_serialize clippy_tfhe_lints clippy_ws_tests clippy_bench clippy_param_dedup \
 clippy_test_vectors clippy_backward_compat_data clippy_wasm_par_mq
 
 .PHONY: clippy_fast # Run main clippy targets
@@ -665,7 +696,7 @@ build_c_api: install_rs_check_toolchain
 .PHONY: build_c_api_gpu # Build the C API for boolean, shortint and integer
 build_c_api_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) build --profile $(CARGO_PROFILE) \
-		--features=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,extended-types,gpu \
+		--features=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,extended-types,gpu,gpu-experimental-zk \
 		-p tfhe
 
 .PHONY: build_c_api_experimental_deterministic_fft # Build the C API for boolean, shortint and integer with experimental deterministic FFT
@@ -674,11 +705,14 @@ build_c_api_experimental_deterministic_fft: install_rs_check_toolchain
 		--features=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,experimental-force_fft_algo_dif4 \
 		-p tfhe
 
-.PHONY: build_web_js_api # Build the js API targeting the web browser
+.PHONY: build_web_js_api # Build the js API targeting the web browser, in sequential or cross origin parallelism modes.
 build_web_js_api: install_wasm_pack
 	cd tfhe && \
 	RUSTFLAGS="$(WASM_RUSTFLAGS)" wasm-pack build --release --target=web \
-		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok,extended-types
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok,extended-types,cross-origin-wasm-api && \
+	find pkg/snippets -type f -iname worker_helpers.js -exec sed -i 's|import("../../..")|import("../../../tfhe.js")|g' {} \;
+	cp utils/wasm-par-mq/js/coordinator.js tfhe/pkg/
+	jq '.files += ["snippets"]' tfhe/pkg/package.json > tmp_pkg.json && mv -f tmp_pkg.json tfhe/pkg/package.json
 
 .PHONY: build_web_js_api_parallel # Build the js API targeting the web browser with parallelism support
 # parallel wasm requires specific build options, see https://github.com/rust-lang/rust/pull/147225
@@ -764,7 +798,7 @@ test_zk_cuda_backend:
 
 
 .PHONY: test_gpu # Run the tests of the core_crypto module including experimental on the gpu backend
-test_gpu: test_core_crypto_gpu test_integer_gpu test_cuda_backend
+test_gpu: test_core_crypto_gpu test_integer_gpu test_cuda_backend test_zk_cuda_backend
 
 .PHONY: test_core_crypto_gpu # Run the tests of the core_crypto module including experimental on the gpu backend
 test_core_crypto_gpu:
@@ -1200,12 +1234,31 @@ test_tfhe_csprng_big_endian: install_cargo_cross
 	RUSTFLAGS="" cross test --profile $(CARGO_PROFILE) \
 		-p tfhe-csprng --target=powerpc64-unknown-linux-gnu
 
-
 .PHONY: test_zk_pok # Run tfhe-zk-pok tests
 test_zk_pok:
 	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
 		-p tfhe-zk-pok --features experimental
 
+.PHONY: test_zk_pok_experimental_gpu # Run tfhe-zk-pok GPU-accelerated tests
+test_zk_pok_experimental_gpu:
+	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
+		-p tfhe-zk-pok --features experimental,gpu-experimental -- gpu
+
+.PHONY: test_integer_zk_gpu # Run tfhe-zk-pok tests
+test_integer_zk_gpu:
+	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
+		--features=integer,zk-pok,gpu -p tfhe -- \
+		integer::gpu::zk::
+
+.PHONY: test_integer_zk_experimental_gpu # Run tfhe-zk-pok tests
+test_integer_zk_experimental_gpu:
+	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
+		--features=integer,zk-pok,gpu,gpu-experimental-zk -p tfhe -- \
+		integer::gpu::zk::
+
+.PHONY: test_zk_cuda # Run all GPU MSM integration tests (CPU vs GPU comparison + integration test)
+test_zk_cuda: test_zk_cuda_backend test_zk_pok_experimental_gpu test_integer_zk_gpu test_integer_zk_experimental_gpu
+
 .PHONY: test_zk_wasm_x86_compat_ci
 test_zk_wasm_x86_compat_ci: check_nvm_installed
 	source ~/.nvm/nvm.sh && \
@@ -1224,6 +1277,11 @@ test_versionable:
 	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
 		--all-targets -p tfhe-versionable
 
+.PHONY: test_safe_serialize # Run tests for tfhe-safe-serialize subcrate
+test_safe_serialize:
+	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
+		--all-targets -p tfhe-safe-serialize
+
 # The backward compat data folder holds historical binary data but also rust code to generate and load them.
 .PHONY: gen_backward_compat_data # Re-generate backward compatibility data
 gen_backward_compat_data:
@@ -1237,11 +1295,19 @@ new_backward_compat_crate:
 .PHONY: test_backward_compatibility_ci
 test_backward_compatibility_ci:
 	TFHE_BACKWARD_COMPAT_DATA_DIR="../$(BACKWARD_COMPAT_DATA_DIR)" RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
-		--features=shortint,integer,zk-pok -p tests test_backward_compatibility -- --nocapture
+		--features=shortint,integer,zk-pok,strings -p tests test_backward_compatibility -- --nocapture
 
 .PHONY: test_backward_compatibility # Same as test_backward_compatibility_ci but tries to clone the data repo first if needed
 test_backward_compatibility: pull_backward_compat_data test_backward_compatibility_ci
 
+.PHONY: test_corrupted_inputs_ci
+test_corrupted_inputs_ci:
+	RUSTFLAGS="$(RUSTFLAGS)" cargo test --profile $(CARGO_PROFILE) \
+		--features=integer,zk-pok,strings -p tests test_corrupted_inputs_deserialization -- --nocapture
+
+.PHONY: test_corrupted_inputs # Same as test_corrupted_inputs_ci but pulls data first
+test_corrupted_inputs: pull_corrupted_inputs_data test_corrupted_inputs_ci
+
 # Generate the test vectors and update the hash file
 .PHONY: gen_test_vectors
 gen_test_vectors:
@@ -1350,6 +1416,19 @@ test_nodejs_wasm_api_ci: build_node_js_api
 
 # This is an internal target, not meant to be called on its own.
 run_web_js_api_parallel: build_web_js_api_parallel setup_venv
+	cd $(WEB_SERVER_DIR) && npm install && npm run build
+	source venv/bin/activate && \
+	python ci/webdriver.py \
+	--browser-path $(browser_path) \
+	--driver-path $(driver_path) \
+	--browser-kind  $(browser_kind) \
+	--server-cmd $(server_cmd) \
+	--server-workdir "$(WEB_SERVER_DIR)" \
+	--id-pattern $(filter) \
+	--id-exclude-pattern asyncMainThread
+
+# This is an internal target, not meant to be called on its own.
+run_web_js_api_cross_origin: build_web_js_api setup_venv
 	cd $(WEB_SERVER_DIR) && npm install && npm run build
 	source venv/bin/activate && \
 	python ci/webdriver.py \
@@ -1392,6 +1471,38 @@ test_web_js_api_parallel_firefox_ci: setup_venv
 	nvm use $(NODE_VERSION) && \
 	$(MAKE) test_web_js_api_parallel_firefox
 
+test_web_js_api_cross_origin_chrome: browser_path = "$(WEB_RUNNER_DIR)/chrome/chrome-linux64/chrome"
+test_web_js_api_cross_origin_chrome: driver_path = "$(WEB_RUNNER_DIR)/chrome/chromedriver-linux64/chromedriver"
+test_web_js_api_cross_origin_chrome: browser_kind = chrome
+test_web_js_api_cross_origin_chrome: server_cmd = "npm run server:cross-origin"
+test_web_js_api_cross_origin_chrome: filter = ZeroKnowledgeTest # Only run zk proof tests in cross-origin mode
+
+.PHONY: test_web_js_api_cross_origin_chrome # Run tests for the web wasm api in cross-origin mode on Chrome
+test_web_js_api_cross_origin_chrome: run_web_js_api_cross_origin
+
+.PHONY: test_web_js_api_cross_origin_chrome_ci # Run tests for the web wasm api in cross-origin mode on Chrome
+test_web_js_api_cross_origin_chrome_ci: setup_venv
+	source ~/.nvm/nvm.sh && \
+	nvm install $(NODE_VERSION) && \
+	nvm use $(NODE_VERSION) && \
+	$(MAKE) test_web_js_api_cross_origin_chrome
+
+test_web_js_api_cross_origin_firefox: browser_path = "$(WEB_RUNNER_DIR)/firefox/firefox/firefox"
+test_web_js_api_cross_origin_firefox: driver_path = "$(WEB_RUNNER_DIR)/firefox/geckodriver"
+test_web_js_api_cross_origin_firefox: browser_kind = firefox
+test_web_js_api_cross_origin_firefox: server_cmd = "npm run server:cross-origin"
+test_web_js_api_cross_origin_firefox: filter = ZeroKnowledgeTest  # Only run zk proof tests in cross-origin mode
+
+.PHONY: test_web_js_api_cross_origin_firefox # Run tests for the web wasm api in cross-origin mode on Firefox
+test_web_js_api_cross_origin_firefox: run_web_js_api_cross_origin
+
+.PHONY: test_web_js_api_cross_origin_firefox_ci # Run tests for the web wasm api in cross-origin mode on Firefox
+test_web_js_api_cross_origin_firefox_ci: setup_venv
+	source ~/.nvm/nvm.sh && \
+	nvm install $(NODE_VERSION) && \
+	nvm use $(NODE_VERSION) && \
+	$(MAKE) test_web_js_api_cross_origin_firefox
+
 WASM_PAR_MQ_TEST_DIR=utils/wasm-par-mq/web_tests
 
 .PHONY: build_wasm_par_mq_tests # Build the wasm-par-mq test WASM package
@@ -1548,27 +1659,57 @@ bench_integer_rerand: install_rs_check_toolchain
 	--bench integer-rerand \
 	--features=integer,internal-keycache,pbs-stats -p tfhe-benchmark --
 
+.PHONY: bench_integer_rerand_gpu # Run benchmarks for integer rerand on GPU backend
+bench_integer_rerand_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench integer-rerand \
+	--features=integer,internal-keycache,gpu,pbs-stats -p tfhe-benchmark --profile release_lto_off --
+
+.PHONY: bench_msm_zk
+bench_msm_zk: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench zk-msm \
+	--features=zk-pok -p tfhe-benchmark --profile release --
+
+# GPU benchmarks need --profile release for correct measurements
+.PHONY: bench_msm_zk_gpu
+bench_msm_zk_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench zk-msm \
+	--features=gpu,gpu-experimental-zk,zk-pok -p tfhe-benchmark --profile release -- zk::cuda::msm
+
+# GPU benchmarks need --profile release for correct measurements
 .PHONY: bench_integer_zk_gpu
 bench_integer_zk_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_BENCH_BIT_SIZES_SET=$(BIT_SIZES_SET) __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) \
-	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-zk-pke \
-	--features=integer,internal-keycache,gpu,pbs-stats,zk-pok -p tfhe-benchmark --profile release_lto_off --
+	--features=integer,internal-keycache,gpu,pbs-stats,zk-pok -p tfhe-benchmark --profile release --
+
+# GPU benchmarks need --profile release for correct measurements
+.PHONY: bench_integer_zk_experimental_gpu
+bench_integer_zk_experimental_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_BENCH_BIT_SIZES_SET=$(BIT_SIZES_SET) __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench integer-zk-pke \
+	--features=integer,internal-keycache,gpu,gpu-experimental-zk,pbs-stats,zk-pok -p tfhe-benchmark --profile release --
 
 .PHONY: bench_integer_aes_gpu # Run benchmarks for AES on GPU backend
 bench_integer_aes_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-aes \
-	--features=integer,internal-keycache,gpu, -p tfhe-benchmark --profile release_lto_off --
+	--features=integer,internal-keycache,gpu -p tfhe-benchmark --profile release_lto_off --
 
 .PHONY: bench_integer_aes256_gpu # Run benchmarks for AES256 on GPU backend
 bench_integer_aes256_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-aes256 \
-	--features=integer,internal-keycache,gpu, -p tfhe-benchmark --profile release_lto_off --
+	--features=integer,internal-keycache,gpu -p tfhe-benchmark --profile release_lto_off --
 
 .PHONY: bench_integer_trivium_gpu # Run benchmarks for trivium on GPU backend
 bench_integer_trivium_gpu: install_rs_check_toolchain
@@ -1732,37 +1873,37 @@ bench_web_js_api_parallel_firefox_ci: setup_venv
 	nvm use $(NODE_VERSION) && \
 	$(MAKE) bench_web_js_api_parallel_firefox
 
-bench_web_js_api_unsafe_coop_chrome: browser_path = "$(WEB_RUNNER_DIR)/chrome/chrome-linux64/chrome"
-bench_web_js_api_unsafe_coop_chrome: driver_path = "$(WEB_RUNNER_DIR)/chrome/chromedriver-linux64/chromedriver"
-bench_web_js_api_unsafe_coop_chrome: browser_kind = chrome
-bench_web_js_api_unsafe_coop_chrome: server_cmd = "npm run server:unsafe-coop"
-bench_web_js_api_unsafe_coop_chrome: filter = ZeroKnowledgeBench # Only bench zk with unsafe coop
+bench_web_js_api_cross_origin_chrome: browser_path = "$(WEB_RUNNER_DIR)/chrome/chrome-linux64/chrome"
+bench_web_js_api_cross_origin_chrome: driver_path = "$(WEB_RUNNER_DIR)/chrome/chromedriver-linux64/chromedriver"
+bench_web_js_api_cross_origin_chrome: browser_kind = chrome
+bench_web_js_api_cross_origin_chrome: server_cmd = "npm run server:cross-origin"
+bench_web_js_api_cross_origin_chrome: filter = ZeroKnowledgeBench # Only bench zk with cross-origin workers
 
-.PHONY: bench_web_js_api_unsafe_coop_chrome # Run benchmarks for the web wasm api without cross-origin isolation
-bench_web_js_api_unsafe_coop_chrome: run_web_js_api_parallel
+.PHONY: bench_web_js_api_cross_origin_chrome # Run benchmarks for the web wasm api without cross-origin isolation
+bench_web_js_api_cross_origin_chrome: run_web_js_api_cross_origin
 
-.PHONY: bench_web_js_api_unsafe_coop_chrome_ci # Run benchmarks for the web wasm api without cross-origin isolation
-bench_web_js_api_unsafe_coop_chrome_ci: setup_venv
+.PHONY: bench_web_js_api_cross_origin_chrome_ci # Run benchmarks for the web wasm api without cross-origin isolation
+bench_web_js_api_cross_origin_chrome_ci: setup_venv
 	source ~/.nvm/nvm.sh && \
 	nvm install $(NODE_VERSION) && \
 	nvm use $(NODE_VERSION) && \
-	$(MAKE) bench_web_js_api_unsafe_coop_chrome
+	$(MAKE) bench_web_js_api_cross_origin_chrome
 
-bench_web_js_api_unsafe_coop_firefox: browser_path = "$(WEB_RUNNER_DIR)/firefox/firefox/firefox"
-bench_web_js_api_unsafe_coop_firefox: driver_path = "$(WEB_RUNNER_DIR)/firefox/geckodriver"
-bench_web_js_api_unsafe_coop_firefox: browser_kind = firefox
-bench_web_js_api_unsafe_coop_firefox: server_cmd = "npm run server:unsafe-coop"
-bench_web_js_api_unsafe_coop_firefox: filter = ZeroKnowledgeBench # Only bench zk with unsafe coop
+bench_web_js_api_cross_origin_firefox: browser_path = "$(WEB_RUNNER_DIR)/firefox/firefox/firefox"
+bench_web_js_api_cross_origin_firefox: driver_path = "$(WEB_RUNNER_DIR)/firefox/geckodriver"
+bench_web_js_api_cross_origin_firefox: browser_kind = firefox
+bench_web_js_api_cross_origin_firefox: server_cmd = "npm run server:cross-origin"
+bench_web_js_api_cross_origin_firefox: filter = ZeroKnowledgeBench # Only bench zk with cross-origin workers
 
-.PHONY: bench_web_js_api_unsafe_coop_firefox # Run benchmarks for the web wasm api without cross-origin isolation
-bench_web_js_api_unsafe_coop_firefox: run_web_js_api_parallel
+.PHONY: bench_web_js_api_cross_origin_firefox # Run benchmarks for the web wasm api without cross-origin isolation
+bench_web_js_api_cross_origin_firefox: run_web_js_api_cross_origin
 
-.PHONY: bench_web_js_api_unsafe_coop_firefox_ci # Run benchmarks for the web wasm api without cross-origin isolation
-bench_web_js_api_unsafe_coop_firefox_ci: setup_venv
+.PHONY: bench_web_js_api_cross_origin_firefox_ci # Run benchmarks for the web wasm api without cross-origin isolation
+bench_web_js_api_cross_origin_firefox_ci: setup_venv
 	source ~/.nvm/nvm.sh && \
 	nvm install $(NODE_VERSION) && \
 	nvm use $(NODE_VERSION) && \
-	$(MAKE) bench_web_js_api_unsafe_coop_firefox
+	$(MAKE) bench_web_js_api_cross_origin_firefox
 
 .PHONY: bench_hlapi_unsigned # Run benchmarks for integer operations
 bench_hlapi_unsigned: install_rs_check_toolchain
@@ -1795,27 +1936,61 @@ bench_hlapi_hpu: install_rs_check_toolchain
 	--bench hlapi \
 	--features=integer,internal-keycache,hpu,hpu-v80,pbs-stats -p tfhe-benchmark --
 
-.PHONY: bench_hlapi_erc20 # Run benchmarks for ERC20 operations
-bench_hlapi_erc20: install_rs_check_toolchain
+.PHONY: bench_hlapi_erc7984 # Run benchmarks for ERC7984 operations
+bench_hlapi_erc7984: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench hlapi-erc20 \
+	--bench hlapi-erc7984 \
 	--features=integer,internal-keycache,pbs-stats -p tfhe-benchmark --
 
-.PHONY: bench_hlapi_erc20_gpu # Run benchmarks for ERC20 operations on GPU
-bench_hlapi_erc20_gpu: install_rs_check_toolchain
+.PHONY: bench_hlapi_erc7984_gpu # Run benchmarks for ERC7984 operations on GPU
+bench_hlapi_erc7984_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench hlapi-erc20 \
+	--bench hlapi-erc7984 \
 	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off --
 
-.PHONY: bench_hlapi_erc20_gpu_classical # Run benchmarks for ERC20 operations on GPU with classical parameters
-bench_hlapi_erc20_gpu_classical: install_rs_check_toolchain
+.PHONY: bench_hlapi_erc7984_gpu_classical # Run benchmarks for ERC7984 operations on GPU with classical parameters
+bench_hlapi_erc7984_gpu_classical: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_PARAM_TYPE=classical \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench hlapi-erc20 \
+	--bench hlapi-erc7984 \
 	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off --
 
+.PHONY: bench_hlapi_erc7984_multi_group_gpu # Runs ERC7984 bench in two processes (half of gpus for each) and aggregates results
+bench_hlapi_erc7984_multi_group_gpu: install_rs_check_toolchain
+	# This next line must be kept here: the code can not remove this file without a risk of concurrency issues
+	# we don't know which process starts first and which one deletes the files - file deletion may also be not atomic)
+	rm -f /dev/shm/sem.tfhe_bench_*
+	NUM_GROUPS=$${__TFHE_RS_BENCH_MULTI_PROC_GROUPS:-2}; \
+	[ "$$NUM_GROUPS" -ge 2 ] || { echo "Error: __TFHE_RS_BENCH_MULTI_PROC_GROUPS must be at least 2, got $$NUM_GROUPS"; exit 1; }; \
+	trap "echo 'User interrupted the benchmark, stopping all workers!'; rm -f /dev/shm/sem.tfhe_bench_*; kill 0" INT TERM; \
+	for i in $$(seq 0 $$((NUM_GROUPS - 1))); do \
+		GPU_LIST=$$(python3 ci/split_gpus.py $$i $$NUM_GROUPS) || exit 1; \
+		echo "Starting benchmark group $$i with CUDA_VISIBLE_DEVICES=$$GPU_LIST"; \
+		CUDA_VISIBLE_DEVICES=$$GPU_LIST CARGO_TARGET_DIR=target_p$$i RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) __TFHE_RS_BENCH_GPU_PROCESS_COUNT=$$NUM_GROUPS \
+		cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+		--bench hlapi-erc7984 \
+		--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off -- '::transfer::overflow' & \
+	done; \
+	wait
+
+.PHONY: bench_hlapi_erc7984_multi_group_fake_multi_gpu # Runs ERC7984 bench in two processes in parallel on a single GPU (use to debug bench_hlapi_erc7984_multi_group_gpu)
+bench_hlapi_erc7984_multi_group_fake_multi_gpu: install_rs_check_toolchain
+	# This next line must be kept here: the code can not remove this file without a risk of concurrency issues
+	# we don't know which process starts first and which one deletes the files - file deletion may also be not atomic)
+	rm -f /dev/shm/sem.tfhe_bench_*
+	trap "echo 'User interrupted the benchmark, stopping all workers!'; rm -f /dev/shm/sem.tfhe_bench_*; kill 0" INT TERM; \
+	CARGO_TARGET_DIR=target_p0 RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=throughput __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) __TFHE_RS_BENCH_GPU_PROCESS_COUNT=2 \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench hlapi-erc7984 \
+	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off -- '::transfer::overflow' & \
+	CARGO_TARGET_DIR=target_p1 RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=throughput __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) __TFHE_RS_BENCH_GPU_PROCESS_COUNT=2 \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench hlapi-erc7984 \
+	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off -- '::transfer::overflow' & \
+	wait
+
 .PHONY: bench_hlapi_dex # Run benchmarks for DEX operations
 bench_hlapi_dex: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
@@ -1837,13 +2012,13 @@ bench_hlapi_dex_gpu_classical: install_rs_check_toolchain
 	--bench hlapi-dex \
 	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off --
 
-.PHONY: bench_hlapi_erc20_hpu # Run benchmarks for ECR20 operations on HPU
-bench_hlapi_erc20_hpu: install_rs_check_toolchain
+.PHONY: bench_hlapi_erc7984_hpu # Run benchmarks for ECR20 operations on HPU
+bench_hlapi_erc7984_hpu: install_rs_check_toolchain
 	source ./setup_hpu.sh --config $(HPU_CONFIG); \
 	export V80_PCIE_DEV=${V80_PCIE_DEV}; \
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench hlapi-erc20 \
+	--bench hlapi-erc7984 \
 	--features=integer,internal-keycache,hpu,hpu-v80,pbs-stats -p tfhe-benchmark --
 
 .PHONY: bench_tfhe_zk_pok # Run benchmarks for the tfhe_zk_pok crate
@@ -1851,6 +2026,13 @@ bench_tfhe_zk_pok: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench -p tfhe-zk-pok --
 
+.PHONY: bench_tfhe_zk_pok_gpu # Run benchmarks for the tfhe_zk_pok crate using GPU acceleration
+bench_tfhe_zk_pok_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--package tfhe-zk-pok \
+	--features=gpu-experimental --profile release
+
 .PHONY: bench_hlapi_noise_squash # Run benchmarks for noise squash operation
 bench_hlapi_noise_squash: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_BENCH_BIT_SIZES_SET=$(BIT_SIZES_SET) \
@@ -1892,10 +2074,10 @@ bench_summary: install_rs_check_toolchain
 	--bench hlapi-noise-squash \
 	--features=integer,internal-keycache,pbs-stats -p tfhe-benchmark -- '::decomp_noise_squash_comp::'
 
-	# ERC20
+	# ERC7984
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench hlapi-erc20 \
+	--bench hlapi-erc7984 \
 	--features=integer,internal-keycache -p tfhe-benchmark -- '::transfer::overflow'
 
 	# DEX
@@ -1937,11 +2119,16 @@ bench_summary_gpu: install_rs_check_toolchain
 	--bench hlapi-noise-squash \
 	--features=integer,gpu,internal-keycache,pbs-stats -p tfhe-benchmark --profile release_lto_off -- '::decomp_noise_squash_comp::'
 
-	# ERC20
-	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) \
+	# This make target only runs the latency benchmark. This is because
+	# summary benchmarks must use the multi-process-multi-group throughput target
+	# to measure throughput. That target must be followed by specific post-processing steps.
+	# Thus that target is run in a separate step in benchmark_summary.yml.
+ifneq ($(filter latency both,$(BENCH_TYPE)),)
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=latency __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench hlapi-erc20 \
+	--bench hlapi-erc7984 \
 	--features=integer,gpu,internal-keycache -p tfhe-benchmark --profile release_lto_off -- '::transfer::overflow'
+endif
 
 	# DEX
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE)  __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) \
@@ -2033,6 +2220,10 @@ write_params_to_file: install_rs_check_toolchain
 pull_backward_compat_data:
 	./scripts/pull_lfs_data.sh $(BACKWARD_COMPAT_DATA_DIR)
 
+.PHONY: pull_corrupted_inputs_data # Pull the data files needed for corrupted inputs deserialization tests
+pull_corrupted_inputs_data:
+	./scripts/pull_lfs_data.sh $(CORRUPTED_INPUTS_TEST)
+
 .PHONY: pull_hpu_files # Pull the hpu files
 pull_hpu_files:
 	./scripts/pull_lfs_data.sh backends/tfhe-hpu-backend/
@@ -2123,8 +2314,10 @@ pcc_batch_6:
 	$(call run_recipe_with_details,clippy_tasks)
 	$(call run_recipe_with_details,clippy_tfhe_csprng)
 	$(call run_recipe_with_details,clippy_zk_pok)
+	$(call run_recipe_with_details,clippy_zk_pok_wasm)
 	$(call run_recipe_with_details,clippy_trivium)
 	$(call run_recipe_with_details,clippy_versionable)
+	$(call run_recipe_with_details,clippy_safe_serialize)
 	$(call run_recipe_with_details,clippy_param_dedup)
 	$(call run_recipe_with_details,docs)
 

_typos.toml

@@ -15,12 +15,3 @@ extend-ignore-identifiers-re = [
     "0x[0-9a-fA-F]+",
     "xrt_coreutil",
 ]
-
-[files]
-extend-exclude = [
-    "backends/tfhe-cuda-backend/cuda/src/fft128/twiddles.cu",
-    "backends/tfhe-cuda-backend/cuda/src/fft/twiddles.cu",
-    "backends/tfhe-hpu-backend/config_store/**/*.link_summary",
-    "*.cbor",
-    "*.bcode",
-]

backends/tfhe-cuda-backend/LICENSE

@@ -1,6 +1,6 @@
 BSD 3-Clause Clear License
 
-Copyright © 2025 ZAMA.
+Copyright © 2026 ZAMA.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without modification,

backends/tfhe-cuda-backend/cuda/.semgrep/release-ordering.yaml

@@ -62,3 +62,29 @@ rules:
             cuda_synchronize_stream(...);
             ...
           }
+
+  - id: tfhe-cuda-unwrapped-cuda-runtime-call
+    message: "CUDA runtime API call is not wrapped in `check_cuda_error(...)`."
+    severity: WARNING
+    languages: [c, cpp]
+    options:
+      generic_ellipsis_max_span: 500
+    paths:
+      include:
+        - "*.cu"
+        - "*.cuh"
+        - "*.cpp"
+        - "*.h"
+      exclude:
+        - backends/tfhe-cuda-backend/cuda/check_cuda.cu # contains cuda checking functions
+        - backends/tfhe-cuda-backend/cuda/include/device.h # contains the cuda_check_error macro (and others)
+    patterns:
+      - pattern: $FUNC(...)
+      - metavariable-regex:
+          metavariable: $FUNC
+          regex: "^cuda[A-Z][A-Za-z0-9]*$" # matches cudaMalloc/cudaMemcpy/... (not project helpers like cuda_set_device)
+      - pattern-not-inside: check_cuda_error(...)
+      - pattern-not-inside: |
+          $FUNC(...);
+          check_cuda_error(cudaGetLastError());
+      - pattern-not-inside: $FUNC(...) == $VAL

backends/tfhe-cuda-backend/cuda/include/ciphertext.h

@@ -36,5 +36,19 @@ void cuda_glwe_sample_extract_128_async(
     void const *glwe_array_in, uint32_t const *nth_array, uint32_t num_nths,
     uint32_t num_lwes_to_extract_per_glwe, uint32_t num_lwes_stored_per_glwe,
     uint32_t glwe_dimension, uint32_t polynomial_size);
+
+void cuda_modulus_switch_multi_bit_64_async(void *stream, uint32_t gpu_index,
+                                            void *lwe_array_out,
+                                            void *lwe_array_in, uint32_t size,
+                                            uint32_t log_modulus,
+                                            uint32_t degree,
+                                            uint32_t grouping_factor);
+
+void cuda_modulus_switch_multi_bit_128_async(void *stream, uint32_t gpu_index,
+                                             void *lwe_array_out,
+                                             void *lwe_array_in, uint32_t size,
+                                             uint32_t log_modulus,
+                                             uint32_t degree,
+                                             uint32_t grouping_factor);
 }
 #endif

backends/tfhe-cuda-backend/cuda/include/integer/div_rem.h

@@ -382,14 +382,17 @@ template <typename Torus> struct unsigned_int_div_rem_2_2_memory {
                        ->use_sequential_algorithm_to_resolve_group_carries;
 
     cuda_set_device(0);
-    cudaEventCreateWithFlags(&create_indexes_done, cudaEventDisableTiming);
+    check_cuda_error(
+        cudaEventCreateWithFlags(&create_indexes_done, cudaEventDisableTiming));
     create_indexes_for_overflow_sub(streams.get_ith(0), num_blocks, group_size,
                                     use_seq, allocate_gpu_memory, size_tracker);
-    cudaEventRecord(create_indexes_done, streams.stream(0));
+    check_cuda_error(cudaEventRecord(create_indexes_done, streams.stream(0)));
     cuda_set_device(1);
-    cudaStreamWaitEvent(streams.stream(1), create_indexes_done, 0);
+    check_cuda_error(
+        cudaStreamWaitEvent(streams.stream(1), create_indexes_done, 0));
     cuda_set_device(2);
-    cudaStreamWaitEvent(streams.stream(2), create_indexes_done, 0);
+    check_cuda_error(
+        cudaStreamWaitEvent(streams.stream(2), create_indexes_done, 0));
 
     scatter_indexes_for_overflowing_sub(
         streams.stream(1), streams.gpu_index(1),
@@ -842,7 +845,7 @@ template <typename Torus> struct unsigned_int_div_rem_2_2_memory {
     free(second_indexes_for_overflow_sub_gpu_2);
     free(scalars_for_overflow_sub_gpu_2);
 
-    cudaEventDestroy(create_indexes_done);
+    check_cuda_error(cudaEventDestroy(create_indexes_done));
 
     // release sub streams
     sub_streams_1.release();

backends/tfhe-cuda-backend/cuda/include/integer/integer.h

@@ -153,7 +153,13 @@ void cuda_full_propagation_64_inplace_async(
 void cleanup_cuda_full_propagation_64_inplace(CudaStreamsFFI streams,
                                               int8_t **mem_ptr_void);
 
-uint64_t scratch_cuda_integer_mult_64_async(
+void cuda_integer_mult_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *radix_lwe_inout,
+    bool const is_bool_left, CudaRadixCiphertextFFI const *radix_lwe_right,
+    bool const is_bool_right, void *const *bsks, void *const *ksks,
+    int8_t *mem_ptr, uint32_t polynomial_size, uint32_t num_blocks);
+
+uint64_t scratch_cuda_integer_mult_inplace_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, bool const is_boolean_left,
     bool const is_boolean_right, uint32_t message_modulus,
     uint32_t carry_modulus, uint32_t glwe_dimension, uint32_t lwe_dimension,
@@ -162,17 +168,8 @@ uint64_t scratch_cuda_integer_mult_64_async(
     uint32_t num_blocks, PBS_TYPE pbs_type, bool allocate_gpu_memory,
     PBS_MS_REDUCTION_T noise_reduction_type);
 
-void cuda_integer_mult_64_async(CudaStreamsFFI streams,
-                                CudaRadixCiphertextFFI *radix_lwe_out,
-                                CudaRadixCiphertextFFI const *radix_lwe_left,
-                                bool const is_bool_left,
-                                CudaRadixCiphertextFFI const *radix_lwe_right,
-                                bool const is_bool_right, void *const *bsks,
-                                void *const *ksks, int8_t *mem_ptr,
-                                uint32_t polynomial_size, uint32_t num_blocks);
-
-void cleanup_cuda_integer_mult_64(CudaStreamsFFI streams,
-                                  int8_t **mem_ptr_void);
+void cleanup_cuda_integer_mult_inplace_64(CudaStreamsFFI streams,
+                                          int8_t **mem_ptr_void);
 
 void cuda_negate_ciphertext_64(CudaStreamsFFI streams,
                                CudaRadixCiphertextFFI *lwe_array_out,
@@ -273,7 +270,12 @@ void cleanup_cuda_integer_comparison_64(CudaStreamsFFI streams,
 void cleanup_cuda_integer_scalar_comparison_64(CudaStreamsFFI streams,
                                                int8_t **mem_ptr_void);
 
-uint64_t scratch_cuda_boolean_bitop_64_async(
+void cuda_boolean_bitop_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *lwe_array_2, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks);
+
+uint64_t scratch_cuda_boolean_bitop_inplace_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t big_lwe_dimension,
     uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
@@ -283,15 +285,8 @@ uint64_t scratch_cuda_boolean_bitop_64_async(
     bool is_unchecked, bool allocate_gpu_memory,
     PBS_MS_REDUCTION_T noise_reduction_type);
 
-void cuda_boolean_bitop_64_async(CudaStreamsFFI streams,
-                                 CudaRadixCiphertextFFI *lwe_array_out,
-                                 CudaRadixCiphertextFFI const *lwe_array_1,
-                                 CudaRadixCiphertextFFI const *lwe_array_2,
-                                 int8_t *mem_ptr, void *const *bsks,
-                                 void *const *ksks);
-
-void cleanup_cuda_boolean_bitop_64(CudaStreamsFFI streams,
-                                   int8_t **mem_ptr_void);
+void cleanup_cuda_boolean_bitop_inplace_64(CudaStreamsFFI streams,
+                                           int8_t **mem_ptr_void);
 
 uint64_t scratch_cuda_boolean_bitnot_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
@@ -316,42 +311,40 @@ void cuda_bitnot_ciphertext_64(CudaStreamsFFI streams,
                                uint32_t param_message_modulus,
                                uint32_t param_carry_modulus);
 
-uint64_t scratch_cuda_integer_bitop_64_async(
-    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
-    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
-
-uint64_t scratch_cuda_integer_scalar_bitop_64_async(
-    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
-    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
-
-void cuda_integer_scalar_bitop_64_async(
-    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_out,
-    CudaRadixCiphertextFFI const *lwe_array_input, void const *clear_blocks,
-    void const *h_clear_blocks, uint32_t num_clear_blocks, int8_t *mem_ptr,
+void cuda_integer_bitop_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *lwe_array_2, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks);
 
-void cuda_integer_bitop_64_async(CudaStreamsFFI streams,
-                                 CudaRadixCiphertextFFI *lwe_array_out,
-                                 CudaRadixCiphertextFFI const *lwe_array_1,
-                                 CudaRadixCiphertextFFI const *lwe_array_2,
-                                 int8_t *mem_ptr, void *const *bsks,
-                                 void *const *ksks);
+void cuda_integer_scalar_bitop_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_inout,
+    void const *clear_blocks, void const *h_clear_blocks,
+    uint32_t num_clear_blocks, int8_t *mem_ptr, void *const *bsks,
+    void *const *ksks);
 
-void cleanup_cuda_integer_bitop_64(CudaStreamsFFI streams,
-                                   int8_t **mem_ptr_void);
+uint64_t scratch_cuda_integer_bitop_inplace_64_async(
+    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t big_lwe_dimension,
+    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
+    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
 
-void cleanup_cuda_integer_scalar_bitop_64(CudaStreamsFFI streams,
-                                          int8_t **mem_ptr_void);
+void cleanup_cuda_integer_bitop_inplace_64(CudaStreamsFFI streams,
+                                           int8_t **mem_ptr_void);
+
+uint64_t scratch_cuda_integer_scalar_bitop_inplace_64_async(
+    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t big_lwe_dimension,
+    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
+    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type);
+
+void cleanup_cuda_integer_scalar_bitop_inplace_64(CudaStreamsFFI streams,
+                                                  int8_t **mem_ptr_void);
 
 uint64_t scratch_cuda_cmux_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,

backends/tfhe-cuda-backend/cuda/include/integer/rerand.h

@@ -2,12 +2,17 @@
 
 #include "integer.h"
 
+enum RERAND_MODE {
+  RERAND_WITH_KS = 0,
+  RERAND_WITHOUT_KS = 1,
+};
+
 extern "C" {
 uint64_t scratch_cuda_rerand_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t big_lwe_dimension,
     uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
     uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, bool allocate_gpu_memory);
+    uint32_t carry_modulus, bool allocate_gpu_memory, RERAND_MODE rerand_type);
 
 void cuda_rerand_64_async(
     CudaStreamsFFI streams, void *lwe_array,

backends/tfhe-cuda-backend/cuda/include/integer/rerand_utilities.h

@@ -3,16 +3,18 @@
 #include "checked_arithmetic.h"
 #include "integer_utilities.h"
 #include "keyswitch/ks_enums.h"
+#include "rerand.h"
 #include "zk/expand.cuh"
 #include "zk/zk_utilities.h"
 
 template <typename Torus> struct int_rerand_mem {
   int_radix_params params;
 
-  Torus *tmp_zero_lwes;
-  Torus *tmp_ksed_zero_lwes;
-  Torus *lwe_trivial_indexes;
+  Torus *tmp_expanded_zero_lwes = nullptr;
+  Torus *tmp_ksed_expanded_zero_lwes = nullptr;
+  Torus *lwe_trivial_indexes = nullptr;
   uint32_t num_lwes;
+  RERAND_MODE rerand_mode;
 
   bool gpu_memory_allocated;
 
@@ -20,24 +22,20 @@ template <typename Torus> struct int_rerand_mem {
       ks_tmp_buf_vec; // not allocated, ReRand not using GEMM KS for now
   // kept empty to pass to the KS function indicating GEMM KS disabled
 
-  expand_job<Torus> *d_expand_jobs;
-  expand_job<Torus> *h_expand_jobs;
+  expand_job<Torus> *d_expand_jobs = nullptr;
+  expand_job<Torus> *h_expand_jobs = nullptr;
 
   int_rerand_mem(CudaStreams streams, int_radix_params params,
-                 const uint32_t num_lwes, const bool allocate_gpu_memory,
-                 uint64_t &size_tracker)
-      : params(params), num_lwes(num_lwes),
+                 const uint32_t num_lwes, const RERAND_MODE rerand_mode,
+                 const bool allocate_gpu_memory, uint64_t &size_tracker)
+      : params(params), num_lwes(num_lwes), rerand_mode(rerand_mode),
         gpu_memory_allocated(allocate_gpu_memory) {
 
-    tmp_zero_lwes = (Torus *)cuda_malloc_with_size_tracking_async(
-        safe_mul_sizeof<Torus>(num_lwes, params.big_lwe_dimension + 1),
-        streams.stream(0), streams.gpu_index(0), size_tracker,
-        allocate_gpu_memory);
-
-    tmp_ksed_zero_lwes = (Torus *)cuda_malloc_with_size_tracking_async(
-        safe_mul_sizeof<Torus>(num_lwes, params.small_lwe_dimension + 1),
-        streams.stream(0), streams.gpu_index(0), size_tracker,
-        allocate_gpu_memory);
+    tmp_expanded_zero_lwes =
+        static_cast<Torus *>(cuda_malloc_with_size_tracking_async(
+            safe_mul_sizeof<Torus>(num_lwes, params.big_lwe_dimension + 1),
+            streams.stream(0), streams.gpu_index(0), size_tracker,
+            allocate_gpu_memory));
 
     d_expand_jobs =
         static_cast<expand_job<Torus> *>(cuda_malloc_with_size_tracking_async(
@@ -46,47 +44,63 @@ template <typename Torus> struct int_rerand_mem {
 
     h_expand_jobs = static_cast<expand_job<Torus> *>(
         malloc(safe_mul_sizeof<expand_job<Torus>>(num_lwes)));
+    PANIC_IF_FALSE(h_expand_jobs != nullptr,
+                   "host allocation failed for h_expand_jobs");
 
-    auto h_lwe_trivial_indexes =
-        static_cast<Torus *>(malloc(safe_mul_sizeof<Torus>(num_lwes)));
-    for (auto i = 0; i < num_lwes; ++i) {
-      h_lwe_trivial_indexes[i] = i;
+    if (rerand_mode == RERAND_MODE::RERAND_WITH_KS) {
+      tmp_ksed_expanded_zero_lwes =
+          static_cast<Torus *>(cuda_malloc_with_size_tracking_async(
+              safe_mul_sizeof<Torus>(num_lwes, params.small_lwe_dimension + 1),
+              streams.stream(0), streams.gpu_index(0), size_tracker,
+              allocate_gpu_memory));
+
+      auto h_lwe_trivial_indexes =
+          static_cast<Torus *>(malloc(safe_mul_sizeof<Torus>(num_lwes)));
+      PANIC_IF_FALSE(h_lwe_trivial_indexes != nullptr,
+                     "host allocation failed for h_lwe_trivial_indexes");
+      for (uint32_t i = 0; i < num_lwes; ++i) {
+        h_lwe_trivial_indexes[i] = i;
+      }
+      lwe_trivial_indexes =
+          static_cast<Torus *>(cuda_malloc_with_size_tracking_async(
+              safe_mul_sizeof<Torus>(num_lwes), streams.stream(0),
+              streams.gpu_index(0), size_tracker, allocate_gpu_memory));
+      cuda_memcpy_async_to_gpu(lwe_trivial_indexes, h_lwe_trivial_indexes,
+                               safe_mul_sizeof<Torus>(num_lwes),
+                               streams.stream(0), streams.gpu_index(0));
+      cuda_synchronize_stream(streams.stream(0), streams.gpu_index(0));
+      free(h_lwe_trivial_indexes);
+    } else {
+      cuda_synchronize_stream(streams.stream(0), streams.gpu_index(0));
     }
-    lwe_trivial_indexes = (Torus *)cuda_malloc_with_size_tracking_async(
-        safe_mul_sizeof<Torus>(num_lwes), streams.stream(0),
-        streams.gpu_index(0), size_tracker, allocate_gpu_memory);
-    cuda_memcpy_async_to_gpu(lwe_trivial_indexes, h_lwe_trivial_indexes,
-                             safe_mul_sizeof<Torus>(num_lwes),
-                             streams.stream(0), streams.gpu_index(0));
-
-    cuda_synchronize_stream(streams.stream(0), streams.gpu_index(0));
-
-    free(h_lwe_trivial_indexes);
   }
 
   void release(CudaStreams streams) {
-    cuda_drop_with_size_tracking_async(tmp_zero_lwes, streams.stream(0),
-                                       streams.gpu_index(0),
+    cuda_drop_with_size_tracking_async(tmp_expanded_zero_lwes,
+                                       streams.stream(0), streams.gpu_index(0),
                                        gpu_memory_allocated);
-    tmp_zero_lwes = nullptr;
-    cuda_drop_with_size_tracking_async(tmp_ksed_zero_lwes, streams.stream(0),
-                                       streams.gpu_index(0),
-                                       gpu_memory_allocated);
-    tmp_ksed_zero_lwes = nullptr;
-    cuda_drop_with_size_tracking_async(lwe_trivial_indexes, streams.stream(0),
-                                       streams.gpu_index(0),
-                                       gpu_memory_allocated);
-    lwe_trivial_indexes = nullptr;
+    tmp_expanded_zero_lwes = nullptr;
     cuda_drop_with_size_tracking_async(d_expand_jobs, streams.stream(0),
                                        streams.gpu_index(0),
                                        gpu_memory_allocated);
     d_expand_jobs = nullptr;
 
-    for (auto i = 0; i < ks_tmp_buf_vec.size(); i++) {
-      cleanup_cuda_keyswitch(streams.stream(i), streams.gpu_index(i),
-                             ks_tmp_buf_vec[i], gpu_memory_allocated);
+    if (rerand_mode == RERAND_MODE::RERAND_WITH_KS) {
+      cuda_drop_with_size_tracking_async(
+          tmp_ksed_expanded_zero_lwes, streams.stream(0), streams.gpu_index(0),
+          gpu_memory_allocated);
+      tmp_ksed_expanded_zero_lwes = nullptr;
+      cuda_drop_with_size_tracking_async(lwe_trivial_indexes, streams.stream(0),
+                                         streams.gpu_index(0),
+                                         gpu_memory_allocated);
+      lwe_trivial_indexes = nullptr;
+
+      for (size_t i = 0; i < ks_tmp_buf_vec.size(); i++) {
+        cleanup_cuda_keyswitch(streams.stream(i), streams.gpu_index(i),
+                               ks_tmp_buf_vec[i], gpu_memory_allocated);
+      }
+      ks_tmp_buf_vec.clear();
     }
-    ks_tmp_buf_vec.clear();
 
     cuda_synchronize_stream(streams.stream(0), streams.gpu_index(0));
     free(h_expand_jobs);

backends/tfhe-cuda-backend/cuda/include/linear_algebra.h

@@ -64,6 +64,12 @@ void cuda_add_lwe_ciphertext_vector_plaintext_64(
     void const *lwe_array_in, const uint64_t plaintext_in,
     const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count);
+void cuda_add_lwe_ciphertext_vector_inplace_32(
+    void *stream, uint32_t gpu_index, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *input_2);
+void cuda_add_lwe_ciphertext_vector_inplace_64(
+    void *stream, uint32_t gpu_index, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *input_2);
 }
 
 #endif // CUDA_LINALG_H_

backends/tfhe-cuda-backend/cuda/include/pbs/programmable_bootstrap_multibit.h

@@ -39,6 +39,28 @@ void cleanup_cuda_multi_bit_programmable_bootstrap_64(void *stream,
                                                       uint32_t gpu_index,
                                                       int8_t **pbs_buffer);
 
+// Noise-tests-namespaced wrappers for scratch/cleanup, so that callers
+// working with the noise-tests PBS variant use a consistent naming scheme.
+uint64_t scratch_cuda_multi_bit_programmable_bootstrap_noise_tests_64_async(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
+
+void cleanup_cuda_multi_bit_programmable_bootstrap_noise_tests_64(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer);
+
+// Noise tests variant: 64-bit torus, polynomial_size=2048 only. Uses the
+// NOISE_TESTS keybundle mode for noise analysis purposes.
+void cuda_multi_bit_programmable_bootstrap_noise_tests_64_async(
+    void *stream, uint32_t gpu_index, void *lwe_array_out,
+    void const *lwe_output_indexes, void const *lut_vector,
+    void const *lut_vector_indexes, void const *lwe_array_in,
+    void const *lwe_input_indexes, void const *bootstrapping_key,
+    int8_t *buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t grouping_factor, uint32_t base_log,
+    uint32_t level_count, uint32_t num_samples, uint32_t num_many_lut,
+    uint32_t lut_stride);
+
 uint64_t scratch_cuda_multi_bit_programmable_bootstrap_128_async(
     void *stream, uint32_t gpu_index, int8_t **buffer, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t level_count,
@@ -56,6 +78,23 @@ void cuda_multi_bit_programmable_bootstrap_128_async(
 void cleanup_cuda_multi_bit_programmable_bootstrap_128(void *stream,
                                                        const uint32_t gpu_index,
                                                        int8_t **buffer);
+
+uint64_t scratch_cuda_multi_bit_programmable_bootstrap_noise_tests_128_async(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
+
+void cleanup_cuda_multi_bit_programmable_bootstrap_noise_tests_128(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer);
+
+void cuda_multi_bit_programmable_bootstrap_noise_tests_128_async(
+    void *stream, uint32_t gpu_index, void *lwe_array_out,
+    void const *lwe_output_indexes, void const *lut_vector,
+    void const *lwe_array_in, void const *lwe_input_indexes,
+    void const *bootstrapping_key, int8_t *buffer, uint32_t lwe_dimension,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride);
 }
 
 #endif // CUDA_MULTI_BIT_H

backends/tfhe-cuda-backend/cuda/include/zk/zk_utilities.h

@@ -105,11 +105,11 @@ template <typename Torus> struct zk_expand_mem {
   uint32_t num_lwes;
   uint32_t num_compact_lists;
 
-  int_radix_lut<Torus> *message_and_carry_extract_luts;
-  int_radix_lut<Torus> *identity_lut;
+  int_radix_lut<Torus> *message_and_carry_extract_luts = nullptr;
+  int_radix_lut<Torus> *identity_lut = nullptr;
 
-  Torus *tmp_expanded_lwes;
-  Torus *tmp_ksed_small_to_big_expanded_lwes;
+  Torus *tmp_expanded_lwes = nullptr;
+  Torus *tmp_ksed_small_to_big_expanded_lwes = nullptr;
 
   bool gpu_memory_allocated;
 
@@ -148,66 +148,6 @@ template <typename Torus> struct zk_expand_mem {
       PANIC("GPU backend requires carry_modulus equal to message_modulus")
     }
 
-    // We create the identity LUT only if we are doing a SANITY_CHECK
-    if (expand_kind == EXPAND_KIND::SANITY_CHECK) {
-      identity_lut =
-          new int_radix_lut<Torus>(streams, computing_params, 1, 2 * num_lwes,
-                                   allocate_gpu_memory, size_tracker);
-
-      auto identity_lut_f = [](Torus x) -> Torus { return x; };
-
-      identity_lut->generate_and_broadcast_lut(streams, {0}, {identity_lut_f},
-                                               LUT_0_FOR_ALL_BLOCKS);
-    }
-
-    auto message_extract_lut_f = [casting_params](Torus x) -> Torus {
-      return x % casting_params.message_modulus;
-    };
-    auto carry_extract_lut_f = [casting_params](Torus x) -> Torus {
-      return (x / casting_params.carry_modulus) %
-             casting_params.message_modulus;
-    };
-
-    // Booleans have to be sanitized
-    auto sanitize_bool_f = [](Torus x) -> Torus { return x == 0 ? 0 : 1; };
-    auto message_extract_and_sanitize_bool_lut_f =
-        [message_extract_lut_f, sanitize_bool_f](Torus x) -> Torus {
-      return sanitize_bool_f(message_extract_lut_f(x));
-    };
-    auto carry_extract_and_sanitize_bool_lut_f =
-        [carry_extract_lut_f, sanitize_bool_f](Torus x) -> Torus {
-      return sanitize_bool_f(carry_extract_lut_f(x));
-    };
-
-    /** In case the casting key casts from BIG to SMALL key we run a single KS
-    to expand using the casting key as ksk. Otherwise, in case the casting key
-    casts from SMALL to BIG key, we first keyswitch from SMALL to BIG using
-    the casting key as ksk, then we keyswitch from BIG to SMALL using the
-    computing ksk, and lastly we apply the PBS. The output is always on the
-    BIG key.
-    **/
-    auto params = casting_params;
-    if (casting_key_type == SMALL_TO_BIG) {
-      params = computing_params;
-    }
-    message_and_carry_extract_luts = new int_radix_lut<Torus>(
-        streams, params, 4, 2 * num_lwes, allocate_gpu_memory, size_tracker);
-
-    // We are always packing two LWEs. We just need to be sure we have enough
-    // space in the carry part to store a message of the same size as is in the
-    // message part.
-    if (params.carry_modulus < params.message_modulus)
-      PANIC("Carry modulus must be at least as large as message modulus");
-    auto num_packed_msgs = 2;
-
-    // Adjust indexes to permute the output and access the correct LUT
-    auto h_indexes_in = static_cast<Torus *>(
-        malloc(safe_mul_sizeof<Torus>(num_packed_msgs, num_lwes)));
-    auto h_indexes_out = static_cast<Torus *>(
-        malloc(safe_mul_sizeof<Torus>(num_packed_msgs, num_lwes)));
-    auto h_lut_indexes = static_cast<Torus *>(
-        malloc(safe_mul_sizeof<Torus>(num_packed_msgs, num_lwes)));
-
     d_expand_jobs =
         static_cast<expand_job<Torus> *>(cuda_malloc_with_size_tracking_async(
             safe_mul_sizeof<expand_job<Torus>>(num_lwes), streams.stream(0),
@@ -216,144 +156,202 @@ template <typename Torus> struct zk_expand_mem {
     h_expand_jobs = static_cast<expand_job<Torus> *>(
         malloc(safe_mul_sizeof<expand_job<Torus>>(num_lwes)));
 
-    /*
-     * Each LWE contains encrypted data in both carry and message spaces
-     * that needs to be extracted.
-     *
-     * The loop processes each compact list (k) and for each LWE within that
-     * list:
-     * 1. Sets input indexes to read each LWE twice (for carry and message
-     * extraction)
-     * 2. Creates output indexes to properly reorder the results
-     * 3. Selects appropriate LUT index based on whether boolean sanitization is
-     * needed
-     *
-     * We want the output to have always first the content of the message part
-     * and then the content of the carry part of each LWE.
-     *
-     * i.e. msg_extract(LWE_0), carry_extract(LWE_0), msg_extract(LWE_1),
-     * carry_extract(LWE_1), ...
-     *
-     * Aiming that behavior, with 4 LWEs we would have:
-     *
-     * // Each LWE is processed twice
-     * h_indexes_in   = {0, 1, 2, 3, 0, 1, 2, 3}
-     *
-     * // First 4 use message LUT, last 4 use carry LUT
-     * h_lut_indexes  = {0, 0, 0, 0, 1, 1, 1, 1}
-     *
-     * // Reorders output so message and carry for each LWE appear together
-     * h_indexes_out  = {0, 2, 4, 6, 1, 3, 5, 7}
-     *
-     * If an LWE contains a boolean value, its LUT index is shifted by
-     * num_packed_msgs to use the sanitization LUT (which ensures output is
-     * exactly 0 or 1).
-     */
-    auto offset = 0;
-    for (int k = 0; k < num_compact_lists; k++) {
-      auto num_lwes_in_kth = this->num_lwes_per_compact_list[k];
-      for (int i = 0; i < num_packed_msgs * num_lwes_in_kth; i++) {
-        auto lwe_index = i + num_packed_msgs * offset;
-        auto lwe_index_in_list = i % num_lwes_in_kth;
-        PANIC_IF_FALSE(lwe_index < num_packed_msgs * num_lwes,
-                       "Cuda error: index %d is beyond the max value %d",
-                       lwe_index, num_packed_msgs * num_lwes);
-        h_indexes_in[lwe_index] = lwe_index_in_list + offset;
-        h_indexes_out[lwe_index] =
-            num_packed_msgs * h_indexes_in[lwe_index] + i / num_lwes_in_kth;
-        PANIC_IF_FALSE(h_indexes_in[lwe_index] < num_packed_msgs * num_lwes,
-                       "Cuda error: index %lu is beyond the max value %lu",
-                       (unsigned long)h_indexes_in[lwe_index],
-                       (unsigned long)(num_packed_msgs * num_lwes));
-        PANIC_IF_FALSE(h_indexes_out[lwe_index] < num_packed_msgs * num_lwes,
-                       "Cuda error: index %lu is beyond the max value %lu",
-                       (unsigned long)h_indexes_out[lwe_index],
-                       (unsigned long)(num_packed_msgs * num_lwes));
-        // is_boolean_array tells us which input is a boolean and thus the
-        // related output needs boolean sanitization. It naturally has
-        // total_blocks entries, but h_indexes_out reaches
-        // message_modulus * ceil(total_blocks/2) - 1. When total_blocks is odd,
-        // the ceiling causes out-of-bounds access. Reading garbage "true" would
-        // set h_lut_indexes to an invalid index pointing to uninitialized
-        // memory instead of a real LUT. Rust pads is_boolean_array with FALSE
-        // to match.
-        PANIC_IF_FALSE(h_indexes_out[lwe_index] < is_boolean_array_len,
-                       "Cuda error: index %lu for is_boolean_array is out of "
-                       "bounds (len is %lu)",
-                       (unsigned long)h_indexes_out[lwe_index],
-                       (unsigned long)is_boolean_array_len);
+    // NO_CASTING expands directly into the output buffer — no LUTs, no PBS,
+    // no intermediate buffers needed.
+    if (expand_kind != EXPAND_KIND::NO_CASTING) {
+      /** In case the casting key casts from BIG to SMALL key we run a single KS
+      to expand using the casting key as ksk. Otherwise, in case the casting key
+      casts from SMALL to BIG key, we first keyswitch from SMALL to BIG using
+      the casting key as ksk, then we keyswitch from BIG to SMALL using the
+      computing ksk, and lastly we apply the PBS. The output is always on the
+      BIG key.
+      **/
+      auto params = casting_params;
+      if (casting_key_type == SMALL_TO_BIG) {
+        params = computing_params;
       }
-      offset += num_lwes_in_kth;
-    }
 
-    message_and_carry_extract_luts->set_lwe_indexes(
-        streams.stream(0), streams.gpu_index(0), h_indexes_in, h_indexes_out);
+      // We always pack two LWEs (message and carry parts per LWE)
+      auto num_packed_msgs = 2;
 
-    auto active_streams =
-        streams.active_gpu_subset(2 * num_lwes, params.pbs_type);
+      // Adjust indexes to permute the output and access the correct LUT.
+      //
+      // The loop below fills h_indexes_in and h_indexes_out so that the output
+      // is ordered as: msg_extract(LWE_0), carry_extract(LWE_0),
+      // msg_extract(LWE_1), carry_extract(LWE_1), ...
+      //
+      // With 4 LWEs the arrays look like:
+      //   h_indexes_in  = {0, 1, 2, 3, 0, 1, 2, 3}  (each LWE read twice)
+      //   h_lut_indexes = {0, 0, 0, 0, 1, 1, 1, 1}  (msg LUT then carry LUT)
+      //   h_indexes_out = {0, 2, 4, 6, 1, 3, 5, 7}  (interleaved output)
+      //
+      // If an LWE contains a boolean its LUT index is shifted by
+      // num_packed_msgs to use the sanitization LUT (output clamped to {0, 1}).
+      auto h_indexes_in = static_cast<Torus *>(
+          malloc(safe_mul_sizeof<Torus>(num_packed_msgs, num_lwes)));
+      auto h_indexes_out = static_cast<Torus *>(
+          malloc(safe_mul_sizeof<Torus>(num_packed_msgs, num_lwes)));
 
-    // Index generator for message/carry extraction LUTs
-    auto index_gen = [num_compact_lists,
-                      num_lwes_per_compact_list =
-                          this->num_lwes_per_compact_list,
-                      num_packed_msgs, is_boolean_array,
-                      h_indexes_out](Torus *h_lut_indexes, uint32_t) {
       auto offset = 0;
       for (int k = 0; k < num_compact_lists; k++) {
-        auto num_lwes_in_kth = num_lwes_per_compact_list[k];
+        auto num_lwes_in_kth = this->num_lwes_per_compact_list[k];
         for (int i = 0; i < num_packed_msgs * num_lwes_in_kth; i++) {
           auto lwe_index = i + num_packed_msgs * offset;
-          auto boolean_offset =
-              is_boolean_array[h_indexes_out[lwe_index]] ? num_packed_msgs : 0;
-          h_lut_indexes[lwe_index] = i / num_lwes_in_kth + boolean_offset;
+          auto lwe_index_in_list = i % num_lwes_in_kth;
+          PANIC_IF_FALSE(lwe_index < num_packed_msgs * num_lwes,
+                         "Cuda error: index %d is beyond the max value %d",
+                         lwe_index, num_packed_msgs * num_lwes);
+          h_indexes_in[lwe_index] = lwe_index_in_list + offset;
+          h_indexes_out[lwe_index] =
+              num_packed_msgs * h_indexes_in[lwe_index] + i / num_lwes_in_kth;
+          PANIC_IF_FALSE(h_indexes_in[lwe_index] < num_packed_msgs * num_lwes,
+                         "Cuda error: index %lu is beyond the max value %lu",
+                         (unsigned long)h_indexes_in[lwe_index],
+                         (unsigned long)(num_packed_msgs * num_lwes));
+          PANIC_IF_FALSE(h_indexes_out[lwe_index] < num_packed_msgs * num_lwes,
+                         "Cuda error: index %lu is beyond the max value %lu",
+                         (unsigned long)h_indexes_out[lwe_index],
+                         (unsigned long)(num_packed_msgs * num_lwes));
+          // is_boolean_array tells us which input is a boolean and thus the
+          // related output needs boolean sanitization. It naturally has
+          // total_blocks entries, but h_indexes_out reaches
+          // message_modulus * ceil(total_blocks/2) - 1. When total_blocks is
+          // odd, the ceiling causes out-of-bounds access. Reading garbage
+          // "true" would set h_lut_indexes to an invalid index pointing to
+          // uninitialized memory instead of a real LUT. Rust pads
+          // is_boolean_array with FALSE to match.
+          PANIC_IF_FALSE(h_indexes_out[lwe_index] < is_boolean_array_len,
+                         "Cuda error: index %lu for is_boolean_array is out of "
+                         "bounds (len is %lu)",
+                         (unsigned long)h_indexes_out[lwe_index],
+                         (unsigned long)is_boolean_array_len);
         }
         offset += num_lwes_in_kth;
       }
-    };
 
-    message_and_carry_extract_luts->generate_and_broadcast_lut(
-        active_streams, {0, 1, 2, 3},
-        {message_extract_lut_f, carry_extract_lut_f,
-         message_extract_and_sanitize_bool_lut_f,
-         carry_extract_and_sanitize_bool_lut_f},
-        index_gen, true, {}, h_lut_indexes);
+      auto active_streams =
+          streams.active_gpu_subset(2 * num_lwes, params.pbs_type);
 
-    message_and_carry_extract_luts->allocate_lwe_vector_for_non_trivial_indexes(
-        active_streams, 2 * num_lwes, size_tracker, allocate_gpu_memory);
-    // The expanded LWEs will always be on the casting key format
-    tmp_expanded_lwes = (Torus *)cuda_malloc_with_size_tracking_async(
-        safe_mul_sizeof<Torus>(num_lwes, casting_params.big_lwe_dimension + 1),
-        streams.stream(0), streams.gpu_index(0), size_tracker,
-        allocate_gpu_memory);
+      // SANITY_CHECK uses identity_lut (skipping the full message/carry
+      // extraction LUT and the SMALL_TO_BIG intermediate buffer).
+      if (expand_kind == EXPAND_KIND::SANITY_CHECK) {
+        identity_lut =
+            new int_radix_lut<Torus>(streams, casting_params, 1, 2 * num_lwes,
+                                     allocate_gpu_memory, size_tracker);
 
-    tmp_ksed_small_to_big_expanded_lwes =
-        (Torus *)cuda_malloc_with_size_tracking_async(
-            safe_mul_sizeof<Torus>(num_lwes,
-                                   casting_params.big_lwe_dimension + 1),
-            streams.stream(0), streams.gpu_index(0), size_tracker,
-            allocate_gpu_memory);
+        auto identity_lut_f = [](Torus x) -> Torus { return x; };
+        identity_lut->generate_and_broadcast_lut(streams, {0}, {identity_lut_f},
+                                                 LUT_0_FOR_ALL_BLOCKS);
+        identity_lut->set_lwe_indexes(streams.stream(0), streams.gpu_index(0),
+                                      h_indexes_in, h_indexes_out);
+        identity_lut->allocate_lwe_vector_for_non_trivial_indexes(
+            active_streams, 2 * num_lwes, size_tracker, allocate_gpu_memory);
+      } else {
+        // We are always packing two LWEs. We just need to be sure we have
+        // enough space in the carry part to store a message of the same size
+        // as is in the message part.
+        if (params.carry_modulus < params.message_modulus)
+          PANIC("Carry modulus must be at least as large as message modulus");
+
+        message_and_carry_extract_luts =
+            new int_radix_lut<Torus>(streams, params, 4, 2 * num_lwes,
+                                     allocate_gpu_memory, size_tracker);
+        message_and_carry_extract_luts->set_lwe_indexes(
+            streams.stream(0), streams.gpu_index(0), h_indexes_in,
+            h_indexes_out);
+
+        auto message_extract_lut_f = [casting_params](Torus x) -> Torus {
+          return x % casting_params.message_modulus;
+        };
+        auto carry_extract_lut_f = [casting_params](Torus x) -> Torus {
+          return (x / casting_params.carry_modulus) %
+                 casting_params.message_modulus;
+        };
+        auto sanitize_bool_f = [](Torus x) -> Torus { return x == 0 ? 0 : 1; };
+        auto message_extract_and_sanitize_bool_lut_f =
+            [message_extract_lut_f, sanitize_bool_f](Torus x) -> Torus {
+          return sanitize_bool_f(message_extract_lut_f(x));
+        };
+        auto carry_extract_and_sanitize_bool_lut_f =
+            [carry_extract_lut_f, sanitize_bool_f](Torus x) -> Torus {
+          return sanitize_bool_f(carry_extract_lut_f(x));
+        };
+
+        auto h_lut_indexes = static_cast<Torus *>(
+            malloc(safe_mul_sizeof<Torus>(num_packed_msgs, num_lwes)));
+
+        auto index_gen = [num_compact_lists,
+                          num_lwes_per_compact_list =
+                              this->num_lwes_per_compact_list,
+                          num_packed_msgs, is_boolean_array,
+                          h_indexes_out](Torus *h_lut_indexes, uint32_t) {
+          auto offset = 0;
+          for (int k = 0; k < num_compact_lists; k++) {
+            auto num_lwes_in_kth = num_lwes_per_compact_list[k];
+            for (int i = 0; i < num_packed_msgs * num_lwes_in_kth; i++) {
+              auto lwe_index = i + num_packed_msgs * offset;
+              auto boolean_offset = is_boolean_array[h_indexes_out[lwe_index]]
+                                        ? num_packed_msgs
+                                        : 0;
+              h_lut_indexes[lwe_index] = i / num_lwes_in_kth + boolean_offset;
+            }
+            offset += num_lwes_in_kth;
+          }
+        };
+
+        message_and_carry_extract_luts->generate_and_broadcast_lut(
+            active_streams, {0, 1, 2, 3},
+            {message_extract_lut_f, carry_extract_lut_f,
+             message_extract_and_sanitize_bool_lut_f,
+             carry_extract_and_sanitize_bool_lut_f},
+            index_gen, true, {}, h_lut_indexes);
+        message_and_carry_extract_luts
+            ->allocate_lwe_vector_for_non_trivial_indexes(
+                active_streams, 2 * num_lwes, size_tracker,
+                allocate_gpu_memory);
+        free(h_lut_indexes);
+
+        // SANITY_CHECK panics on SMALL_TO_BIG, so this buffer is only needed
+        // on the full casting path.
+        tmp_ksed_small_to_big_expanded_lwes =
+            (Torus *)cuda_malloc_with_size_tracking_async(
+                safe_mul_sizeof<Torus>(num_lwes,
+                                       casting_params.big_lwe_dimension + 1),
+                streams.stream(0), streams.gpu_index(0), size_tracker,
+                allocate_gpu_memory);
+      }
+
+      // The expanded LWEs will always be on the casting key format
+      tmp_expanded_lwes = (Torus *)cuda_malloc_with_size_tracking_async(
+          safe_mul_sizeof<Torus>(num_lwes,
+                                 casting_params.big_lwe_dimension + 1),
+          streams.stream(0), streams.gpu_index(0), size_tracker,
+          allocate_gpu_memory);
+
+      free(h_indexes_in);
+      free(h_indexes_out);
+    }
 
     cuda_synchronize_stream(streams.stream(0), streams.gpu_index(0));
-    free(h_indexes_in);
-    free(h_indexes_out);
-    free(h_lut_indexes);
   }
 
   void release(CudaStreams streams) {
-    message_and_carry_extract_luts->release(streams);
-    delete message_and_carry_extract_luts;
-
-    if (expand_kind == EXPAND_KIND::SANITY_CHECK) {
-      identity_lut->release(streams);
-      delete identity_lut;
+    if (expand_kind != EXPAND_KIND::NO_CASTING) {
+      if (expand_kind == EXPAND_KIND::SANITY_CHECK) {
+        identity_lut->release(streams);
+        delete identity_lut;
+      } else {
+        message_and_carry_extract_luts->release(streams);
+        delete message_and_carry_extract_luts;
+        cuda_drop_with_size_tracking_async(
+            tmp_ksed_small_to_big_expanded_lwes, streams.stream(0),
+            streams.gpu_index(0), gpu_memory_allocated);
+      }
+      cuda_drop_with_size_tracking_async(tmp_expanded_lwes, streams.stream(0),
+                                         streams.gpu_index(0),
+                                         gpu_memory_allocated);
     }
 
-    cuda_drop_with_size_tracking_async(tmp_expanded_lwes, streams.stream(0),
-                                       streams.gpu_index(0),
-                                       gpu_memory_allocated);
-    cuda_drop_with_size_tracking_async(tmp_ksed_small_to_big_expanded_lwes,
-                                       streams.stream(0), streams.gpu_index(0),
-                                       gpu_memory_allocated);
     cuda_drop_with_size_tracking_async(d_expand_jobs, streams.stream(0),
                                        streams.gpu_index(0),
                                        gpu_memory_allocated);

backends/tfhe-cuda-backend/cuda/src/crypto/ciphertext.cu

@@ -83,6 +83,8 @@ void cuda_modulus_switch_inplace_64_async(void *stream, uint32_t gpu_index,
 void cuda_modulus_switch_64_async(void *stream, uint32_t gpu_index,
                                   void *lwe_out, const void *lwe_in,
                                   uint32_t size, uint32_t log_modulus) {
+  PANIC_IF_FALSE(lwe_out != lwe_in, "Output and input pointers must be "
+                                    "different for out-of-place operations");
   host_modulus_switch<uint64_t>(static_cast<cudaStream_t>(stream), gpu_index,
                                 static_cast<uint64_t *>(lwe_out),
                                 static_cast<const uint64_t *>(lwe_in), size,
@@ -93,6 +95,8 @@ void cuda_centered_modulus_switch_64_async(void *stream, uint32_t gpu_index,
                                            void *lwe_out, const void *lwe_in,
                                            uint32_t lwe_dimension,
                                            uint32_t log_modulus) {
+  PANIC_IF_FALSE(lwe_out != lwe_in, "Output and input pointers must be "
+                                    "different for out-of-place operations");
   host_centered_modulus_switch_inplace<uint64_t>(
       static_cast<cudaStream_t>(stream), gpu_index,
       static_cast<uint64_t *>(lwe_out), static_cast<const uint64_t *>(lwe_in),
@@ -146,3 +150,31 @@ void cuda_glwe_sample_extract_128_async(
           "N's are powers of two in the interval [256..4096].")
   }
 }
+
+void cuda_modulus_switch_multi_bit_64_async(void *stream, uint32_t gpu_index,
+                                            void *lwe_array_out,
+                                            void *lwe_array_in, uint32_t size,
+                                            uint32_t log_modulus,
+                                            uint32_t degree,
+                                            uint32_t grouping_factor) {
+
+  host_modulus_switch_multi_bit<uint64_t>(
+      static_cast<cudaStream_t>(stream), gpu_index,
+      static_cast<uint64_t *>(lwe_array_out),
+      static_cast<uint64_t *>(lwe_array_in), size, log_modulus, degree,
+      grouping_factor);
+}
+
+void cuda_modulus_switch_multi_bit_128_async(void *stream, uint32_t gpu_index,
+                                             void *lwe_array_out,
+                                             void *lwe_array_in, uint32_t size,
+                                             uint32_t log_modulus,
+                                             uint32_t degree,
+                                             uint32_t grouping_factor) {
+
+  host_modulus_switch_multi_bit<__uint128_t>(
+      static_cast<cudaStream_t>(stream), gpu_index,
+      static_cast<__uint128_t *>(lwe_array_out),
+      static_cast<__uint128_t *>(lwe_array_in), size, log_modulus, degree,
+      grouping_factor);
+}

backends/tfhe-cuda-backend/cuda/src/crypto/keyswitch.cu

@@ -217,6 +217,8 @@ void cuda_closest_representable_64_async(void *stream, uint32_t gpu_index,
                                          void const *input, void *output,
                                          uint32_t base_log,
                                          uint32_t level_count) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
   host_cuda_closest_representable(static_cast<cudaStream_t>(stream), gpu_index,
                                   static_cast<const uint64_t *>(input),
                                   static_cast<uint64_t *>(output), base_log,

backends/tfhe-cuda-backend/cuda/src/crypto/torus.cuh

@@ -463,5 +463,48 @@ __global__ void __launch_bounds__(512)
       return;
   }
 }
+// This function is only used for noise tests, it follows the same logic
+// that is embedded in the keybundle just we need a global function to
+// be able to test it individually.
+template <typename Torus, class params>
+__global__ void
+modulus_switch_multi_bit(Torus *array_out, const Torus *array_in, int size,
+                         uint32_t log_modulus, uint32_t grouping_factor) {
+  const int tid = threadIdx.x + blockIdx.x * blockDim.x;
+  if (tid < size) {
+    int num_monomials = 1 << grouping_factor;
+    int input_offset = tid * grouping_factor;
+    int output_offset = tid * num_monomials;
+    // We calculate all monomials even if the first one is never used.
+    for (int ggsw_idx = 0; ggsw_idx < num_monomials; ggsw_idx++) {
+      array_out[ggsw_idx + output_offset] =
+          calculates_monomial_degree<Torus, params>(&array_in[input_offset],
+                                                    ggsw_idx, grouping_factor);
+    }
+  }
+}
+// This aims to be launched only from the noise tests.
+//  That is why we support a specific set of parameters
+template <typename Torus>
+__host__ void host_modulus_switch_multi_bit(
+    cudaStream_t stream, uint32_t gpu_index, Torus *array_out, Torus *array_in,
+    int size, uint32_t log_modulus, uint32_t degree, uint32_t grouping_factor) {
+  check_cuda_error(cudaSetDevice(gpu_index));
+  int multibit_size = size / grouping_factor;
+  int num_threads = 0, num_blocks = 0;
+  getNumBlocksAndThreads(multibit_size, 1024, num_blocks, num_threads);
+  switch (degree) {
+  case 2048:
+    modulus_switch_multi_bit<Torus, Degree<2048>>
+        <<<num_blocks, num_threads, 0, stream>>>(
+            array_out, array_in, multibit_size, log_modulus, grouping_factor);
+    break;
+  default:
+    PANIC("Cuda error: unsupported polynomial size. Supported "
+          "N's are powers of two in the interval [2048].")
+  };
+
+  check_cuda_error(cudaGetLastError());
+}
 
 #endif // CNCRT_TORUS_H

backends/tfhe-cuda-backend/cuda/src/device.cu

@@ -326,6 +326,10 @@ void cuda_memcpy_gpu_to_gpu(void *dest, void const *src, uint64_t size,
                             uint32_t gpu_index) {
   if (size == 0)
     return;
+
+  GPU_ASSERT(src != nullptr, "Cuda error: null device ptr");
+  GPU_ASSERT(dest != nullptr, "Cuda error: null device ptr");
+
   cudaPointerAttributes attr_dest;
   check_cuda_error(cudaPointerGetAttributes(&attr_dest, dest));
   PANIC_IF_FALSE(

backends/tfhe-cuda-backend/cuda/src/integer/bitwise_ops.cu

@@ -1,6 +1,16 @@
 #include "integer/bitwise_ops.cuh"
 
-uint64_t scratch_cuda_boolean_bitop_64_async(
+void cuda_boolean_bitop_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *lwe_array_2, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks) {
+  // In-place variant: lwe_array_inout op= lwe_array_2, no aliasing check needed
+  host_boolean_bitop<uint64_t>(
+      CudaStreams(streams), lwe_array_inout, lwe_array_inout, lwe_array_2,
+      (boolean_bitop_buffer<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks));
+}
+
+uint64_t scratch_cuda_boolean_bitop_inplace_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t big_lwe_dimension,
     uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
@@ -20,20 +30,8 @@ uint64_t scratch_cuda_boolean_bitop_64_async(
       lwe_ciphertext_count, params, op_type, is_unchecked, allocate_gpu_memory);
 }
 
-void cuda_boolean_bitop_64_async(CudaStreamsFFI streams,
-                                 CudaRadixCiphertextFFI *lwe_array_out,
-                                 CudaRadixCiphertextFFI const *lwe_array_1,
-                                 CudaRadixCiphertextFFI const *lwe_array_2,
-                                 int8_t *mem_ptr, void *const *bsks,
-                                 void *const *ksks) {
-
-  host_boolean_bitop<uint64_t>(
-      CudaStreams(streams), lwe_array_out, lwe_array_1, lwe_array_2,
-      (boolean_bitop_buffer<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks));
-}
-
-void cleanup_cuda_boolean_bitop_64(CudaStreamsFFI streams,
-                                   int8_t **mem_ptr_void) {
+void cleanup_cuda_boolean_bitop_inplace_64(CudaStreamsFFI streams,
+                                           int8_t **mem_ptr_void) {
 
   boolean_bitop_buffer<uint64_t> *mem_ptr =
       (boolean_bitop_buffer<uint64_t> *)(*mem_ptr_void);
@@ -80,44 +78,6 @@ void cleanup_cuda_boolean_bitnot_64(CudaStreamsFFI streams,
   *mem_ptr_void = nullptr;
 }
 
-uint64_t scratch_cuda_integer_bitop_64_async(
-    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
-    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type) {
-
-  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
-                          big_lwe_dimension, small_lwe_dimension, ks_level,
-                          ks_base_log, pbs_level, pbs_base_log, grouping_factor,
-                          message_modulus, carry_modulus, noise_reduction_type);
-
-  return scratch_cuda_bitop<uint64_t>(
-      CudaStreams(streams), (int_bitop_buffer<uint64_t> **)mem_ptr,
-      lwe_ciphertext_count, params, op_type, allocate_gpu_memory);
-}
-
-uint64_t scratch_cuda_integer_scalar_bitop_64_async(
-    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
-    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type) {
-
-  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
-                          big_lwe_dimension, small_lwe_dimension, ks_level,
-                          ks_base_log, pbs_level, pbs_base_log, grouping_factor,
-                          message_modulus, carry_modulus, noise_reduction_type);
-
-  return scratch_cuda_bitop<uint64_t>(
-      CudaStreams(streams), (int_bitop_buffer<uint64_t> **)mem_ptr,
-      lwe_ciphertext_count, params, op_type, allocate_gpu_memory);
-}
-
 void cuda_bitnot_ciphertext_64(CudaStreamsFFI streams,
                                CudaRadixCiphertextFFI *radix_ciphertext,
                                uint32_t ct_message_modulus,
@@ -129,20 +89,37 @@ void cuda_bitnot_ciphertext_64(CudaStreamsFFI streams,
   cuda_synchronize_stream(cuda_streams.stream(0), cuda_streams.gpu_index(0));
 }
 
-void cuda_integer_bitop_64_async(CudaStreamsFFI streams,
-                                 CudaRadixCiphertextFFI *lwe_array_out,
-                                 CudaRadixCiphertextFFI const *lwe_array_1,
-                                 CudaRadixCiphertextFFI const *lwe_array_2,
-                                 int8_t *mem_ptr, void *const *bsks,
-                                 void *const *ksks) {
-
-  host_bitop<uint64_t>(CudaStreams(streams), lwe_array_out, lwe_array_1,
+void cuda_integer_bitop_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *lwe_array_2, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks) {
+  // In-place variant: lwe_array_inout op= lwe_array_2, no aliasing check needed
+  host_bitop<uint64_t>(CudaStreams(streams), lwe_array_inout, lwe_array_inout,
                        lwe_array_2, (int_bitop_buffer<uint64_t> *)mem_ptr, bsks,
                        (uint64_t **)(ksks));
 }
 
-void cleanup_cuda_integer_bitop_64(CudaStreamsFFI streams,
-                                   int8_t **mem_ptr_void) {
+uint64_t scratch_cuda_integer_bitop_inplace_64_async(
+    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t big_lwe_dimension,
+    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
+    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type) {
+
+  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
+                          big_lwe_dimension, small_lwe_dimension, ks_level,
+                          ks_base_log, pbs_level, pbs_base_log, grouping_factor,
+                          message_modulus, carry_modulus, noise_reduction_type);
+
+  return scratch_cuda_bitop<uint64_t>(
+      CudaStreams(streams), (int_bitop_buffer<uint64_t> **)mem_ptr,
+      lwe_ciphertext_count, params, op_type, allocate_gpu_memory);
+}
+
+void cleanup_cuda_integer_bitop_inplace_64(CudaStreamsFFI streams,
+                                           int8_t **mem_ptr_void) {
 
   int_bitop_buffer<uint64_t> *mem_ptr =
       (int_bitop_buffer<uint64_t> *)(*mem_ptr_void);
@@ -151,8 +128,27 @@ void cleanup_cuda_integer_bitop_64(CudaStreamsFFI streams,
   *mem_ptr_void = nullptr;
 }
 
-void cleanup_cuda_integer_scalar_bitop_64(CudaStreamsFFI streams,
-                                          int8_t **mem_ptr_void) {
+uint64_t scratch_cuda_integer_scalar_bitop_inplace_64_async(
+    CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t big_lwe_dimension,
+    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
+    bool allocate_gpu_memory, PBS_MS_REDUCTION_T noise_reduction_type) {
+
+  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
+                          big_lwe_dimension, small_lwe_dimension, ks_level,
+                          ks_base_log, pbs_level, pbs_base_log, grouping_factor,
+                          message_modulus, carry_modulus, noise_reduction_type);
+
+  return scratch_cuda_bitop<uint64_t>(
+      CudaStreams(streams), (int_bitop_buffer<uint64_t> **)mem_ptr,
+      lwe_ciphertext_count, params, op_type, allocate_gpu_memory);
+}
+
+void cleanup_cuda_integer_scalar_bitop_inplace_64(CudaStreamsFFI streams,
+                                                  int8_t **mem_ptr_void) {
 
   int_bitop_buffer<uint64_t> *mem_ptr =
       (int_bitop_buffer<uint64_t> *)(*mem_ptr_void);

backends/tfhe-cuda-backend/cuda/src/integer/cast.cu

@@ -3,6 +3,8 @@
 void extend_radix_with_trivial_zero_blocks_msb_64(
     CudaRadixCiphertextFFI *output, CudaRadixCiphertextFFI const *input,
     CudaStreamsFFI streams) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
   auto cuda_streams = CudaStreams(streams);
   host_extend_radix_with_trivial_zero_blocks_msb<uint64_t>(output, input,
                                                            cuda_streams);
@@ -12,6 +14,8 @@ void extend_radix_with_trivial_zero_blocks_msb_64(
 void trim_radix_blocks_lsb_64(CudaRadixCiphertextFFI *output,
                               CudaRadixCiphertextFFI const *input,
                               CudaStreamsFFI streams) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
 
   auto cuda_streams = CudaStreams(streams);
   host_trim_radix_blocks_lsb<uint64_t>(output, input, cuda_streams);
@@ -21,6 +25,8 @@ void trim_radix_blocks_lsb_64(CudaRadixCiphertextFFI *output,
 void trim_radix_blocks_msb_64(CudaRadixCiphertextFFI *output,
                               CudaRadixCiphertextFFI const *input,
                               CudaStreamsFFI streams) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
 
   auto cuda_streams = CudaStreams(streams);
   host_trim_radix_blocks_msb<uint64_t>(output, input, cuda_streams);
@@ -54,6 +60,8 @@ void cuda_cast_to_unsigned_64_async(CudaStreamsFFI streams,
                                     int8_t *mem_ptr, uint32_t target_num_blocks,
                                     bool input_is_signed, void *const *bsks,
                                     void *const *ksks) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
 
   host_cast_to_unsigned<uint64_t>(
       CudaStreams(streams), output, input,
@@ -97,6 +105,8 @@ void cuda_cast_to_signed_64_async(CudaStreamsFFI streams,
                                   CudaRadixCiphertextFFI const *input,
                                   int8_t *mem, bool input_is_signed,
                                   void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
 
   host_cast_to_signed<uint64_t>(CudaStreams(streams), output, input,
                                 (int_cast_to_signed_buffer<uint64_t> *)mem,

backends/tfhe-cuda-backend/cuda/src/integer/cmux.cu

@@ -30,6 +30,18 @@ void cuda_cmux_64_async(CudaStreamsFFI streams,
                         CudaRadixCiphertextFFI const *lwe_array_true,
                         CudaRadixCiphertextFFI const *lwe_array_false,
                         int8_t *mem_ptr, void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(
+      lwe_array_out != lwe_condition,
+      "Output and condition pointers must be different for out-of-place "
+      "operations");
+  PANIC_IF_FALSE(
+      lwe_array_out != lwe_array_true,
+      "Output and true-branch pointers must be different for out-of-place "
+      "operations");
+  PANIC_IF_FALSE(
+      lwe_array_out != lwe_array_false,
+      "Output and false-branch pointers must be different for out-of-place "
+      "operations");
   PUSH_RANGE("cmux")
   host_cmux<uint64_t>(CudaStreams(streams), lwe_array_out, lwe_condition,
                       lwe_array_true, lwe_array_false,

backends/tfhe-cuda-backend/cuda/src/integer/comparison.cu

@@ -80,6 +80,12 @@ void cuda_integer_comparison_64_async(CudaStreamsFFI streams,
                                       CudaRadixCiphertextFFI const *lwe_array_2,
                                       int8_t *mem_ptr, void *const *bsks,
                                       void *const *ksks) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_1,
+                 "Output and first input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_2,
+                 "Output and second input pointers must be different for "
+                 "out-of-place operations");
   PUSH_RANGE("comparison")
   if (lwe_array_1->num_radix_blocks != lwe_array_2->num_radix_blocks)
     PANIC("Cuda error: input num radix blocks must be the same")
@@ -167,6 +173,9 @@ void cuda_integer_are_all_comparisons_block_true_64_async(
     CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_out,
     CudaRadixCiphertextFFI const *lwe_array_in, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks, uint32_t num_radix_blocks) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   int_comparison_buffer<uint64_t> *buffer =
       (int_comparison_buffer<uint64_t> *)mem_ptr;
@@ -209,6 +218,9 @@ void cuda_integer_is_at_least_one_comparisons_block_true_64_async(
     CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_out,
     CudaRadixCiphertextFFI const *lwe_array_in, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks, uint32_t num_radix_blocks) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   int_comparison_buffer<uint64_t> *buffer =
       (int_comparison_buffer<uint64_t> *)mem_ptr;

backends/tfhe-cuda-backend/cuda/src/integer/div_rem.cu

@@ -27,6 +27,21 @@ void cuda_integer_div_rem_64_async(CudaStreamsFFI streams,
                                    CudaRadixCiphertextFFI const *divisor,
                                    bool is_signed, int8_t *mem_ptr,
                                    void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(quotient != numerator,
+                 "Quotient and numerator pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(quotient != divisor,
+                 "Quotient and divisor pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(remainder != numerator,
+                 "Remainder and numerator pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(remainder != divisor,
+                 "Remainder and divisor pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(quotient != remainder,
+                 "Quotient and remainder pointers must be different for "
+                 "out-of-place operations");
   PUSH_RANGE("div")
   auto mem = (int_div_rem_memory<uint64_t> *)mem_ptr;
 

backends/tfhe-cuda-backend/cuda/src/integer/ilog2.cu

@@ -30,6 +30,9 @@ void cuda_integer_count_of_consecutive_bits_64_async(
     CudaStreamsFFI streams, CudaRadixCiphertextFFI *output_ct,
     CudaRadixCiphertextFFI const *input_ct, int8_t *mem_ptr, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(output_ct != input_ct,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_integer_count_of_consecutive_bits<uint64_t, uint64_t>(
       CudaStreams(streams), output_ct, input_ct,
@@ -81,6 +84,9 @@ void cuda_integer_ilog2_64_async(
     CudaRadixCiphertextFFI const *trivial_ct_2,
     CudaRadixCiphertextFFI const *trivial_ct_m_minus_1_block, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(output_ct != input_ct,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_integer_ilog2<uint64_t, uint64_t>(
       CudaStreams(streams), output_ct, input_ct, trivial_ct_neg_n, trivial_ct_2,

backends/tfhe-cuda-backend/cuda/src/integer/integer.cu

@@ -212,6 +212,9 @@ void cuda_apply_univariate_lut_64_async(
     CudaStreamsFFI streams, CudaRadixCiphertextFFI *output_radix_lwe,
     CudaRadixCiphertextFFI const *input_radix_lwe, int8_t *mem_ptr,
     void *const *ksks, void *const *bsks) {
+  PANIC_IF_FALSE(output_radix_lwe != input_radix_lwe,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_apply_univariate_lut<uint64_t>(
       CudaStreams(streams), output_radix_lwe, input_radix_lwe,
@@ -243,6 +246,9 @@ void cuda_apply_many_univariate_lut_64_async(
     CudaRadixCiphertextFFI const *input_radix_lwe, int8_t *mem_ptr,
     void *const *ksks, void *const *bsks, uint32_t num_many_lut,
     uint32_t lut_stride) {
+  PANIC_IF_FALSE(output_radix_lwe != input_radix_lwe,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_apply_many_univariate_lut<uint64_t>(
       CudaStreams(streams), output_radix_lwe, input_radix_lwe,
@@ -312,6 +318,9 @@ void cuda_apply_noise_squashing_async(
     CudaStreamsFFI streams, CudaRadixCiphertextFFI *output_radix_lwe,
     CudaRadixCiphertextFFI const *input_radix_lwe, int8_t *mem_ptr,
     void *const *ksks, void *const *bsks) {
+  PANIC_IF_FALSE(output_radix_lwe != input_radix_lwe,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   PUSH_RANGE("apply noise squashing")
   integer_radix_apply_noise_squashing<uint64_t>(

backends/tfhe-cuda-backend/cuda/src/integer/multiplication.cu

@@ -61,11 +61,65 @@ void generate_ids_update_degrees(uint64_t *terms_degree, size_t *h_lwe_idx_in,
 
   total_count = message_count + carry_count;
 }
-/*
- * This scratch function allocates the necessary amount of data on the GPU for
- * the integer radix multiplication in keyswitch->bootstrap order.
- */
-uint64_t scratch_cuda_integer_mult_64_async(
+void cuda_integer_mult_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *radix_lwe_inout,
+    bool const is_bool_left, CudaRadixCiphertextFFI const *radix_lwe_right,
+    bool const is_bool_right, void *const *bsks, void *const *ksks,
+    int8_t *mem_ptr, uint32_t polynomial_size, uint32_t num_blocks) {
+  // In-place variant: radix_lwe_inout *= radix_lwe_right, no aliasing check
+  // needed
+  PUSH_RANGE("mul_inplace")
+  switch (polynomial_size) {
+  case 256:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<256>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  case 512:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<512>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  case 1024:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<1024>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  case 2048:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<2048>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  case 4096:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<4096>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  case 8192:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<8192>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  case 16384:
+    host_integer_mult_radix<uint64_t, AmortizedDegree<16384>>(
+        CudaStreams(streams), radix_lwe_inout, radix_lwe_inout, is_bool_left,
+        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
+    break;
+  default:
+    PANIC("Cuda error (integer multiplication): unsupported polynomial size. "
+          "Supported N's are powers of two in the interval [256..16384].")
+  }
+  POP_RANGE()
+}
+
+uint64_t scratch_cuda_integer_mult_inplace_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, bool const is_boolean_left,
     bool const is_boolean_right, uint32_t message_modulus,
     uint32_t carry_modulus, uint32_t glwe_dimension, uint32_t lwe_dimension,
@@ -97,94 +151,8 @@ uint64_t scratch_cuda_integer_mult_64_async(
   }
 }
 
-/*
- * Computes a multiplication between two 64 bit radix lwe ciphertexts
- * encrypting integer values. keyswitch -> bootstrap pattern is used, function
- * works for single pair of radix ciphertexts, 'v_stream' can be used for
- * parallelization
- * - 'v_stream' is a void pointer to the Cuda stream to be used in the kernel
- * launch
- * - 'gpu_index' is the index of the GPU to be used in the kernel launch
- * - 'radix_lwe_out' is 64 bit radix big lwe ciphertext, product of
- * multiplication
- * - 'radix_lwe_left' left radix big lwe ciphertext
- * - 'radix_lwe_right' right radix big lwe ciphertext
- * - 'bsk' bootstrapping key in fourier domain
- * - 'ksk' keyswitching key
- * - 'mem_ptr'
- * - 'message_modulus' message_modulus
- * - 'carry_modulus' carry_modulus
- * - 'glwe_dimension' glwe_dimension
- * - 'lwe_dimension' is the dimension of small lwe ciphertext
- * - 'polynomial_size' polynomial size
- * - 'pbs_base_log' base log used in the pbs
- * - 'pbs_level' decomposition level count used in the pbs
- * - 'ks_level' decomposition level count used in the keyswitch
- * - 'num_blocks' is the number of big lwe ciphertext blocks inside radix
- * ciphertext
- * - 'pbs_type' selects which PBS implementation should be used
- */
-void cuda_integer_mult_64_async(CudaStreamsFFI streams,
-                                CudaRadixCiphertextFFI *radix_lwe_out,
-                                CudaRadixCiphertextFFI const *radix_lwe_left,
-                                bool const is_bool_left,
-                                CudaRadixCiphertextFFI const *radix_lwe_right,
-                                bool const is_bool_right, void *const *bsks,
-                                void *const *ksks, int8_t *mem_ptr,
-                                uint32_t polynomial_size, uint32_t num_blocks) {
-  PUSH_RANGE("mul")
-  switch (polynomial_size) {
-  case 256:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<256>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  case 512:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<512>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  case 1024:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<1024>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  case 2048:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<2048>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  case 4096:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<4096>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  case 8192:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<8192>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  case 16384:
-    host_integer_mult_radix<uint64_t, AmortizedDegree<16384>>(
-        CudaStreams(streams), radix_lwe_out, radix_lwe_left, is_bool_left,
-        radix_lwe_right, is_bool_right, bsks, (uint64_t **)(ksks),
-        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
-    break;
-  default:
-    PANIC("Cuda error (integer multiplication): unsupported polynomial size. "
-          "Supported N's are powers of two in the interval [256..16384].")
-  }
-  POP_RANGE()
-}
-
-void cleanup_cuda_integer_mult_64(CudaStreamsFFI streams,
-                                  int8_t **mem_ptr_void) {
+void cleanup_cuda_integer_mult_inplace_64(CudaStreamsFFI streams,
+                                          int8_t **mem_ptr_void) {
   PUSH_RANGE("cleanup mul")
   int_mul_memory<uint64_t> *mem_ptr =
       (int_mul_memory<uint64_t> *)(*mem_ptr_void);
@@ -221,6 +189,9 @@ void cuda_partial_sum_ciphertexts_vec_64_async(
     CudaStreamsFFI streams, CudaRadixCiphertextFFI *radix_lwe_out,
     CudaRadixCiphertextFFI *radix_lwe_vec, int8_t *mem_ptr, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(radix_lwe_out != radix_lwe_vec,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   auto mem = (int_sum_ciphertexts_vec_memory<uint64_t> *)mem_ptr;
   if (radix_lwe_vec->num_radix_blocks % radix_lwe_out->num_radix_blocks != 0)

backends/tfhe-cuda-backend/cuda/src/integer/negation.cu

@@ -5,6 +5,9 @@ void cuda_negate_ciphertext_64(CudaStreamsFFI streams,
                                CudaRadixCiphertextFFI const *lwe_array_in,
                                uint32_t message_modulus, uint32_t carry_modulus,
                                uint32_t num_radix_blocks) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   auto cuda_streams = CudaStreams(streams);
   host_negation<uint64_t>(cuda_streams, lwe_array_out, lwe_array_in,

backends/tfhe-cuda-backend/cuda/src/integer/rerand.cu

@@ -5,7 +5,7 @@ uint64_t scratch_cuda_rerand_64_async(
     CudaStreamsFFI streams, int8_t **mem_ptr, uint32_t big_lwe_dimension,
     uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
     uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, bool allocate_gpu_memory) {
+    uint32_t carry_modulus, bool allocate_gpu_memory, RERAND_MODE rerand_type) {
   PUSH_RANGE("scratch rerand")
   int_radix_params params(PBS_TYPE::CLASSICAL, 0, 0, big_lwe_dimension,
                           small_lwe_dimension, ks_level, ks_base_log, 0, 0, 0,
@@ -13,8 +13,9 @@ uint64_t scratch_cuda_rerand_64_async(
                           PBS_MS_REDUCTION_T::NO_REDUCTION);
 
   uint64_t ret = scratch_cuda_rerand<uint64_t>(
-      CudaStreams(streams), (int_rerand_mem<uint64_t> **)mem_ptr,
-      lwe_ciphertext_count, params, allocate_gpu_memory);
+      CudaStreams(streams),
+      reinterpret_cast<int_rerand_mem<uint64_t> **>(mem_ptr),
+      lwe_ciphertext_count, params, allocate_gpu_memory, rerand_type);
   POP_RANGE()
   return ret;
 }
@@ -28,7 +29,7 @@ void cuda_rerand_64_async(
     CudaStreamsFFI streams, void *lwe_array,
     const void *lwe_flattened_encryptions_of_zero_compact_array_in,
     int8_t *mem_ptr, void *const *ksk) {
-
+  PUSH_RANGE("rerand")
   auto rerand_buffer = reinterpret_cast<int_rerand_mem<uint64_t> *>(mem_ptr);
 
   switch (rerand_buffer->params.big_lwe_dimension) {
@@ -37,49 +38,49 @@ void cuda_rerand_64_async(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   case 512:
     host_rerand_inplace<uint64_t, AmortizedDegree<512>>(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   case 1024:
     host_rerand_inplace<uint64_t, AmortizedDegree<1024>>(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   case 2048:
     host_rerand_inplace<uint64_t, AmortizedDegree<2048>>(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   case 4096:
     host_rerand_inplace<uint64_t, AmortizedDegree<4096>>(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   case 8192:
     host_rerand_inplace<uint64_t, AmortizedDegree<8192>>(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   case 16384:
     host_rerand_inplace<uint64_t, AmortizedDegree<16384>>(
         streams, static_cast<uint64_t *>(lwe_array),
         static_cast<const uint64_t *>(
             lwe_flattened_encryptions_of_zero_compact_array_in),
-        (uint64_t **)(ksk), rerand_buffer);
+        reinterpret_cast<uint64_t *const *>(ksk), rerand_buffer);
     break;
   default:
     PANIC("CUDA error: lwe_dimension not supported."
@@ -87,12 +88,12 @@ void cuda_rerand_64_async(
           " in the interval [256..16384].");
     break;
   }
+  POP_RANGE()
 }
 
 void cleanup_cuda_rerand_64(CudaStreamsFFI streams, int8_t **mem_ptr_void) {
   PUSH_RANGE("cleanup rerand")
-  int_rerand_mem<uint64_t> *mem_ptr =
-      (int_rerand_mem<uint64_t> *)(*mem_ptr_void);
+  auto *mem_ptr = reinterpret_cast<int_rerand_mem<uint64_t> *>(*mem_ptr_void);
   mem_ptr->release(CudaStreams(streams));
   delete mem_ptr;
   *mem_ptr_void = nullptr;

backends/tfhe-cuda-backend/cuda/src/integer/rerand.cuh

@@ -14,26 +14,25 @@ void host_rerand_inplace(
     CudaStreams const streams, Torus *lwe_array,
     const Torus *lwe_flattened_encryptions_of_zero_compact_array_in,
     Torus *const *ksk, int_rerand_mem<Torus> *mem_ptr) {
-  auto zero_lwes = mem_ptr->tmp_zero_lwes;
+  auto rerand_mode = mem_ptr->rerand_mode;
+  auto expanded_zero_lwes = mem_ptr->tmp_expanded_zero_lwes;
   auto num_lwes = mem_ptr->num_lwes;
-  auto ksed_zero_lwes = mem_ptr->tmp_ksed_zero_lwes;
-  auto lwe_trivial_indexes = mem_ptr->lwe_trivial_indexes;
-  auto ksk_params = mem_ptr->params;
-  auto output_dimension = ksk_params.small_lwe_dimension;
-  auto input_dimension = ksk_params.big_lwe_dimension;
-  auto ks_level = ksk_params.ks_level;
-  auto ks_base_log = ksk_params.ks_base_log;
-  auto message_modulus = ksk_params.message_modulus;
-  auto carry_modulus = ksk_params.carry_modulus;
 
-  GPU_ASSERT(sizeof(Torus) == 8,
-             "Cuda error: expand is only supported on 64 bits");
+  auto rerand_params = mem_ptr->params;
+  auto message_modulus = rerand_params.message_modulus;
+  auto carry_modulus = rerand_params.carry_modulus;
+  auto input_dimension = rerand_params.big_lwe_dimension;
+  // Default to input dimension; overridden to small_lwe_dimension in the KS
+  // path
+  auto output_dimension = input_dimension;
+
+  static_assert(sizeof(Torus) == 8, "expand is only supported on 64 bits");
 
   // Expand encryptions of zero
   // Wraps the input into a flattened_compact_lwe_lists type
   auto compact_lwe_lists = flattened_compact_lwe_lists<Torus>(
       const_cast<Torus *>(lwe_flattened_encryptions_of_zero_compact_array_in),
-      &num_lwes, (uint32_t)1, input_dimension);
+      &num_lwes, static_cast<uint32_t>(1), input_dimension);
   auto h_expand_jobs = mem_ptr->h_expand_jobs;
   auto d_expand_jobs = mem_ptr->d_expand_jobs;
 
@@ -53,20 +52,30 @@ void host_rerand_inplace(
       streams.stream(0), streams.gpu_index(0), true);
 
   host_lwe_expand<Torus, params>(streams.stream(0), streams.gpu_index(0),
-                                 zero_lwes, d_expand_jobs, num_lwes);
+                                 expanded_zero_lwes, d_expand_jobs, num_lwes);
 
-  // Keyswitch
-  execute_keyswitch_async<Torus>(
-      streams.get_ith(0), ksed_zero_lwes, lwe_trivial_indexes, zero_lwes,
-      lwe_trivial_indexes, ksk, input_dimension, output_dimension, ks_base_log,
-      ks_level, num_lwes, true, mem_ptr->ks_tmp_buf_vec);
+  auto lwes_to_be_added = expanded_zero_lwes;
+  if (rerand_mode == RERAND_MODE::RERAND_WITH_KS) {
+    lwes_to_be_added = mem_ptr->tmp_ksed_expanded_zero_lwes;
+    output_dimension = rerand_params.small_lwe_dimension;
+    auto ks_level = rerand_params.ks_level;
+    auto ks_base_log = rerand_params.ks_base_log;
+    auto lwe_trivial_indexes = mem_ptr->lwe_trivial_indexes;
+
+    // Keyswitch
+    execute_keyswitch_async<Torus>(streams.get_ith(0), lwes_to_be_added,
+                                   lwe_trivial_indexes, expanded_zero_lwes,
+                                   lwe_trivial_indexes, ksk, input_dimension,
+                                   output_dimension, ks_base_log, ks_level,
+                                   num_lwes, true, mem_ptr->ks_tmp_buf_vec);
+  }
 
   // Add ks output to ct
   // Check sizes
   CudaRadixCiphertextFFI lwes_ffi;
   into_radix_ciphertext(&lwes_ffi, lwe_array, num_lwes, output_dimension);
   CudaRadixCiphertextFFI ksed_zero_lwes_ffi;
-  into_radix_ciphertext(&ksed_zero_lwes_ffi, ksed_zero_lwes, num_lwes,
+  into_radix_ciphertext(&ksed_zero_lwes_ffi, lwes_to_be_added, num_lwes,
                         output_dimension);
   host_addition<Torus>(streams.stream(0), streams.gpu_index(0), &lwes_ffi,
                        &lwes_ffi, &ksed_zero_lwes_ffi, num_lwes,
@@ -81,10 +90,11 @@ __host__ uint64_t scratch_cuda_rerand(CudaStreams streams,
                                       int_rerand_mem<Torus> **mem_ptr,
                                       uint32_t num_lwes,
                                       int_radix_params params,
-                                      bool allocate_gpu_memory) {
+                                      bool allocate_gpu_memory,
+                                      RERAND_MODE rerand_mode) {
 
   uint64_t size_tracker = 0;
-  *mem_ptr = new int_rerand_mem<Torus>(streams, params, num_lwes,
+  *mem_ptr = new int_rerand_mem<Torus>(streams, params, num_lwes, rerand_mode,
                                        allocate_gpu_memory, size_tracker);
   return size_tracker;
 }

backends/tfhe-cuda-backend/cuda/src/integer/scalar_bitops.cu

@@ -1,13 +1,13 @@
 #include "integer/scalar_bitops.cuh"
 
-void cuda_integer_scalar_bitop_64_async(
-    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_out,
-    CudaRadixCiphertextFFI const *lwe_array_input, void const *clear_blocks,
-    void const *h_clear_blocks, uint32_t num_clear_blocks, int8_t *mem_ptr,
-    void *const *bsks, void *const *ksks) {
-
+void cuda_integer_scalar_bitop_inplace_64_async(
+    CudaStreamsFFI streams, CudaRadixCiphertextFFI *lwe_array_inout,
+    void const *clear_blocks, void const *h_clear_blocks,
+    uint32_t num_clear_blocks, int8_t *mem_ptr, void *const *bsks,
+    void *const *ksks) {
+  // In-place variant: lwe_array_inout op= scalar, no aliasing check needed
   host_scalar_bitop<uint64_t>(
-      CudaStreams(streams), lwe_array_out, lwe_array_input,
+      CudaStreams(streams), lwe_array_inout, lwe_array_inout,
       static_cast<const uint64_t *>(clear_blocks),
       static_cast<const uint64_t *>(h_clear_blocks), num_clear_blocks,
       (int_bitop_buffer<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks));

backends/tfhe-cuda-backend/cuda/src/integer/scalar_comparison.cu

@@ -36,6 +36,9 @@ void cuda_integer_scalar_comparison_64_async(
     CudaRadixCiphertextFFI const *lwe_array_in, void const *scalar_blocks,
     void const *h_scalar_blocks, int8_t *mem_ptr, void *const *bsks,
     void *const *ksks, uint32_t num_scalar_blocks) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   // The output ciphertext might be a boolean block or a radix ciphertext
   // depending on the case (eq/gt vs max/min) so the amount of blocks to

backends/tfhe-cuda-backend/cuda/src/integer/scalar_div.cu

@@ -118,6 +118,9 @@ void cuda_integer_unsigned_scalar_div_rem_radix_64_async(
     uint64_t const *decomposed_divisor, uint32_t const num_scalars_divisor,
     void const *clear_blocks, void const *h_clear_blocks,
     uint32_t num_clear_blocks) {
+  PANIC_IF_FALSE(quotient_ct != remainder_ct,
+                 "Quotient and remainder pointers must be different for "
+                 "out-of-place operations");
 
   host_integer_unsigned_scalar_div_rem_radix<uint64_t>(
       CudaStreams(streams), quotient_ct, remainder_ct,
@@ -168,6 +171,9 @@ void cuda_integer_signed_scalar_div_rem_radix_64_async(
     uint64_t const *divisor_has_at_least_one_set,
     uint64_t const *decomposed_divisor, uint32_t const num_scalars_divisor,
     uint32_t numerator_bits) {
+  PANIC_IF_FALSE(quotient_ct != remainder_ct,
+                 "Quotient and remainder pointers must be different for "
+                 "out-of-place operations");
 
   host_integer_signed_scalar_div_rem_radix<uint64_t>(
       CudaStreams(streams), quotient_ct, remainder_ct,

backends/tfhe-cuda-backend/cuda/src/integer/vector_comparison.cu

@@ -25,6 +25,12 @@ void cuda_unchecked_all_eq_slices_64_async(
     CudaRadixCiphertextFFI const *lhs, CudaRadixCiphertextFFI const *rhs,
     uint32_t num_inputs, uint32_t num_blocks, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(match_ct != lhs,
+                 "Output and first input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(match_ct != rhs,
+                 "Output and second input pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_all_eq_slices<uint64_t>(
       CudaStreams(streams), match_ct, lhs, rhs, num_inputs, num_blocks,
@@ -68,6 +74,12 @@ void cuda_unchecked_contains_sub_slice_64_async(
     CudaRadixCiphertextFFI const *lhs, CudaRadixCiphertextFFI const *rhs,
     uint32_t num_rhs, uint32_t num_blocks, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(match_ct != lhs,
+                 "Output and first input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(match_ct != rhs,
+                 "Output and second input pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_contains_sub_slice<uint64_t>(
       CudaStreams(streams), match_ct, lhs, rhs, num_rhs, num_blocks,

backends/tfhe-cuda-backend/cuda/src/integer/vector_find.cu

@@ -27,6 +27,15 @@ void cuda_unchecked_match_value_64_async(
     CudaRadixCiphertextFFI const *lwe_array_in_ct,
     const uint64_t *h_match_inputs, const uint64_t *h_match_outputs,
     int8_t *mem, void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(lwe_array_out_result != lwe_array_in_ct,
+                 "Output result and input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(lwe_array_out_boolean != lwe_array_in_ct,
+                 "Output boolean and input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(lwe_array_out_result != lwe_array_out_boolean,
+                 "Result and boolean output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_match_value<uint64_t>(
       CudaStreams(streams), lwe_array_out_result, lwe_array_out_boolean,
@@ -75,6 +84,9 @@ void cuda_unchecked_match_value_or_64_async(
     const uint64_t *h_match_inputs, const uint64_t *h_match_outputs,
     const uint64_t *h_or_value, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in_ct,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_unchecked_match_value_or<uint64_t>(
       CudaStreams(streams), lwe_array_out, lwe_array_in_ct, h_match_inputs,
@@ -120,6 +132,12 @@ void cuda_unchecked_contains_64_async(CudaStreamsFFI streams,
                                       uint32_t num_inputs, uint32_t num_blocks,
                                       int8_t *mem, void *const *bsks,
                                       void *const *ksks) {
+  PANIC_IF_FALSE(output != inputs,
+                 "Output and first input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(output != value,
+                 "Output and second input pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_contains<uint64_t>(
       CudaStreams(streams), output, inputs, value, num_inputs, num_blocks,
@@ -163,6 +181,8 @@ void cuda_unchecked_contains_clear_64_async(
     CudaRadixCiphertextFFI const *inputs, const uint64_t *h_clear_val,
     uint32_t num_inputs, uint32_t num_blocks, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(output != inputs, "Output and input pointers must be "
+                                   "different for out-of-place operations");
 
   host_unchecked_contains_clear<uint64_t>(
       CudaStreams(streams), output, inputs, h_clear_val, num_inputs, num_blocks,
@@ -206,6 +226,8 @@ void cuda_unchecked_is_in_clears_64_async(
     CudaRadixCiphertextFFI const *input, const uint64_t *h_cleartexts,
     uint32_t num_clears, uint32_t num_blocks, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(output != input, "Output and input pointers must be different "
+                                  "for out-of-place operations");
 
   host_unchecked_is_in_clears<uint64_t>(
       CudaStreams(streams), output, input, h_cleartexts, num_clears, num_blocks,
@@ -250,6 +272,13 @@ void cuda_unchecked_index_in_clears_64_async(
     const uint64_t *h_cleartexts, uint32_t num_clears, uint32_t num_blocks,
     uint32_t num_blocks_index, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(index_ct != input, "Output and input pointers must be "
+                                    "different for out-of-place operations");
+  PANIC_IF_FALSE(match_ct != input, "Output and input pointers must be "
+                                    "different for out-of-place operations");
+  PANIC_IF_FALSE(index_ct != match_ct,
+                 "Index and match output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_index_in_clears<uint64_t>(
       CudaStreams(streams), index_ct, match_ct, input, h_cleartexts, num_clears,
@@ -295,6 +324,13 @@ void cuda_unchecked_first_index_in_clears_64_async(
     const uint64_t *h_unique_values, const uint64_t *h_unique_indices,
     uint32_t num_unique, uint32_t num_blocks, uint32_t num_blocks_index,
     int8_t *mem, void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(index_ct != input, "Output and input pointers must be "
+                                    "different for out-of-place operations");
+  PANIC_IF_FALSE(match_ct != input, "Output and input pointers must be "
+                                    "different for out-of-place operations");
+  PANIC_IF_FALSE(index_ct != match_ct,
+                 "Index and match output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_first_index_in_clears<uint64_t>(
       CudaStreams(streams), index_ct, match_ct, input, h_unique_values,
@@ -340,6 +376,13 @@ void cuda_unchecked_first_index_of_clear_64_async(
     const uint64_t *h_clear_val, uint32_t num_inputs, uint32_t num_blocks,
     uint32_t num_blocks_index, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(index_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(match_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(index_ct != match_ct,
+                 "Index and match output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_first_index_of_clear<uint64_t>(
       CudaStreams(streams), index_ct, match_ct, inputs, h_clear_val, num_inputs,
@@ -385,6 +428,13 @@ void cuda_unchecked_first_index_of_64_async(
     CudaRadixCiphertextFFI const *value, uint32_t num_inputs,
     uint32_t num_blocks, uint32_t num_blocks_index, int8_t *mem,
     void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(index_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(match_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(index_ct != match_ct,
+                 "Index and match output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_first_index_of<uint64_t>(
       CudaStreams(streams), index_ct, match_ct, inputs, value, num_inputs,
@@ -431,6 +481,13 @@ void cuda_unchecked_index_of_64_async(CudaStreamsFFI streams,
                                       uint32_t num_inputs, uint32_t num_blocks,
                                       uint32_t num_blocks_index, int8_t *mem,
                                       void *const *bsks, void *const *ksks) {
+  PANIC_IF_FALSE(index_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(match_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(index_ct != match_ct,
+                 "Index and match output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_index_of<uint64_t>(
       CudaStreams(streams), index_ct, match_ct, inputs, value, num_inputs,
@@ -477,6 +534,13 @@ void cuda_unchecked_index_of_clear_64_async(
     uint32_t num_inputs, uint32_t num_blocks, uint32_t num_scalar_blocks,
     uint32_t num_blocks_index, int8_t *mem, void *const *bsks,
     void *const *ksks) {
+  PANIC_IF_FALSE(index_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(match_ct != inputs, "Output and input pointers must be "
+                                     "different for out-of-place operations");
+  PANIC_IF_FALSE(index_ct != match_ct,
+                 "Index and match output pointers must be different for "
+                 "out-of-place operations");
 
   host_unchecked_index_of_clear<uint64_t>(
       CudaStreams(streams), index_ct, match_ct, inputs,

backends/tfhe-cuda-backend/cuda/src/linearalgebra/addition.cu

@@ -5,6 +5,12 @@ void cuda_add_lwe_ciphertext_vector_32(void *stream, uint32_t gpu_index,
                                        CudaRadixCiphertextFFI *output,
                                        CudaRadixCiphertextFFI const *input_1,
                                        CudaRadixCiphertextFFI const *input_2) {
+  PANIC_IF_FALSE(output != input_1,
+                 "Output and first input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(output != input_2,
+                 "Output and second input pointers must be different for "
+                 "out-of-place operations");
 
   if (output->num_radix_blocks != input_1->num_radix_blocks ||
       output->num_radix_blocks != input_2->num_radix_blocks)
@@ -44,6 +50,12 @@ void cuda_add_lwe_ciphertext_vector_64(void *stream, uint32_t gpu_index,
                                        CudaRadixCiphertextFFI *output,
                                        CudaRadixCiphertextFFI const *input_1,
                                        CudaRadixCiphertextFFI const *input_2) {
+  PANIC_IF_FALSE(output != input_1,
+                 "Output and first input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(output != input_2,
+                 "Output and second input pointers must be different for "
+                 "out-of-place operations");
 
   if (output->num_radix_blocks != input_1->num_radix_blocks ||
       output->num_radix_blocks != input_2->num_radix_blocks)
@@ -53,6 +65,30 @@ void cuda_add_lwe_ciphertext_vector_64(void *stream, uint32_t gpu_index,
   cuda_synchronize_stream(static_cast<cudaStream_t>(stream), gpu_index);
 }
 
+void cuda_add_lwe_ciphertext_vector_inplace_32(
+    void *stream, uint32_t gpu_index, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *input_2) {
+  // In-place variant: lwe_array_inout += input_2, no aliasing check needed
+  if (lwe_array_inout->num_radix_blocks != input_2->num_radix_blocks)
+    PANIC("Cuda error: input and output num radix blocks must be the same")
+  host_addition<uint32_t>(static_cast<cudaStream_t>(stream), gpu_index,
+                          lwe_array_inout, lwe_array_inout, input_2,
+                          lwe_array_inout->num_radix_blocks, 0, 0);
+  cuda_synchronize_stream(static_cast<cudaStream_t>(stream), gpu_index);
+}
+
+void cuda_add_lwe_ciphertext_vector_inplace_64(
+    void *stream, uint32_t gpu_index, CudaRadixCiphertextFFI *lwe_array_inout,
+    CudaRadixCiphertextFFI const *input_2) {
+  // In-place variant: lwe_array_inout += input_2, no aliasing check needed
+  if (lwe_array_inout->num_radix_blocks != input_2->num_radix_blocks)
+    PANIC("Cuda error: input and output num radix blocks must be the same")
+  host_addition<uint64_t>(static_cast<cudaStream_t>(stream), gpu_index,
+                          lwe_array_inout, lwe_array_inout, input_2,
+                          lwe_array_inout->num_radix_blocks, 0, 0);
+  cuda_synchronize_stream(static_cast<cudaStream_t>(stream), gpu_index);
+}
+
 /*
  * Perform the addition of a u32 input LWE ciphertext vector with a u32
  * plaintext vector. See the equivalent operation on u64 data for more details.
@@ -62,6 +98,9 @@ void cuda_add_lwe_ciphertext_vector_plaintext_vector_32(
     void const *lwe_array_in, void const *plaintext_array_in,
     const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_addition_plaintext<uint32_t>(
       static_cast<cudaStream_t>(stream), gpu_index,
@@ -104,6 +143,9 @@ void cuda_add_lwe_ciphertext_vector_plaintext_vector_64(
     void const *lwe_array_in, void const *plaintext_array_in,
     const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_addition_plaintext<uint64_t>(
       static_cast<cudaStream_t>(stream), gpu_index,
@@ -144,6 +186,9 @@ void cuda_add_lwe_ciphertext_vector_plaintext_64(
     void const *lwe_array_in, const uint64_t plaintext_in,
     const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_addition_plaintext_scalar<uint64_t>(
       static_cast<cudaStream_t>(stream), gpu_index,

backends/tfhe-cuda-backend/cuda/src/linearalgebra/multiplication.cu

@@ -10,6 +10,9 @@ void cuda_mult_lwe_ciphertext_vector_cleartext_vector_32(
     void const *lwe_array_in, void const *cleartext_array_in,
     const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_cleartext_vec_multiplication<uint32_t>(
       static_cast<cudaStream_t>(stream), gpu_index,
@@ -52,6 +55,9 @@ void cuda_mult_lwe_ciphertext_vector_cleartext_vector_64(
     void const *lwe_array_in, void const *cleartext_array_in,
     const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_cleartext_vec_multiplication<uint64_t>(
       static_cast<cudaStream_t>(stream), gpu_index,
@@ -80,6 +86,12 @@ void cuda_wrapping_polynomial_mul_one_to_many_64_async(
     void *stream, uint32_t gpu_index, void *result, void const *poly_lhs,
     int8_t *circulant, void const *poly_rhs, uint32_t polynomial_size,
     uint32_t n_rhs) {
+  PANIC_IF_FALSE(result != poly_lhs,
+                 "Output and left input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(result != poly_rhs,
+                 "Output and right input pointers must be different for "
+                 "out-of-place operations");
 
   host_wrapping_polynomial_mul_one_to_many<uint64_t, ulonglong4>(
       static_cast<cudaStream_t>(stream), gpu_index,
@@ -92,6 +104,13 @@ void cuda_glwe_wrapping_polynomial_mul_one_to_many_64_async(
     void *stream, uint32_t gpu_index, void *result, void const *glwe_lhs,
     int8_t *circulant, void const *poly_rhs, uint32_t polynomial_size,
     uint32_t glwe_dimension, uint32_t n_rhs) {
+  PANIC_IF_FALSE(result != glwe_lhs,
+                 "Output and left input pointers must be different for "
+                 "out-of-place operations");
+  PANIC_IF_FALSE(result != poly_rhs,
+                 "Output and right input pointers must be different for "
+                 "out-of-place operations");
+
   host_glwe_wrapping_polynomial_mul_one_to_many<uint64_t, ulonglong4>(
       static_cast<cudaStream_t>(stream), gpu_index,
       static_cast<uint64_t *>(result), static_cast<uint64_t const *>(glwe_lhs),

backends/tfhe-cuda-backend/cuda/src/linearalgebra/negation.cu

@@ -8,6 +8,9 @@ void cuda_negate_lwe_ciphertext_vector_32(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
     void const *lwe_array_in, const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_negation<uint32_t>(static_cast<cudaStream_t>(stream), gpu_index,
                           static_cast<uint32_t *>(lwe_array_out),
@@ -42,6 +45,9 @@ void cuda_negate_lwe_ciphertext_vector_64(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
     void const *lwe_array_in, const uint32_t input_lwe_dimension,
     const uint32_t input_lwe_ciphertext_count) {
+  PANIC_IF_FALSE(lwe_array_out != lwe_array_in,
+                 "Output and input pointers must be different for out-of-place "
+                 "operations");
 
   host_negation<uint64_t>(static_cast<cudaStream_t>(stream), gpu_index,
                           static_cast<uint64_t *>(lwe_array_out),

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_cg_classic.cuh

@@ -373,7 +373,8 @@ __host__ bool verify_cuda_programmable_bootstrap_cg_grid_size(
 
   // Get the number of streaming multiprocessors
   int number_of_sm = 0;
-  cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0);
+  check_cuda_error(
+      cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0));
   return number_of_blocks <= max_active_blocks_per_sm * number_of_sm;
 }
 

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_cg_multibit.cuh

@@ -420,6 +420,39 @@ __host__ void host_cg_multi_bit_programmable_bootstrap(
   }
 }
 
+// Noise tests variant: identical to host_cg_multi_bit_programmable_bootstrap
+// but uses NOISE_TESTS keybundle mode.
+template <typename Torus, class params>
+__host__ void host_cg_multi_bit_programmable_bootstrap_noise_tests(
+    cudaStream_t stream, uint32_t gpu_index, Torus *lwe_array_out,
+    Torus const *lwe_output_indexes, Torus const *lut_vector,
+    Torus const *lut_vector_indexes, Torus const *lwe_array_in,
+    Torus const *lwe_input_indexes, uint64_t const *bootstrapping_key,
+    pbs_buffer<Torus, MULTI_BIT> *buffer, uint32_t glwe_dimension,
+    uint32_t lwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride) {
+
+  auto lwe_chunk_size = buffer->lwe_chunk_size;
+
+  for (uint32_t lwe_offset = 0; lwe_offset < (lwe_dimension / grouping_factor);
+       lwe_offset += lwe_chunk_size) {
+
+    // Compute a keybundle with NOISE_TESTS mode instead of GENERIC
+    execute_compute_keybundle_noise_tests<Torus, params>(
+        stream, gpu_index, lwe_array_in, lwe_input_indexes, bootstrapping_key,
+        buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, level_count, lwe_offset);
+
+    execute_cg_external_product_loop<Torus, params>(
+        stream, gpu_index, lut_vector, lut_vector_indexes, lwe_array_in,
+        lwe_input_indexes, lwe_array_out, lwe_output_indexes, buffer,
+        num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, base_log, level_count, lwe_offset, num_many_lut,
+        lut_stride);
+  }
+}
+
 // Verify if the grid size satisfies the cooperative group constraints
 template <typename Torus, class params>
 __host__ bool verify_cuda_programmable_bootstrap_cg_multi_bit_grid_size(
@@ -484,7 +517,8 @@ __host__ bool verify_cuda_programmable_bootstrap_cg_multi_bit_grid_size(
 
   // Get the number of streaming multiprocessors
   int number_of_sm = 0;
-  cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0);
+  check_cuda_error(
+      cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0));
   return number_of_blocks <= max_active_blocks_per_sm * number_of_sm;
 }
 

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_classic_128.cuh

@@ -784,9 +784,9 @@ __host__ uint64_t scratch_programmable_bootstrap_tbc_128(
       device_programmable_bootstrap_tbc_128<InputTorus, params, FULLSM>,
       cudaFuncAttributeMaxDynamicSharedMemorySize,
       full_sm)); // full_sm + minimum_sm_tbc));
-  cudaFuncSetCacheConfig(
+  check_cuda_error(cudaFuncSetCacheConfig(
       device_programmable_bootstrap_tbc_128<InputTorus, params, FULLSM>,
-      cudaFuncCachePreferShared);
+      cudaFuncCachePreferShared));
   check_cuda_error(cudaFuncSetAttribute(
       device_programmable_bootstrap_tbc_128<InputTorus, params, FULLSM>,
       cudaFuncAttributeNonPortableClusterSizeAllowed, true));
@@ -1271,7 +1271,8 @@ __host__ bool verify_cuda_programmable_bootstrap_128_cg_grid_size(
 
   // Get the number of streaming multiprocessors
   int number_of_sm = 0;
-  cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0);
+  check_cuda_error(
+      cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0));
 
   return number_of_blocks <= max_active_blocks_per_sm * number_of_sm;
 }

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_multibit.cu

@@ -645,6 +645,103 @@ void cleanup_cuda_multi_bit_programmable_bootstrap_64(void *stream,
   *buffer = nullptr;
 }
 
+// Noise-tests-namespaced wrappers: delegate to the standard scratch/cleanup so
+// that callers using the noise-tests PBS variant have a consistent API.
+uint64_t scratch_cuda_multi_bit_programmable_bootstrap_noise_tests_64_async(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory) {
+  return scratch_cuda_multi_bit_programmable_bootstrap_64_async(
+      stream, gpu_index, pbs_buffer, glwe_dimension, polynomial_size,
+      level_count, input_lwe_ciphertext_count, allocate_gpu_memory);
+}
+
+void cleanup_cuda_multi_bit_programmable_bootstrap_noise_tests_64(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer) {
+  cleanup_cuda_multi_bit_programmable_bootstrap_64(stream, gpu_index,
+                                                   pbs_buffer);
+}
+
+// Noise tests variant of the 64-bit multi-bit PBS, restricted to
+// polynomial_size=2048. The main difference is that the input
+// is assumed to be modulus switched before bootstrapping.
+void cuda_multi_bit_programmable_bootstrap_noise_tests_64_async(
+    void *stream, uint32_t gpu_index, void *lwe_array_out,
+    void const *lwe_output_indexes, void const *lut_vector,
+    void const *lut_vector_indexes, void const *lwe_array_in,
+    void const *lwe_input_indexes, void const *bootstrapping_key,
+    int8_t *mem_ptr, uint32_t lwe_dimension, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t grouping_factor, uint32_t base_log,
+    uint32_t level_count, uint32_t num_samples, uint32_t num_many_lut,
+    uint32_t lut_stride) {
+
+  PANIC_IF_FALSE(num_samples == 1,
+                 "Cuda error (multi-bit PBS): num_samples (%d) should be 1",
+                 num_samples);
+
+  PANIC_IF_FALSE(base_log <= 64,
+                 "Cuda error (multi-bit PBS): base log (%d) should be <= 64",
+                 base_log);
+  PANIC_IF_FALSE(polynomial_size == 2048,
+                 "Cuda error (multi-bit PBS noise tests): only polynomial "
+                 "size 2048 is supported, got %d.",
+                 polynomial_size);
+
+  pbs_buffer<uint64_t, MULTI_BIT> *buffer =
+      (pbs_buffer<uint64_t, MULTI_BIT> *)mem_ptr;
+
+  switch (buffer->pbs_variant) {
+  case PBS_VARIANT::TBC:
+#if CUDA_ARCH >= 900
+  {
+    host_tbc_multi_bit_programmable_bootstrap_noise_tests<uint64_t,
+                                                          Degree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index,
+        static_cast<uint64_t *>(lwe_array_out),
+        static_cast<const uint64_t *>(lwe_output_indexes),
+        static_cast<const uint64_t *>(lut_vector),
+        static_cast<const uint64_t *>(lut_vector_indexes),
+        static_cast<const uint64_t *>(lwe_array_in),
+        static_cast<const uint64_t *>(lwe_input_indexes),
+        static_cast<const uint64_t *>(bootstrapping_key), buffer,
+        glwe_dimension, lwe_dimension, polynomial_size, grouping_factor,
+        base_log, level_count, num_samples, num_many_lut, lut_stride);
+  } break;
+#else
+    PANIC("Cuda error (multi-bit PBS): TBC pbs is not supported.")
+#endif
+  case PBS_VARIANT::CG:
+    host_cg_multi_bit_programmable_bootstrap_noise_tests<uint64_t,
+                                                         Degree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index,
+        static_cast<uint64_t *>(lwe_array_out),
+        static_cast<const uint64_t *>(lwe_output_indexes),
+        static_cast<const uint64_t *>(lut_vector),
+        static_cast<const uint64_t *>(lut_vector_indexes),
+        static_cast<const uint64_t *>(lwe_array_in),
+        static_cast<const uint64_t *>(lwe_input_indexes),
+        static_cast<const uint64_t *>(bootstrapping_key), buffer,
+        glwe_dimension, lwe_dimension, polynomial_size, grouping_factor,
+        base_log, level_count, num_samples, num_many_lut, lut_stride);
+    break;
+  case PBS_VARIANT::DEFAULT:
+    host_multi_bit_programmable_bootstrap_noise_tests<uint64_t, Degree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index,
+        static_cast<uint64_t *>(lwe_array_out),
+        static_cast<const uint64_t *>(lwe_output_indexes),
+        static_cast<const uint64_t *>(lut_vector),
+        static_cast<const uint64_t *>(lut_vector_indexes),
+        static_cast<const uint64_t *>(lwe_array_in),
+        static_cast<const uint64_t *>(lwe_input_indexes),
+        static_cast<const uint64_t *>(bootstrapping_key), buffer,
+        glwe_dimension, lwe_dimension, polynomial_size, grouping_factor,
+        base_log, level_count, num_samples, num_many_lut, lut_stride);
+    break;
+  default:
+    PANIC("Cuda error (multi-bit PBS): unsupported implementation variant.")
+  }
+}
+
 /**
  * Computes divisors of the product of num_sms (streaming multiprocessors on the
  * GPU) and max_blocks_per_sm (maximum active blocks per SM to launch

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_multibit.cuh

@@ -25,7 +25,8 @@ get_start_ith_ggsw_offset(uint32_t polynomial_size, int glwe_dimension,
          level_count;
 }
 
-template <typename Torus, class params, sharedMemDegree SMD>
+template <typename Torus, class params, sharedMemDegree SMD,
+          bool runs_noise_test = false>
 __global__ void device_multi_bit_programmable_bootstrap_keybundle(
     const Torus *__restrict__ lwe_array_in,
     const Torus *__restrict__ lwe_input_indexes, double2 *keybundle_array,
@@ -55,9 +56,6 @@ __global__ void device_multi_bit_programmable_bootstrap_keybundle(
 
   if (lwe_iteration < (lwe_dimension / grouping_factor)) {
 
-    const Torus *block_lwe_array_in =
-        &lwe_array_in[lwe_input_indexes[input_idx] * (lwe_dimension + 1)];
-
     double2 *keybundle = keybundle_array +
                          // select the input
                          input_idx * keybundle_size_per_input;
@@ -86,10 +84,40 @@ __global__ void device_multi_bit_programmable_bootstrap_keybundle(
     // Precalculate the monomial degrees and store them in shared memory
     uint32_t *monomial_degrees = (uint32_t *)selected_memory;
     if (threadIdx.x < (1 << grouping_factor)) {
-      const Torus *lwe_array_group =
-          block_lwe_array_in + rev_lwe_iteration * grouping_factor;
-      monomial_degrees[threadIdx.x] = calculates_monomial_degree<Torus, params>(
-          lwe_array_group, threadIdx.x, grouping_factor);
+      if constexpr (runs_noise_test == true) {
+        // For noise tests the input array contains the input lwe but also the
+        // modswitched results. This allows to avoid changing the accumulation
+        // kernel for the noise tests since the input body will stay in the same
+        // position. The layout of the input array is the following:
+        // | input lwe     | modswitched inputs       |
+        // | lwe size      | lwe_size*grouping_factor |
+
+        // This offset allows to jump directly to the modswitched inputs,
+        // skipping the input lwe
+        const Torus modswitched_offset = lwe_dimension + 1;
+
+        const Torus *block_lwe_array_in =
+            &lwe_array_in[lwe_input_indexes[input_idx] *
+                              (lwe_dimension / grouping_factor) *
+                              (1 << grouping_factor) +
+                          modswitched_offset];
+
+        const Torus *lwe_array_group =
+            block_lwe_array_in + rev_lwe_iteration * (1 << grouping_factor);
+        monomial_degrees[threadIdx.x] = lwe_array_group[threadIdx.x];
+
+      } else {
+        // In production we calculate the monomial degrees on the fly, since
+        // they are not stored in the input array.
+        const Torus *block_lwe_array_in =
+            &lwe_array_in[lwe_input_indexes[input_idx] * (lwe_dimension + 1)];
+
+        const Torus *lwe_array_group =
+            block_lwe_array_in + rev_lwe_iteration * grouping_factor;
+        monomial_degrees[threadIdx.x] =
+            calculates_monomial_degree<Torus, params>(
+                lwe_array_group, threadIdx.x, grouping_factor);
+      }
     }
     __syncthreads();
 
@@ -145,7 +173,8 @@ __global__ void device_multi_bit_programmable_bootstrap_keybundle(
 // Then we can just calculate the offset needed to apply this coefficients, and
 // the operation transforms into a pointwise vector multiplication, avoiding to
 // perform extra instructions other than MADD
-template <typename Torus, class params, sharedMemDegree SMD>
+template <typename Torus, class params, sharedMemDegree SMD,
+          bool runs_noise_test = false>
 __global__ void device_multi_bit_programmable_bootstrap_keybundle_2_2_params(
     const Torus *__restrict__ lwe_array_in,
     const Torus *__restrict__ lwe_input_indexes, double2 *keybundle_array,
@@ -219,10 +248,40 @@ __global__ void device_multi_bit_programmable_bootstrap_keybundle_2_2_params(
     uint32_t *monomial_degrees = (uint32_t *)selected_memory;
 
     if (threadIdx.x < (1 << grouping_factor)) {
-      const Torus *lwe_array_group =
-          block_lwe_array_in + rev_lwe_iteration * grouping_factor;
-      monomial_degrees[threadIdx.x] = calculates_monomial_degree<Torus, params>(
-          lwe_array_group, threadIdx.x, grouping_factor);
+      if constexpr (runs_noise_test == true) {
+        // For noise tests the input array contains the input lwe but also the
+        // modswitched results. This allows to avoid changing the accumulation
+        // kernel for the noise tests since the input body will stay in the same
+        // position. The layout of the input array is the following:
+        // | input lwe     | modswitched inputs       |
+        // | lwe size      | lwe_size*grouping_factor |
+
+        // This offset allows to jump directly to the modswitched inputs,
+        // skipping the input lwe
+        const Torus modswitched_offset = lwe_dimension + 1;
+
+        const Torus *block_lwe_array_in =
+            &lwe_array_in[lwe_input_indexes[input_idx] *
+                              (lwe_dimension / grouping_factor) *
+                              (1 << grouping_factor) +
+                          modswitched_offset];
+
+        const Torus *lwe_array_group =
+            block_lwe_array_in + rev_lwe_iteration * (1 << grouping_factor);
+        monomial_degrees[threadIdx.x] = lwe_array_group[threadIdx.x];
+
+      } else {
+        // In production we calculate the monomial degrees on the fly, since
+        // they are not stored in the input array.
+        const Torus *block_lwe_array_in =
+            &lwe_array_in[lwe_input_indexes[input_idx] * (lwe_dimension + 1)];
+
+        const Torus *lwe_array_group =
+            block_lwe_array_in + rev_lwe_iteration * grouping_factor;
+        monomial_degrees[threadIdx.x] =
+            calculates_monomial_degree<Torus, params>(
+                lwe_array_group, threadIdx.x, grouping_factor);
+      }
     }
     __syncthreads();
 
@@ -662,6 +721,7 @@ enum class MultiBitKeybundleLaunchMode {
   AUTO,
   GENERIC,
   SPECIALIZED_2_2,
+  NOISE_TESTS,
 };
 
 template <typename Torus, class params>
@@ -726,30 +786,65 @@ __host__ void execute_compute_keybundle_with_mode(
     bool use_specialized =
         launch_mode == MultiBitKeybundleLaunchMode::SPECIALIZED_2_2 ||
         (launch_mode == MultiBitKeybundleLaunchMode::AUTO &&
+         can_use_specialized) ||
+        (launch_mode == MultiBitKeybundleLaunchMode::NOISE_TESTS &&
          can_use_specialized);
+    bool use_noise_test_template =
+        launch_mode == MultiBitKeybundleLaunchMode::NOISE_TESTS;
     if (use_specialized) {
       dim3 thds_new_keybundle(512, 1, 1);
-      check_cuda_error(cudaFuncSetAttribute(
-          device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
-              Torus, Degree<2048>, FULLSM>,
-          cudaFuncAttributeMaxDynamicSharedMemorySize, 3 * full_sm_keybundle));
-      check_cuda_error(cudaFuncSetCacheConfig(
-          device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
-              Torus, Degree<2048>, FULLSM>,
-          cudaFuncCachePreferShared));
-      check_cuda_error(cudaGetLastError());
-      device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
-          Torus, Degree<2048>, FULLSM><<<grid_keybundle, thds_new_keybundle,
-                                         3 * full_sm_keybundle, stream>>>(
-          lwe_array_in, lwe_input_indexes, keybundle_fft, bootstrapping_key,
-          lwe_dimension, lwe_offset, chunk_size, keybundle_size_per_input);
+      if (use_noise_test_template) {
+        // Set up the noise-test variant of the specialized 2_2 kernel
+        check_cuda_error(cudaFuncSetAttribute(
+            device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
+                Torus, Degree<2048>, FULLSM, true>,
+            cudaFuncAttributeMaxDynamicSharedMemorySize,
+            3 * full_sm_keybundle));
+        check_cuda_error(cudaFuncSetCacheConfig(
+            device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
+                Torus, Degree<2048>, FULLSM, true>,
+            cudaFuncCachePreferShared));
+        check_cuda_error(cudaGetLastError());
+        device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
+            Torus, Degree<2048>, FULLSM, true>
+            <<<grid_keybundle, thds_new_keybundle, 3 * full_sm_keybundle,
+               stream>>>(lwe_array_in, lwe_input_indexes, keybundle_fft,
+                         bootstrapping_key, lwe_dimension, lwe_offset,
+                         chunk_size, keybundle_size_per_input);
+      } else {
+        check_cuda_error(cudaFuncSetAttribute(
+            device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
+                Torus, Degree<2048>, FULLSM>,
+            cudaFuncAttributeMaxDynamicSharedMemorySize,
+            3 * full_sm_keybundle));
+        check_cuda_error(cudaFuncSetCacheConfig(
+            device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
+                Torus, Degree<2048>, FULLSM>,
+            cudaFuncCachePreferShared));
+        check_cuda_error(cudaGetLastError());
+        device_multi_bit_programmable_bootstrap_keybundle_2_2_params<
+            Torus, Degree<2048>, FULLSM><<<grid_keybundle, thds_new_keybundle,
+                                           3 * full_sm_keybundle, stream>>>(
+            lwe_array_in, lwe_input_indexes, keybundle_fft, bootstrapping_key,
+            lwe_dimension, lwe_offset, chunk_size, keybundle_size_per_input);
+      }
     } else {
-      device_multi_bit_programmable_bootstrap_keybundle<Torus, params, FULLSM>
-          <<<grid_keybundle, thds, full_sm_keybundle, stream>>>(
-              lwe_array_in, lwe_input_indexes, keybundle_fft, bootstrapping_key,
-              lwe_dimension, glwe_dimension, polynomial_size, grouping_factor,
-              level_count, lwe_offset, chunk_size, keybundle_size_per_input,
-              d_mem, 0);
+      if (use_noise_test_template) {
+        device_multi_bit_programmable_bootstrap_keybundle<Torus, params, FULLSM,
+                                                          true>
+            <<<grid_keybundle, thds, full_sm_keybundle, stream>>>(
+                lwe_array_in, lwe_input_indexes, keybundle_fft,
+                bootstrapping_key, lwe_dimension, glwe_dimension,
+                polynomial_size, grouping_factor, level_count, lwe_offset,
+                chunk_size, keybundle_size_per_input, d_mem, 0);
+      } else {
+        device_multi_bit_programmable_bootstrap_keybundle<Torus, params, FULLSM>
+            <<<grid_keybundle, thds, full_sm_keybundle, stream>>>(
+                lwe_array_in, lwe_input_indexes, keybundle_fft,
+                bootstrapping_key, lwe_dimension, glwe_dimension,
+                polynomial_size, grouping_factor, level_count, lwe_offset,
+                chunk_size, keybundle_size_per_input, d_mem, 0);
+      }
     }
   }
   check_cuda_error(cudaGetLastError());
@@ -796,6 +891,20 @@ __host__ void execute_compute_keybundle_2_2_specialized(
       grouping_factor, level_count, lwe_offset,
       MultiBitKeybundleLaunchMode::SPECIALIZED_2_2);
 }
+// Used only to run noise tests
+template <typename Torus, class params>
+__host__ void execute_compute_keybundle_noise_tests(
+    cudaStream_t stream, uint32_t gpu_index, Torus const *lwe_array_in,
+    Torus const *lwe_input_indexes, Torus const *bootstrapping_key,
+    pbs_buffer<Torus, MULTI_BIT> *buffer, uint32_t num_samples,
+    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t grouping_factor, uint32_t level_count, uint32_t lwe_offset) {
+  execute_compute_keybundle_with_mode<Torus, params>(
+      stream, gpu_index, lwe_array_in, lwe_input_indexes, bootstrapping_key,
+      buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+      grouping_factor, level_count, lwe_offset,
+      MultiBitKeybundleLaunchMode::NOISE_TESTS);
+}
 
 template <typename Torus, class params, bool is_first_iter>
 __host__ void execute_step_one(
@@ -955,4 +1064,62 @@ __host__ void host_multi_bit_programmable_bootstrap(
     }
   }
 }
+
+template <typename Torus, class params>
+__host__ void host_multi_bit_programmable_bootstrap_noise_tests(
+    cudaStream_t stream, uint32_t gpu_index, Torus *lwe_array_out,
+    Torus const *lwe_output_indexes, Torus const *lut_vector,
+    Torus const *lut_vector_indexes, Torus const *lwe_array_in,
+    Torus const *lwe_input_indexes, Torus const *bootstrapping_key,
+    pbs_buffer<Torus, MULTI_BIT> *buffer, uint32_t glwe_dimension,
+    uint32_t lwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride) {
+
+  auto lwe_chunk_size = buffer->lwe_chunk_size;
+
+  for (uint32_t lwe_offset = 0; lwe_offset < (lwe_dimension / grouping_factor);
+       lwe_offset += lwe_chunk_size) {
+
+    // Compute a keybundle with NOISE_TESTS mode to enable the specialized
+    // runs_noise_test=true kernel variant for noise measurement
+    execute_compute_keybundle_with_mode<Torus, params>(
+        stream, gpu_index, lwe_array_in, lwe_input_indexes, bootstrapping_key,
+        buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, level_count, lwe_offset,
+        MultiBitKeybundleLaunchMode::NOISE_TESTS);
+    // Accumulate (same as standard path)
+    uint32_t chunk_size =
+        std::min((uint32_t)lwe_chunk_size,
+                 (lwe_dimension / grouping_factor) - lwe_offset);
+    for (uint32_t j = 0; j < chunk_size; j++) {
+      bool is_first_iter = (j + lwe_offset) == 0;
+      bool is_last_iter =
+          (j + lwe_offset) + 1 == (lwe_dimension / grouping_factor);
+      if (is_first_iter) {
+        execute_step_one<Torus, params, true>(
+            stream, gpu_index, lut_vector, lut_vector_indexes, lwe_array_in,
+            lwe_input_indexes, buffer, num_samples, lwe_dimension,
+            glwe_dimension, polynomial_size, base_log, level_count);
+      } else {
+        execute_step_one<Torus, params, false>(
+            stream, gpu_index, lut_vector, lut_vector_indexes, lwe_array_in,
+            lwe_input_indexes, buffer, num_samples, lwe_dimension,
+            glwe_dimension, polynomial_size, base_log, level_count);
+      }
+
+      if (is_last_iter) {
+        execute_step_two<Torus, params, true>(
+            stream, gpu_index, lwe_array_out, lwe_output_indexes, buffer,
+            num_samples, glwe_dimension, polynomial_size, level_count, j,
+            num_many_lut, lut_stride);
+      } else {
+        execute_step_two<Torus, params, false>(
+            stream, gpu_index, lwe_array_out, lwe_output_indexes, buffer,
+            num_samples, glwe_dimension, polynomial_size, level_count, j,
+            num_many_lut, lut_stride);
+      }
+    }
+  }
+}
 #endif // MULTIBIT_PBS_H

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_multibit_128.cu

@@ -293,6 +293,81 @@ void cleanup_cuda_multi_bit_programmable_bootstrap_128(void *stream,
   *buffer = nullptr;
 }
 
+// Noise-tests-namespaced wrappers: delegate to the standard scratch/cleanup so
+// that callers using the noise-tests PBS128 variant have a consistent API.
+uint64_t scratch_cuda_multi_bit_programmable_bootstrap_noise_tests_128_async(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory) {
+  return scratch_cuda_multi_bit_programmable_bootstrap_128_async(
+      stream, gpu_index, pbs_buffer, glwe_dimension, polynomial_size,
+      level_count, input_lwe_ciphertext_count, allocate_gpu_memory);
+}
+
+void cleanup_cuda_multi_bit_programmable_bootstrap_noise_tests_128(
+    void *stream, uint32_t gpu_index, int8_t **pbs_buffer) {
+  cleanup_cuda_multi_bit_programmable_bootstrap_128(stream, gpu_index,
+                                                    pbs_buffer);
+  cuda_synchronize_stream(static_cast<cudaStream_t>(stream), gpu_index);
+}
+
+// Noise tests variant of the 128-bit multi-bit PBS, restricted to
+// polynomial_size=2048. The input is assumed to contain precomputed
+// modswitched values in the extended input array layout.
+void cuda_multi_bit_programmable_bootstrap_noise_tests_128_async(
+    void *stream, uint32_t gpu_index, void *lwe_array_out,
+    void const *lwe_output_indexes, void const *lut_vector,
+    void const *lwe_array_in, void const *lwe_input_indexes,
+    void const *bootstrapping_key, int8_t *mem_ptr, uint32_t lwe_dimension,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride) {
+
+  PANIC_IF_FALSE(num_samples == 1,
+                 "Cuda error (multi-bit PBS): num_samples (%d) should be 1",
+                 num_samples);
+  PANIC_IF_FALSE(base_log <= 64,
+                 "Cuda error (multi-bit PBS): base log (%d) should be <= 64",
+                 base_log);
+  PANIC_IF_FALSE(polynomial_size == 2048,
+                 "Cuda error (multi-bit PBS128 noise tests): only polynomial "
+                 "size 2048 is supported, got %d.",
+                 polynomial_size);
+
+  auto *buffer =
+      reinterpret_cast<pbs_buffer_128<uint64_t, MULTI_BIT> *>(mem_ptr);
+  switch (buffer->pbs_variant) {
+  case PBS_VARIANT::CG:
+    host_cg_multi_bit_programmable_bootstrap_noise_tests_128<uint64_t,
+                                                             Degree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index,
+        static_cast<__uint128_t *>(lwe_array_out),
+        static_cast<const uint64_t *>(lwe_output_indexes),
+        static_cast<const __uint128_t *>(lut_vector),
+        static_cast<const uint64_t *>(lwe_array_in),
+        static_cast<const uint64_t *>(lwe_input_indexes),
+        static_cast<const __uint128_t *>(bootstrapping_key), buffer,
+        glwe_dimension, lwe_dimension, polynomial_size, grouping_factor,
+        base_log, level_count, num_samples, num_many_lut, lut_stride);
+    break;
+  case PBS_VARIANT::DEFAULT:
+    host_multi_bit_programmable_bootstrap_noise_tests_128<uint64_t,
+                                                          Degree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index,
+        static_cast<__uint128_t *>(lwe_array_out),
+        static_cast<const uint64_t *>(lwe_output_indexes),
+        static_cast<const __uint128_t *>(lut_vector),
+        static_cast<const uint64_t *>(lwe_array_in),
+        static_cast<const uint64_t *>(lwe_input_indexes),
+        static_cast<const __uint128_t *>(bootstrapping_key), buffer,
+        glwe_dimension, lwe_dimension, polynomial_size, grouping_factor,
+        base_log, level_count, num_samples, num_many_lut, lut_stride);
+    break;
+  default:
+    PANIC("Cuda error (multi-bit PBS): unsupported implementation variant.")
+  }
+}
+
 /**
  * Computes divisors of the product of num_sms (streaming multiprocessors on the
  * GPU) and max_blocks_per_sm (maximum active blocks per SM to launch

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_multibit_128.cuh

@@ -18,7 +18,8 @@ uint64_t get_buffer_size_full_sm_multibit_programmable_bootstrap_128_keybundle(
                                       (size_t)2); // accumulator
 }
 
-template <typename InputTorus, class params, sharedMemDegree SMD>
+template <typename InputTorus, class params, sharedMemDegree SMD,
+          bool runs_noise_test = false>
 __global__ void device_multi_bit_programmable_bootstrap_keybundle_128(
     const InputTorus *__restrict__ lwe_array_in,
     const InputTorus *__restrict__ lwe_input_indexes, double *keybundle_array,
@@ -80,11 +81,35 @@ __global__ void device_multi_bit_programmable_bootstrap_keybundle_128(
     // Precalculate the monomial degrees and store them in shared memory
     uint32_t *monomial_degrees = (uint32_t *)selected_memory;
     if (threadIdx.x < (1 << grouping_factor)) {
-      auto lwe_array_group =
-          block_lwe_array_in + rev_lwe_iteration * grouping_factor;
-      monomial_degrees[threadIdx.x] =
-          calculates_monomial_degree<InputTorus, params>(
-              lwe_array_group, threadIdx.x, grouping_factor);
+      if constexpr (runs_noise_test == true) {
+        // For noise tests the input array contains the input lwe but also the
+        // modswitched results. This allows to avoid changing the accumulation
+        // kernel for the noise tests since the input body will stay in the same
+        // position. The layout of the input array is the following:
+        // | input lwe     | modswitched inputs       |
+        // | lwe size      | lwe_size*grouping_factor |
+
+        // This offset allows to jump directly to the modswitched inputs,
+        // skipping the input lwe
+        const InputTorus modswitched_offset = lwe_dimension + 1;
+
+        const InputTorus *block_lwe_array_in_noise =
+            &lwe_array_in[lwe_input_indexes[input_idx] *
+                              (lwe_dimension / grouping_factor) *
+                              (1 << grouping_factor) +
+                          modswitched_offset];
+
+        const InputTorus *lwe_array_group =
+            block_lwe_array_in_noise +
+            rev_lwe_iteration * (1 << grouping_factor);
+        monomial_degrees[threadIdx.x] = lwe_array_group[threadIdx.x];
+      } else {
+        auto lwe_array_group =
+            block_lwe_array_in + rev_lwe_iteration * grouping_factor;
+        monomial_degrees[threadIdx.x] =
+            calculates_monomial_degree<InputTorus, params>(
+                lwe_array_group, threadIdx.x, grouping_factor);
+      }
     }
     __syncthreads();
 
@@ -588,6 +613,74 @@ __host__ void execute_compute_keybundle_128(
   check_cuda_error(cudaGetLastError());
 }
 
+// Used only to run noise tests: launches the keybundle kernel with the
+// runs_noise_test=true variant, which reads modswitched inputs from the
+// extended input array layout instead of computing them on-the-fly
+template <typename InputTorus, class params>
+__host__ void execute_compute_keybundle_noise_tests_128(
+    cudaStream_t stream, uint32_t gpu_index, InputTorus const *lwe_array_in,
+    InputTorus const *lwe_input_indexes, __uint128_t const *bootstrapping_key,
+    pbs_buffer_128<InputTorus, MULTI_BIT> *buffer, uint32_t num_samples,
+    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t grouping_factor, uint32_t level_count, uint32_t lwe_offset) {
+  cuda_set_device(gpu_index);
+
+  auto lwe_chunk_size = buffer->lwe_chunk_size;
+  uint64_t chunk_size = std::min(
+      lwe_chunk_size, (uint64_t)(lwe_dimension / grouping_factor) - lwe_offset);
+
+  uint64_t keybundle_size_per_input =
+      lwe_chunk_size * level_count * (glwe_dimension + 1) *
+      (glwe_dimension + 1) * (polynomial_size / 2) * 4;
+
+  uint64_t full_sm_keybundle =
+      get_buffer_size_full_sm_multibit_programmable_bootstrap_128_keybundle<
+          __uint128_t>(polynomial_size);
+  auto max_shared_memory = cuda_get_max_shared_memory(gpu_index);
+
+  auto d_mem = buffer->d_mem_keybundle;
+  auto keybundle_fft = buffer->keybundle_fft;
+
+  dim3 grid_keybundle(num_samples * chunk_size,
+                      (glwe_dimension + 1) * (glwe_dimension + 1), level_count);
+  dim3 thds(polynomial_size / params::opt, 1, 1);
+
+  if (max_shared_memory < full_sm_keybundle) {
+    check_cuda_error(cudaFuncSetAttribute(
+        device_multi_bit_programmable_bootstrap_keybundle_128<
+            InputTorus, params, NOSM, true>,
+        cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
+    check_cuda_error(cudaFuncSetCacheConfig(
+        device_multi_bit_programmable_bootstrap_keybundle_128<
+            InputTorus, params, NOSM, true>,
+        cudaFuncCachePreferShared));
+    device_multi_bit_programmable_bootstrap_keybundle_128<InputTorus, params,
+                                                          NOSM, true>
+        <<<grid_keybundle, thds, 0, stream>>>(
+            lwe_array_in, lwe_input_indexes, keybundle_fft, bootstrapping_key,
+            lwe_dimension, glwe_dimension, polynomial_size, grouping_factor,
+            level_count, lwe_offset, chunk_size, keybundle_size_per_input,
+            d_mem, full_sm_keybundle);
+  } else {
+    check_cuda_error(cudaFuncSetAttribute(
+        device_multi_bit_programmable_bootstrap_keybundle_128<
+            InputTorus, params, FULLSM, true>,
+        cudaFuncAttributeMaxDynamicSharedMemorySize, full_sm_keybundle));
+    check_cuda_error(cudaFuncSetCacheConfig(
+        device_multi_bit_programmable_bootstrap_keybundle_128<
+            InputTorus, params, FULLSM, true>,
+        cudaFuncCachePreferShared));
+    device_multi_bit_programmable_bootstrap_keybundle_128<InputTorus, params,
+                                                          FULLSM, true>
+        <<<grid_keybundle, thds, full_sm_keybundle, stream>>>(
+            lwe_array_in, lwe_input_indexes, keybundle_fft, bootstrapping_key,
+            lwe_dimension, glwe_dimension, polynomial_size, grouping_factor,
+            level_count, lwe_offset, chunk_size, keybundle_size_per_input,
+            d_mem, 0);
+  }
+  check_cuda_error(cudaGetLastError());
+}
+
 template <typename InputTorus, class params, bool is_first_iter>
 __host__ void execute_step_one_128(
     cudaStream_t stream, uint32_t gpu_index, __uint128_t const *lut_vector,
@@ -1119,46 +1212,47 @@ __host__ bool verify_cuda_programmable_bootstrap_cg_multi_bit_grid_size_128(
   int max_active_blocks_per_sm;
 
   if (max_shared_memory < partial_sm_cg_accumulate) {
-    cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+    check_cuda_error(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
         &max_active_blocks_per_sm,
         (void *)device_multi_bit_programmable_bootstrap_cg_accumulate_128<
             Torus, params, NOSM>,
-        thds, 0);
+        thds, 0));
   } else if (max_shared_memory < full_sm_cg_accumulate) {
     check_cuda_error(cudaFuncSetAttribute(
         device_multi_bit_programmable_bootstrap_cg_accumulate_128<Torus, params,
                                                                   PARTIALSM>,
         cudaFuncAttributeMaxDynamicSharedMemorySize, partial_sm_cg_accumulate));
-    cudaFuncSetCacheConfig(
+    check_cuda_error(cudaFuncSetCacheConfig(
         device_multi_bit_programmable_bootstrap_cg_accumulate_128<Torus, params,
                                                                   PARTIALSM>,
-        cudaFuncCachePreferShared);
-    cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+        cudaFuncCachePreferShared));
+    check_cuda_error(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
         &max_active_blocks_per_sm,
         (void *)device_multi_bit_programmable_bootstrap_cg_accumulate_128<
             Torus, params, PARTIALSM>,
-        thds, partial_sm_cg_accumulate);
+        thds, partial_sm_cg_accumulate));
     check_cuda_error(cudaGetLastError());
   } else {
     check_cuda_error(cudaFuncSetAttribute(
         device_multi_bit_programmable_bootstrap_cg_accumulate_128<Torus, params,
                                                                   FULLSM>,
         cudaFuncAttributeMaxDynamicSharedMemorySize, full_sm_cg_accumulate));
-    cudaFuncSetCacheConfig(
+    check_cuda_error(cudaFuncSetCacheConfig(
         device_multi_bit_programmable_bootstrap_cg_accumulate_128<Torus, params,
                                                                   FULLSM>,
-        cudaFuncCachePreferShared);
-    cudaOccupancyMaxActiveBlocksPerMultiprocessor(
+        cudaFuncCachePreferShared));
+    check_cuda_error(cudaOccupancyMaxActiveBlocksPerMultiprocessor(
         &max_active_blocks_per_sm,
         (void *)device_multi_bit_programmable_bootstrap_cg_accumulate_128<
             Torus, params, FULLSM>,
-        thds, full_sm_cg_accumulate);
+        thds, full_sm_cg_accumulate));
     check_cuda_error(cudaGetLastError());
   }
 
   // Get the number of streaming multiprocessors
   int number_of_sm = 0;
-  cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0);
+  check_cuda_error(
+      cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0));
   return number_of_blocks <= max_active_blocks_per_sm * number_of_sm;
 }
 
@@ -1199,4 +1293,96 @@ supports_cooperative_groups_on_multibit_programmable_bootstrap_128(
   }
 }
 
+// Noise tests variant: identical to
+// host_cg_multi_bit_programmable_bootstrap_128 but uses the noise-test
+// keybundle (runs_noise_test=true) instead of the standard one.
+template <typename InputTorus, class params>
+__host__ void host_cg_multi_bit_programmable_bootstrap_noise_tests_128(
+    cudaStream_t stream, uint32_t gpu_index, __uint128_t *lwe_array_out,
+    InputTorus const *lwe_output_indexes, __uint128_t const *lut_vector,
+    InputTorus const *lwe_array_in, InputTorus const *lwe_input_indexes,
+    __uint128_t const *bootstrapping_key,
+    pbs_buffer_128<InputTorus, MULTI_BIT> *buffer, uint32_t glwe_dimension,
+    uint32_t lwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride) {
+
+  auto lwe_chunk_size = buffer->lwe_chunk_size;
+
+  for (uint32_t lwe_offset = 0; lwe_offset < (lwe_dimension / grouping_factor);
+       lwe_offset += lwe_chunk_size) {
+
+    // Compute a keybundle with the noise-test kernel variant
+    // (runs_noise_test=true) to read precomputed modswitched values
+    execute_compute_keybundle_noise_tests_128<InputTorus, params>(
+        stream, gpu_index, lwe_array_in, lwe_input_indexes, bootstrapping_key,
+        buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, level_count, lwe_offset);
+
+    execute_cg_external_product_loop_128<InputTorus, params>(
+        stream, gpu_index, lut_vector, lwe_array_in, lwe_input_indexes,
+        lwe_array_out, lwe_output_indexes, buffer, num_samples, lwe_dimension,
+        glwe_dimension, polynomial_size, grouping_factor, base_log, level_count,
+        lwe_offset, num_many_lut, lut_stride);
+  }
+}
+
+template <typename InputTorus, class params>
+__host__ void host_multi_bit_programmable_bootstrap_noise_tests_128(
+    cudaStream_t stream, uint32_t gpu_index, __uint128_t *lwe_array_out,
+    InputTorus const *lwe_output_indexes, __uint128_t const *lut_vector,
+    InputTorus const *lwe_array_in, InputTorus const *lwe_input_indexes,
+    __uint128_t const *bootstrapping_key,
+    pbs_buffer_128<InputTorus, MULTI_BIT> *buffer, uint32_t glwe_dimension,
+    uint32_t lwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride) {
+
+  auto lwe_chunk_size = buffer->lwe_chunk_size;
+
+  for (uint32_t lwe_offset = 0; lwe_offset < (lwe_dimension / grouping_factor);
+       lwe_offset += lwe_chunk_size) {
+
+    // Compute a keybundle with the noise-test kernel variant
+    // (runs_noise_test=true) to read precomputed modswitched values
+    execute_compute_keybundle_noise_tests_128<InputTorus, params>(
+        stream, gpu_index, lwe_array_in, lwe_input_indexes, bootstrapping_key,
+        buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, level_count, lwe_offset);
+
+    // Accumulate (same as standard path)
+    uint64_t chunk_size =
+        std::min((uint32_t)lwe_chunk_size,
+                 (lwe_dimension / grouping_factor) - lwe_offset);
+    for (uint32_t j = 0; j < chunk_size; j++) {
+      bool is_first_iter = (j + lwe_offset) == 0;
+      bool is_last_iter =
+          (j + lwe_offset) + 1 == (lwe_dimension / grouping_factor);
+      if (is_first_iter) {
+        execute_step_one_128<InputTorus, params, true>(
+            stream, gpu_index, lut_vector, lwe_array_in, lwe_input_indexes,
+            buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+            base_log, level_count);
+      } else {
+        execute_step_one_128<InputTorus, params, false>(
+            stream, gpu_index, lut_vector, lwe_array_in, lwe_input_indexes,
+            buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+            base_log, level_count);
+      }
+
+      if (is_last_iter) {
+        execute_step_two_128<InputTorus, params, true>(
+            stream, gpu_index, lwe_array_out, lwe_output_indexes, buffer,
+            num_samples, glwe_dimension, polynomial_size, level_count, j,
+            num_many_lut, lut_stride);
+      } else {
+        execute_step_two_128<InputTorus, params, false>(
+            stream, gpu_index, lwe_array_out, lwe_output_indexes, buffer,
+            num_samples, glwe_dimension, polynomial_size, level_count, j,
+            num_many_lut, lut_stride);
+      }
+    }
+  }
+}
+
 #endif // PROGRAMMABLE_BOOTSTRAP_MULTIBIT_128_CUH

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_tbc_classic.cuh

@@ -739,7 +739,8 @@ __host__ bool verify_cuda_programmable_bootstrap_tbc_grid_size(
 
   // Get the number of streaming multiprocessors
   int number_of_sm = 0;
-  cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0);
+  check_cuda_error(
+      cudaDeviceGetAttribute(&number_of_sm, cudaDevAttrMultiProcessorCount, 0));
   return number_of_blocks <= max_active_blocks_per_sm * number_of_sm;
 }
 

backends/tfhe-cuda-backend/cuda/src/pbs/programmable_bootstrap_tbc_multibit.cuh

@@ -795,6 +795,40 @@ __host__ void host_tbc_multi_bit_programmable_bootstrap_2_2_specialized(
       MultiBitTbcLaunchMode::SPECIALIZED_2_2);
 }
 
+// Noise tests variant: uses NOISE_TESTS keybundle mode for the keybundle step
+// while keeping the standard AUTO accumulate behaviour for the TBC loop.
+template <typename Torus, class params>
+__host__ void host_tbc_multi_bit_programmable_bootstrap_noise_tests(
+    cudaStream_t stream, uint32_t gpu_index, Torus *lwe_array_out,
+    Torus const *lwe_output_indexes, Torus const *lut_vector,
+    Torus const *lut_vector_indexes, Torus const *lwe_array_in,
+    Torus const *lwe_input_indexes, Torus const *bootstrapping_key,
+    pbs_buffer<Torus, MULTI_BIT> *buffer, uint32_t glwe_dimension,
+    uint32_t lwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
+    uint32_t num_many_lut, uint32_t lut_stride) {
+  cuda_set_device(gpu_index);
+
+  auto lwe_chunk_size = buffer->lwe_chunk_size;
+  for (uint32_t lwe_offset = 0; lwe_offset < (lwe_dimension / grouping_factor);
+       lwe_offset += lwe_chunk_size) {
+
+    // Keybundle with NOISE_TESTS mode; the TBC accumulate uses AUTO as usual
+    execute_compute_keybundle_noise_tests<Torus, params>(
+        stream, gpu_index, lwe_array_in, lwe_input_indexes, bootstrapping_key,
+        buffer, num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, level_count, lwe_offset);
+
+    // Accumulate (unchanged from standard TBC path)
+    execute_tbc_external_product_loop<Torus, params>(
+        stream, gpu_index, lut_vector, lut_vector_indexes, lwe_array_in,
+        lwe_input_indexes, lwe_array_out, lwe_output_indexes, buffer,
+        num_samples, lwe_dimension, glwe_dimension, polynomial_size,
+        grouping_factor, base_log, level_count, lwe_offset, num_many_lut,
+        lut_stride, MultiBitTbcLaunchMode::AUTO);
+  }
+}
+
 template <typename Torus>
 bool supports_distributed_shared_memory_on_multibit_programmable_bootstrap(
     uint32_t polynomial_size, uint32_t max_shared_memory) {