chores(docs): minor corrections to the doc

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/actions/gpu_setup/action.yml

@@ -82,11 +82,10 @@ runs:
         sudo apt update
         sudo apt -y install cuda-toolkit-"${TOOLKIT_VERSION}"
 
-# Command to put back in once nvcc check is fixed
-#        find /usr/local -executable -name "nvcc"
     - name: Export CUDA variables
       shell: bash
       run: |
+        find /usr/local -executable -name "nvcc"
         CUDA_PATH=/usr/local/cuda-"${CUDA_VERSION}"
         {
           echo "CUDA_PATH=$CUDA_PATH";

.github/workflows/aws_data_tests.yml

@@ -54,7 +54,7 @@ jobs:
 
       - 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
@@ -89,7 +89,7 @@ jobs:
       - 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

.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

@@ -6,6 +6,9 @@ name: backward_compat_pr_change_report
 on:
   pull_request:
 
+env:
+  CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
+
 permissions:
   contents: read
 
@@ -14,9 +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:

.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
@@ -89,10 +93,9 @@ jobs:
     uses: ./.github/workflows/benchmark_gpu_common.yml
     if: inputs.run-gpu-integer-benchmarks
     with:
-      backend: terraform
-      profile: scaleway-multi-h100-sxm5
+      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
@@ -104,14 +107,14 @@ jobs:
       JOB_SECRET: ${{ secrets.JOB_SECRET }}
       SLAB_ACTION_TOKEN: ${{ secrets.SLAB_ACTION_TOKEN }}
       SLAB_URL: ${{ secrets.SLAB_URL }}
-      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL_PRE_PROD }}
+      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL }}
 
   run-benchmarks-hpu-integer:
     name: benchmark_documentation/run-benchmarks-hpu-integer
     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
@@ -151,8 +154,7 @@ jobs:
     uses: ./.github/workflows/benchmark_gpu_common.yml
     if: inputs.run-gpu-core-crypto-benchmarks
     with:
-      backend: terraform
-      profile: scaleway-multi-h100-sxm5
+      profile: multi-h100-sxm5
       hardware_name: n3-H100-SXM5x8
       command: pbs, ks_pbs
       bench_type: latency
@@ -165,23 +167,44 @@ jobs:
       JOB_SECRET: ${{ secrets.JOB_SECRET }}
       SLAB_ACTION_TOKEN: ${{ secrets.SLAB_ACTION_TOKEN }}
       SLAB_URL: ${{ secrets.SLAB_URL }}
-      SLAB_BASE_URL: ${{ secrets.SLAB_BASE_URL_PRE_PROD }}
+      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 }}
@@ -190,7 +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 +31,13 @@ on:
           - pbs128
           - ks
           - ks_pbs
+          - tfhe_zk_pok
+          - msm_zk
           - integer_zk
+          - integer_zk_experimental
           - integer_aes
           - integer_aes256
-          - hlapi_erc20
+          - hlapi_erc7984
           - hlapi_dex
           - hlapi_noise_squash
       op_flavor:
@@ -120,8 +123,8 @@ 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 }}
@@ -333,13 +307,13 @@ 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
           {

.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_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

@@ -50,7 +50,7 @@ jobs:
           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

@@ -73,18 +73,12 @@ jobs:
       (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == '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 +87,6 @@ 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 +119,6 @@ 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 +162,14 @@ 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

@@ -74,18 +74,12 @@ jobs:
       (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == '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 +88,6 @@ 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 +120,6 @@ 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 +170,14 @@ 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

@@ -38,7 +38,7 @@ 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 }}
@@ -112,7 +112,7 @@ 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 }}
@@ -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

@@ -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 }}
@@ -169,7 +169,7 @@ 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

@@ -75,18 +75,12 @@ jobs:
       (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == '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 +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
@@ -134,7 +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 +162,14 @@ 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

@@ -81,7 +81,7 @@ 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 +178,7 @@ 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

@@ -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 }}
@@ -169,7 +169,7 @@ 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

@@ -75,18 +75,12 @@ jobs:
       (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == '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 +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
@@ -134,7 +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 +162,14 @@ 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

@@ -81,7 +81,7 @@ 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 +178,7 @@ 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 +51,16 @@ jobs:
         with:
           files_yaml: |
             gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
               - backends/zk-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/src/shortint/parameters/**
+              - tfhe/src/zk/**
+              - tfhe-zk-pok/**
+              - 'tfhe/docs/**/**.md'
               - '.github/workflows/gpu_zk_tests.yml'
               - ci/slab.toml
 
@@ -67,7 +76,7 @@ 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 +135,9 @@ 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 +170,7 @@ 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

@@ -107,7 +107,7 @@ jobs:
           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}";
@@ -177,22 +199,23 @@ jobs:
       - name: Download artifact
         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 }}

.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__

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

@@ -122,6 +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 || \
@@ -350,23 +356,23 @@ 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
 
@@ -380,7 +386,7 @@ 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
 
@@ -473,7 +479,7 @@ 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
@@ -484,11 +490,17 @@ 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.
@@ -529,6 +541,15 @@ 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 \
@@ -536,6 +557,11 @@ 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 \
@@ -566,10 +592,12 @@ 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
@@ -666,7 +694,7 @@ 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
@@ -675,11 +703,14 @@ 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
@@ -765,7 +796,7 @@ 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:
@@ -1201,12 +1232,31 @@ 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 $(CARGO_PROFILE) \
+		--features=integer,zk-pok,gpu,gpu-experimental-zk -p tfhe -- \
+		integer::gpu::zk::
+
+.PHONY: test_zk_cuda # Run all GPU MSM integration tests (CPU vs GPU comparison + integration test)
+test_zk_cuda: test_zk_cuda_backend test_zk_pok_experimental_gpu test_integer_zk_gpu test_integer_zk_experimental_gpu
+
 .PHONY: test_zk_wasm_x86_compat_ci
 test_zk_wasm_x86_compat_ci: check_nvm_installed
 	source ~/.nvm/nvm.sh && \
@@ -1225,6 +1275,11 @@ test_versionable:
 	RUSTFLAGS="$(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:
@@ -1359,6 +1414,19 @@ 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 \
@@ -1401,6 +1469,38 @@ 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
@@ -1564,27 +1664,50 @@ bench_integer_rerand_gpu: install_rs_check_toolchain
 	--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
@@ -1748,37 +1871,37 @@ 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
@@ -1811,25 +1934,25 @@ 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_dex # Run benchmarks for DEX operations
@@ -1853,13 +1976,13 @@ 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
@@ -1867,6 +1990,13 @@ 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) \
@@ -1908,10 +2038,10 @@ 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
@@ -1953,10 +2083,10 @@ 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
+	# 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,gpu,internal-keycache -p tfhe-benchmark --profile release_lto_off -- '::transfer::overflow'
 
 	# DEX
@@ -2143,8 +2273,10 @@ 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/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

@@ -150,3 +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/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) {
+  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/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 +293,80 @@ 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);
+}
+
+// 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) {

backends/tfhe-cuda-backend/cuda/src/zk/zk.cuh

@@ -119,71 +119,73 @@ __host__ void host_expand_without_verification(
       streams.stream(0), streams.gpu_index(0), true);
 
   if (mem_ptr->expand_kind == EXPAND_KIND::NO_CASTING) {
+    // This path is added to mimic the CPU fallback behaviour for the no_casting
+    // expand, which is needed for the noise sanity checks.
     host_lwe_expand<Torus, params>(streams.stream(0), streams.gpu_index(0),
                                    lwe_array_out, d_expand_jobs, num_lwes);
-    return;
-  }
 
-  host_lwe_expand<Torus, params>(streams.stream(0), streams.gpu_index(0),
-                                 expanded_lwes, d_expand_jobs, num_lwes);
+  } else {
+    // This is our default path for the expand with casting if needed.
+    host_lwe_expand<Torus, params>(streams.stream(0), streams.gpu_index(0),
+                                   expanded_lwes, d_expand_jobs, num_lwes);
 
-  auto lwe_array_input = expanded_lwes;
-  auto ksks = casting_keys;
-  auto message_and_carry_extract_luts = mem_ptr->message_and_carry_extract_luts;
+    auto lwe_array_input = expanded_lwes;
+    auto ksks = casting_keys;
+    auto message_and_carry_extract_luts =
+        mem_ptr->message_and_carry_extract_luts;
 
-  auto lut = mem_ptr->message_and_carry_extract_luts;
-  if (casting_key_type == SMALL_TO_BIG) {
-    if (mem_ptr->expand_kind == EXPAND_KIND::SANITY_CHECK) {
-      PANIC("SANITY_CHECK not supported for SMALL_TO_BIG casting");
+    auto lut = mem_ptr->message_and_carry_extract_luts;
+    if (casting_key_type == SMALL_TO_BIG) {
+      if (mem_ptr->expand_kind == EXPAND_KIND::SANITY_CHECK) {
+        PANIC("SANITY_CHECK not supported for SMALL_TO_BIG casting");
+      }
+      // Keyswitch from small to big key if needed
+      auto ksed_small_to_big_expanded_lwes =
+          mem_ptr->tmp_ksed_small_to_big_expanded_lwes;
+      std::vector<Torus *> lwe_trivial_indexes_vec =
+          lut->lwe_trivial_indexes_vec;
+
+      auto casting_params = mem_ptr->casting_params;
+      auto casting_output_dimension = casting_params.big_lwe_dimension;
+      auto casting_input_dimension = casting_params.small_lwe_dimension;
+      auto casting_ks_level = casting_params.ks_level;
+      auto casting_ks_base_log = casting_params.ks_base_log;
+
+      // apply keyswitch to BIG
+      execute_keyswitch_async<Torus>(
+          streams.get_ith(0), ksed_small_to_big_expanded_lwes,
+          lwe_trivial_indexes_vec[0], expanded_lwes, lwe_trivial_indexes_vec[0],
+          casting_keys, casting_input_dimension, casting_output_dimension,
+          casting_ks_base_log, casting_ks_level, num_lwes,
+          lut->using_trivial_lwe_indexes, lut->ks_tmp_buf_vec);
+
+      // In this case, the next keyswitch will use the compute ksk
+      ksks = compute_ksks;
+      lwe_array_input = ksed_small_to_big_expanded_lwes;
     }
-    // Keyswitch from small to big key if needed
-    auto ksed_small_to_big_expanded_lwes =
-        mem_ptr->tmp_ksed_small_to_big_expanded_lwes;
-    std::vector<Torus *> lwe_trivial_indexes_vec = lut->lwe_trivial_indexes_vec;
 
-    auto casting_params = mem_ptr->casting_params;
-    auto casting_output_dimension = casting_params.big_lwe_dimension;
-    auto casting_input_dimension = casting_params.small_lwe_dimension;
-    auto casting_ks_level = casting_params.ks_level;
-    auto casting_ks_base_log = casting_params.ks_base_log;
+    // Apply LUT
+    cuda_memset_async(lwe_array_out, 0,
+                      safe_mul_sizeof<Torus>((size_t)(lwe_dimension + 1),
+                                             (size_t)num_lwes, (size_t)2),
+                      streams.stream(0), streams.gpu_index(0));
+    CudaRadixCiphertextFFI output;
+    into_radix_ciphertext(&output, lwe_array_out, 2 * num_lwes, lwe_dimension);
+    CudaRadixCiphertextFFI input;
+    into_radix_ciphertext(&input, lwe_array_input, 2 * num_lwes, lwe_dimension);
+    // This is a special case only for our noise sanity checks
+    // If we are doing a SANITY_CHECK expand, we just apply the identity LUT
+    // This replicates the CPU fallback behaviour of the casting expand
+    auto final_lut = (mem_ptr->expand_kind == EXPAND_KIND::SANITY_CHECK
+                          ? mem_ptr->identity_lut
+                          : message_and_carry_extract_luts);
 
-    // apply keyswitch to BIG
-    execute_keyswitch_async<Torus>(
-        streams.get_ith(0), ksed_small_to_big_expanded_lwes,
-        lwe_trivial_indexes_vec[0], expanded_lwes, lwe_trivial_indexes_vec[0],
-        casting_keys, casting_input_dimension, casting_output_dimension,
-        casting_ks_base_log, casting_ks_level, num_lwes,
-        lut->using_trivial_lwe_indexes, lut->ks_tmp_buf_vec);
-
-    // In this case, the next keyswitch will use the compute ksk
-    ksks = compute_ksks;
-    lwe_array_input = ksed_small_to_big_expanded_lwes;
-  }
-
-  // Apply LUT
-  cuda_memset_async(lwe_array_out, 0,
-                    safe_mul_sizeof<Torus>((size_t)(lwe_dimension + 1),
-                                           (size_t)num_lwes, (size_t)2),
-                    streams.stream(0), streams.gpu_index(0));
-  CudaRadixCiphertextFFI output;
-  into_radix_ciphertext(&output, lwe_array_out, 2 * num_lwes, lwe_dimension);
-  CudaRadixCiphertextFFI input;
-  into_radix_ciphertext(&input, lwe_array_input, 2 * num_lwes, lwe_dimension);
-  // This is a special case only for our noise sanity checks
-  // If we are doing a SANITY_CHECK expand, we just apply the identity LUT
-  // This replicates the CPU fallback behaviour of the casting expand
-  if (mem_ptr->expand_kind == EXPAND_KIND::SANITY_CHECK) {
     integer_radix_apply_univariate_lookup_table<Torus>(
-        streams, &output, &input, bsks, ksks, mem_ptr->identity_lut,
-        2 * num_lwes);
-    return;
-  }
+        streams, &output, &input, bsks, ksks, final_lut, 2 * num_lwes);
 
-  integer_radix_apply_univariate_lookup_table<Torus>(
-      streams, &output, &input, bsks, ksks, message_and_carry_extract_luts,
-      2 * num_lwes);
-  release_cpu_radix_ciphertext_async(&input);
-  release_cpu_radix_ciphertext_async(&output);
+    release_cpu_radix_ciphertext_async(&input);
+    release_cpu_radix_ciphertext_async(&output);
+  }
   compact_lwe_lists.release();
 }
 

backends/tfhe-cuda-backend/src/bindings.rs

@@ -79,6 +79,30 @@ unsafe extern "C" {
         polynomial_size: u32,
     );
 }
+unsafe extern "C" {
+    pub fn cuda_modulus_switch_multi_bit_64_async(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        lwe_array_out: *mut ffi::c_void,
+        lwe_array_in: *mut ffi::c_void,
+        size: u32,
+        log_modulus: u32,
+        degree: u32,
+        grouping_factor: u32,
+    );
+}
+unsafe extern "C" {
+    pub fn cuda_modulus_switch_multi_bit_128_async(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        lwe_array_out: *mut ffi::c_void,
+        lwe_array_in: *mut ffi::c_void,
+        size: u32,
+        log_modulus: u32,
+        degree: u32,
+        grouping_factor: u32,
+    );
+}
 pub const PBS_TYPE_MULTI_BIT: PBS_TYPE = 0;
 pub const PBS_TYPE_CLASSICAL: PBS_TYPE = 1;
 pub type PBS_TYPE = ffi::c_uint;
@@ -112,9 +136,6 @@ pub type Direction = ffi::c_uint;
 pub const BitValue_Zero: BitValue = 0;
 pub const BitValue_One: BitValue = 1;
 pub type BitValue = ffi::c_uint;
-pub const RERAND_MODE_RERAND_WITH_KS: RERAND_MODE = 0;
-pub const RERAND_MODE_RERAND_WITHOUT_KS: RERAND_MODE = 1;
-pub type RERAND_MODE = ffi::c_uint;
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
 pub struct CudaStreamsFFI {
@@ -2455,6 +2476,9 @@ unsafe extern "C" {
         glwe_index: u32,
     );
 }
+pub const RERAND_MODE_RERAND_WITH_KS: RERAND_MODE = 0;
+pub const RERAND_MODE_RERAND_WITHOUT_KS: RERAND_MODE = 1;
+pub type RERAND_MODE = ffi::c_uint;
 unsafe extern "C" {
     pub fn scratch_cuda_rerand_64_async(
         streams: CudaStreamsFFI,
@@ -2467,7 +2491,7 @@ unsafe extern "C" {
         message_modulus: u32,
         carry_modulus: u32,
         allocate_gpu_memory: bool,
-        rerand_type: u32,
+        rerand_type: RERAND_MODE,
     ) -> u64;
 }
 unsafe extern "C" {
@@ -3367,6 +3391,48 @@ unsafe extern "C" {
         pbs_buffer: *mut *mut i8,
     );
 }
+unsafe extern "C" {
+    pub fn scratch_cuda_multi_bit_programmable_bootstrap_noise_tests_64_async(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        pbs_buffer: *mut *mut i8,
+        glwe_dimension: u32,
+        polynomial_size: u32,
+        level_count: u32,
+        input_lwe_ciphertext_count: u32,
+        allocate_gpu_memory: bool,
+    ) -> u64;
+}
+unsafe extern "C" {
+    pub fn cleanup_cuda_multi_bit_programmable_bootstrap_noise_tests_64(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        pbs_buffer: *mut *mut i8,
+    );
+}
+unsafe extern "C" {
+    pub fn cuda_multi_bit_programmable_bootstrap_noise_tests_64_async(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        lwe_array_out: *mut ffi::c_void,
+        lwe_output_indexes: *const ffi::c_void,
+        lut_vector: *const ffi::c_void,
+        lut_vector_indexes: *const ffi::c_void,
+        lwe_array_in: *const ffi::c_void,
+        lwe_input_indexes: *const ffi::c_void,
+        bootstrapping_key: *const ffi::c_void,
+        buffer: *mut i8,
+        lwe_dimension: u32,
+        glwe_dimension: u32,
+        polynomial_size: u32,
+        grouping_factor: u32,
+        base_log: u32,
+        level_count: u32,
+        num_samples: u32,
+        num_many_lut: u32,
+        lut_stride: u32,
+    );
+}
 unsafe extern "C" {
     pub fn scratch_cuda_multi_bit_programmable_bootstrap_128_async(
         stream: *mut ffi::c_void,
@@ -3408,3 +3474,44 @@ unsafe extern "C" {
         buffer: *mut *mut i8,
     );
 }
+unsafe extern "C" {
+    pub fn scratch_cuda_multi_bit_programmable_bootstrap_noise_tests_128_async(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        pbs_buffer: *mut *mut i8,
+        glwe_dimension: u32,
+        polynomial_size: u32,
+        level_count: u32,
+        input_lwe_ciphertext_count: u32,
+        allocate_gpu_memory: bool,
+    ) -> u64;
+}
+unsafe extern "C" {
+    pub fn cleanup_cuda_multi_bit_programmable_bootstrap_noise_tests_128(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        pbs_buffer: *mut *mut i8,
+    );
+}
+unsafe extern "C" {
+    pub fn cuda_multi_bit_programmable_bootstrap_noise_tests_128_async(
+        stream: *mut ffi::c_void,
+        gpu_index: u32,
+        lwe_array_out: *mut ffi::c_void,
+        lwe_output_indexes: *const ffi::c_void,
+        lut_vector: *const ffi::c_void,
+        lwe_array_in: *const ffi::c_void,
+        lwe_input_indexes: *const ffi::c_void,
+        bootstrapping_key: *const ffi::c_void,
+        buffer: *mut i8,
+        lwe_dimension: u32,
+        glwe_dimension: u32,
+        polynomial_size: u32,
+        grouping_factor: u32,
+        base_log: u32,
+        level_count: u32,
+        num_samples: u32,
+        num_many_lut: u32,
+        lut_stride: u32,
+    );
+}

backends/tfhe-hpu-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-hpu-backend/README.md

@@ -297,8 +297,8 @@ source setup_hpu.sh --config v80 -p
 # Run hlapi benches
 make test_high_level_api_hpu
 
-# Run hlapi erc20 benches
-make bench_hlapi_erc20_hpu 
+# Run hlapi erc7984 benches
+make bench_hlapi_erc7984_hpu 
 
 # Run integer level benches
 make bench_integer_hpu

backends/tfhe-hpu-backend/config_store/sim/hpu_config.toml

@@ -109,7 +109,7 @@
   flush_behaviour = "Patient"
   flush = true
 
-[firmware.op_cfg.by_op.ERC_20]
+[firmware.op_cfg.by_op.ERC_7984]
   fill_batch_fifo = true
   min_batch_size = false
   use_tiers = true

backends/tfhe-hpu-backend/config_store/v80/hpu_config.toml

@@ -121,7 +121,7 @@
   flush_behaviour = "Patient"
   flush = true
 
-[firmware.op_cfg.by_op.ERC_20]
+[firmware.op_cfg.by_op.ERC_7984]
   fill_batch_fifo = true
   min_batch_size = false
   use_tiers = true

backends/tfhe-hpu-backend/src/asm/iop/mod.rs

@@ -230,7 +230,7 @@ iop!(
     [IOP_CMP -> "CMP_NEQ", opcode::CMP_NEQ],
     [IOP_CT_F_CT_BOOL -> "IF_THEN_ZERO", opcode::IF_THEN_ZERO],
     [IOP_CT_F_2CT_BOOL -> "IF_THEN_ELSE", opcode::IF_THEN_ELSE],
-    [IOP_2CT_F_3CT -> "ERC_20", opcode::ERC_20],
+    [IOP_2CT_F_3CT -> "ERC_7984", opcode::ERC_7984],
     [IOP_CT_F_CT -> "MEMCPY", opcode::MEMCPY],
     [IOP_CT_F_CT -> "ILOG2", opcode::ILOG2],
     [IOP_CT_F_CT -> "COUNT0", opcode::COUNT0],
@@ -240,5 +240,5 @@ iop!(
     [IOP_CT_F_CT -> "TRAIL0", opcode::TRAIL0],
     [IOP_CT_F_CT -> "TRAIL1", opcode::TRAIL1],
     [IOP_NCT_F_2NCT -> "ADD_SIMD", opcode::ADD_SIMD],
-    [IOP_2NCT_F_3NCT -> "ERC_20_SIMD", opcode::ERC_20_SIMD],
+    [IOP_2NCT_F_3NCT -> "ERC_7984_SIMD", opcode::ERC_7984_SIMD],
 );

backends/tfhe-hpu-backend/src/asm/iop/opcode.rs

@@ -74,9 +74,9 @@ pub const IF_THEN_ZERO: u8 = 0xCA;
 pub const IF_THEN_ELSE: u8 = 0xCB;
 
 // Custom algorithm
-// ERC20 -> Found xfer algorithm
+// ERC7984 -> Found xfer algorithm
 // 2Ct <- func(3Ct)
-pub const ERC_20: u8 = 0x80;
+pub const ERC_7984: u8 = 0x80;
 
 // Count bits
 pub const COUNT0: u8 = 0x81;
@@ -89,7 +89,7 @@ pub const TRAIL1: u8 = 0x87;
 
 // SIMD for maximum throughput
 pub const ADD_SIMD: u8 = 0xF0;
-pub const ERC_20_SIMD: u8 = 0xF1;
+pub const ERC_7984_SIMD: u8 = 0xF1;
 //
 // Utility operations
 // Used to handle real clone of ciphertext already uploaded in the Hpu memory

backends/tfhe-hpu-backend/src/fw/fw_impl/demo.rs

@@ -31,7 +31,7 @@ crate::impl_fw!("Demo" [
     IF_THEN_ZERO => fw_impl::ilp::iop_if_then_zero;
     IF_THEN_ELSE => fw_impl::ilp::iop_if_then_else;
 
-    ERC_20 => fw_impl::ilp::iop_erc_20;
+    ERC_7984 => fw_impl::ilp::iop_erc_7984;
 
     CMP_GT  => cmp_gt;
     CMP_GTE => cmp_gte;

backends/tfhe-hpu-backend/src/fw/fw_impl/ilp.rs

@@ -61,7 +61,7 @@ crate::impl_fw!("Ilp" [
     IF_THEN_ZERO => fw_impl::ilp::iop_if_then_zero;
     IF_THEN_ELSE => fw_impl::ilp::iop_if_then_else;
 
-    ERC_20 => fw_impl::ilp::iop_erc_20;
+    ERC_7984 => fw_impl::ilp::iop_erc_7984;
 
     MEMCPY => fw_impl::ilp::iop_memcpy;
 
@@ -74,7 +74,7 @@ crate::impl_fw!("Ilp" [
     TRAIL1 => fw_impl::ilp_log::iop_trail1;
     // SIMD Implementations
     ADD_SIMD     => fw_impl::llt::iop_add_simd;
-    ERC_20_SIMD  => fw_impl::llt::iop_erc_20_simd;
+    ERC_7984_SIMD  => fw_impl::llt::iop_erc_7984_simd;
 ]);
 
 #[instrument(level = "trace", skip(prog))]
@@ -1296,13 +1296,13 @@ pub fn iop_if_then_else(prog: &mut Program) {
         });
 }
 
-/// Implement erc_20 fund xfer
+/// Implement erc_7984 fund xfer
 /// Targeted algorithm is as follow:
 /// 1. Check that from has enough funds
 /// 2. Compute real_amount to xfer (i.e. amount or 0)
 /// 3. Compute new amount (from - new_amount, to + new_amount)
 #[instrument(level = "info", skip(prog))]
-pub fn iop_erc_20(prog: &mut Program) {
+pub fn iop_erc_7984(prog: &mut Program) {
     // Allocate metavariables:
     // Dest -> Operand
     let mut dst_from = prog.iop_template_var(OperandKind::Dst, 0);
@@ -1314,7 +1314,7 @@ pub fn iop_erc_20(prog: &mut Program) {
     let src_amount = prog.iop_template_var(OperandKind::Src, 2);
 
     // Add Comment header
-    prog.push_comment("ERC_20 (new_from, new_to) <- (from, to, amount)".to_string());
+    prog.push_comment("ERC_7984 (new_from, new_to) <- (from, to, amount)".to_string());
 
     let props = prog.params();
     let tfhe_params: asm::DigitParameters = props.clone().into();

backends/tfhe-hpu-backend/src/fw/fw_impl/llt/mod.rs

@@ -70,7 +70,7 @@ crate::impl_fw!("Llt" [
     IF_THEN_ZERO => fw_impl::ilp::iop_if_then_zero;
     IF_THEN_ELSE => fw_impl::ilp::iop_if_then_else;
 
-    ERC_20 => fw_impl::llt::iop_erc_20;
+    ERC_7984 => fw_impl::llt::iop_erc_7984;
     MEMCPY => fw_impl::ilp::iop_memcpy;
 
     COUNT0 => fw_impl::ilp_log::iop_count0;
@@ -83,7 +83,7 @@ crate::impl_fw!("Llt" [
 
     // SIMD Implementations
     ADD_SIMD     => fw_impl::llt::iop_add_simd;
-    ERC_20_SIMD  => fw_impl::llt::iop_erc_20_simd;
+    ERC_7984_SIMD  => fw_impl::llt::iop_erc_7984_simd;
 ]);
 
 // ----------------------------------------------------------------------------
@@ -225,24 +225,24 @@ pub fn iop_muls(prog: &mut Program) {
 }
 
 #[instrument(level = "trace", skip(prog))]
-pub fn iop_erc_20(prog: &mut Program) {
+pub fn iop_erc_7984(prog: &mut Program) {
     // Add Comment header
-    prog.push_comment("ERC_20 (new_from, new_to) <- (from, to, amount)".to_string());
+    prog.push_comment("ERC_7984 (new_from, new_to) <- (from, to, amount)".to_string());
     // TODO: Make sweep of kogge_blk_w
     // All these little parameters would be very handy to write an
     // exploration/compilation program which would try to minimize latency by
     // playing with these.
-    iop_erc_20_rtl(prog, 0, Some(10)).add_to_prog(prog);
+    iop_erc_7984_rtl(prog, 0, Some(10)).add_to_prog(prog);
 }
 
 #[instrument(level = "trace", skip(prog))]
-pub fn iop_erc_20_simd(prog: &mut Program) {
+pub fn iop_erc_7984_simd(prog: &mut Program) {
     // Add Comment header
-    prog.push_comment("ERC_20_SIMD (new_from, new_to) <- (from, to, amount)".to_string());
+    prog.push_comment("ERC_7984_SIMD (new_from, new_to) <- (from, to, amount)".to_string());
     simd(
         prog,
         crate::asm::iop::SIMD_N,
-        fw_impl::llt::iop_erc_20_rtl,
+        fw_impl::llt::iop_erc_7984_rtl,
         None,
     );
 }
@@ -379,7 +379,7 @@ pub fn iop_rotate_scalar_left(prog: &mut Program) {
 // Helper Functions
 // ----------------------------------------------------------------------------
 
-/// Implement erc_20 fund xfer
+/// Implement erc_7984 fund xfer
 /// Targeted algorithm is as follow:
 /// 1. Check that from has enough funds
 /// 2. Compute real_amount to xfer (i.e. amount or 0)
@@ -391,7 +391,7 @@ pub fn iop_rotate_scalar_left(prog: &mut Program) {
 ///     (dst_from[0], dst_to[0], ..., dst_from[N-1], dst_to[N-1])
 /// Where N is the batch size
 #[instrument(level = "trace", skip(prog))]
-pub fn iop_erc_20_rtl(prog: &mut Program, batch_index: u8, kogge_blk_w: Option<usize>) -> Rtl {
+pub fn iop_erc_7984_rtl(prog: &mut Program, batch_index: u8, kogge_blk_w: Option<usize>) -> Rtl {
     // Allocate metavariables:
     // Dest -> Operand
     let dst_from = prog.iop_template_var(OperandKind::Dst, 2 * batch_index);

backends/zk-cuda-backend/Cargo.toml

@@ -24,7 +24,7 @@ bindgen.workspace = true
 [dependencies]
 ark-ec.workspace = true
 ark-ff.workspace = true
-tfhe-cuda-backend = { version = "=0.14.0", path = "../tfhe-cuda-backend" }
+tfhe-cuda-backend = { version = "0.14.0", path = "../tfhe-cuda-backend" }
 
 [features]
 default = []

backends/zk-cuda-backend/cuda/include/msm.h

@@ -97,28 +97,23 @@ size_t pippenger_scratch_size_g2(uint32_t n, uint32_t gpu_index);
 //   d_scalars: Device pointer to input BigInt scalars (array of n scalars)
 //   n: Number of points/scalars
 //   d_scratch: Caller-provided device scratch buffer for intermediate results
-//   size_tracker: Reference for tracking GPU memory allocation sizes
 void point_msm_g1_async(cudaStream_t stream, uint32_t gpu_index,
                         G1Projective *h_result, const G1Affine *d_points,
                         const Scalar *d_scalars, uint32_t n,
-                        G1Projective *d_scratch, uint64_t &size_tracker,
-                        bool gpu_memory_allocated);
+                        G1Projective *d_scratch);
 
 void point_msm_g1(cudaStream_t stream, uint32_t gpu_index,
                   G1Projective *h_result, const G1Affine *d_points,
-                  const Scalar *d_scalars, uint32_t n, G1Projective *d_scratch,
-                  uint64_t &size_tracker, bool gpu_memory_allocated);
+                  const Scalar *d_scalars, uint32_t n, G1Projective *d_scratch);
 
 // MSM for G2 points with BigInt scalars (projective result)
 // Result is written directly to a host pointer.
 void point_msm_g2_async(cudaStream_t stream, uint32_t gpu_index,
                         G2ProjectivePoint *h_result, const G2Point *d_points,
                         const Scalar *d_scalars, uint32_t n,
-                        G2ProjectivePoint *d_scratch, uint64_t &size_tracker,
-                        bool gpu_memory_allocated);
+                        G2ProjectivePoint *d_scratch);
 
 void point_msm_g2(cudaStream_t stream, uint32_t gpu_index,
                   G2ProjectivePoint *h_result, const G2Point *d_points,
                   const Scalar *d_scalars, uint32_t n,
-                  G2ProjectivePoint *d_scratch, uint64_t &size_tracker,
-                  bool gpu_memory_allocated);
+                  G2ProjectivePoint *d_scratch);

backends/zk-cuda-backend/cuda/src/msm/msm.cu

@@ -8,17 +8,16 @@
 // Multi-Scalar Multiplication (MSM) using Pippenger algorithm for BLS12-446
 
 // Forward declarations for Pippenger implementations
-void point_msm_g1_pippenger_async(
-    cudaStream_t stream, uint32_t gpu_index, G1Projective *h_result,
-    const G1Affine *d_points, const Scalar *d_scalars, uint32_t n,
-    G1Projective *d_scratch, uint64_t &size_tracker, bool gpu_memory_allocated);
+void point_msm_g1_pippenger_async(cudaStream_t stream, uint32_t gpu_index,
+                                  G1Projective *h_result,
+                                  const G1Affine *d_points,
+                                  const Scalar *d_scalars, uint32_t n,
+                                  G1Projective *d_scratch);
 void point_msm_g2_pippenger_async(cudaStream_t stream, uint32_t gpu_index,
                                   G2ProjectivePoint *h_result,
                                   const G2Point *d_points,
                                   const Scalar *d_scalars, uint32_t n,
-                                  G2ProjectivePoint *d_scratch,
-                                  uint64_t &size_tracker,
-                                  bool gpu_memory_allocated);
+                                  G2ProjectivePoint *d_scratch);
 
 // ============================================================================
 // Public MSM API for BigInt scalars
@@ -29,11 +28,9 @@ void point_msm_g2_pippenger_async(cudaStream_t stream, uint32_t gpu_index,
 void point_msm_g1_async(cudaStream_t stream, uint32_t gpu_index,
                         G1Projective *h_result, const G1Affine *d_points,
                         const Scalar *d_scalars, uint32_t n,
-                        G1Projective *d_scratch, uint64_t &size_tracker,
-                        bool gpu_memory_allocated) {
+                        G1Projective *d_scratch) {
   point_msm_g1_pippenger_async(stream, gpu_index, h_result, d_points, d_scalars,
-                               n, d_scratch, size_tracker,
-                               gpu_memory_allocated);
+                               n, d_scratch);
 }
 
 // MSM with BigInt scalars for G2 (projective coordinates internally)
@@ -41,19 +38,17 @@ void point_msm_g1_async(cudaStream_t stream, uint32_t gpu_index,
 void point_msm_g2_async(cudaStream_t stream, uint32_t gpu_index,
                         G2ProjectivePoint *h_result, const G2Point *d_points,
                         const Scalar *d_scalars, uint32_t n,
-                        G2ProjectivePoint *d_scratch, uint64_t &size_tracker,
-                        bool gpu_memory_allocated) {
+                        G2ProjectivePoint *d_scratch) {
   point_msm_g2_pippenger_async(stream, gpu_index, h_result, d_points, d_scalars,
-                               n, d_scratch, size_tracker,
-                               gpu_memory_allocated);
+                               n, d_scratch);
 }
 
 void point_msm_g1(cudaStream_t stream, uint32_t gpu_index,
                   G1Projective *h_result, const G1Affine *d_points,
-                  const Scalar *d_scalars, uint32_t n, G1Projective *d_scratch,
-                  uint64_t &size_tracker, bool gpu_memory_allocated) {
+                  const Scalar *d_scalars, uint32_t n,
+                  G1Projective *d_scratch) {
   point_msm_g1_async(stream, gpu_index, h_result, d_points, d_scalars, n,
-                     d_scratch, size_tracker, gpu_memory_allocated);
+                     d_scratch);
   // The async impl already syncs internally before the CPU-side Horner phase,
   // so the stream is idle here. This sync is kept for defensive correctness.
   cuda_synchronize_stream(stream, gpu_index);
@@ -62,10 +57,9 @@ void point_msm_g1(cudaStream_t stream, uint32_t gpu_index,
 void point_msm_g2(cudaStream_t stream, uint32_t gpu_index,
                   G2ProjectivePoint *h_result, const G2Point *d_points,
                   const Scalar *d_scalars, uint32_t n,
-                  G2ProjectivePoint *d_scratch, uint64_t &size_tracker,
-                  bool gpu_memory_allocated) {
+                  G2ProjectivePoint *d_scratch) {
   point_msm_g2_async(stream, gpu_index, h_result, d_points, d_scalars, n,
-                     d_scratch, size_tracker, gpu_memory_allocated);
+                     d_scratch);
   // See comment in point_msm_g1 above.
   cuda_synchronize_stream(stream, gpu_index);
 }

backends/zk-cuda-backend/cuda/src/msm/pippenger/msm_pippenger.cu

@@ -493,12 +493,13 @@ void horner_combine_cpu(ProjectiveType &result,
 // window sums. The caller is responsible for allocating and freeing this
 // buffer.
 template <typename AffineType, typename ProjectiveType>
-void point_msm_pippenger_impl_async(
-    cudaStream_t stream, uint32_t gpu_index, ProjectiveType *h_result,
-    const AffineType *d_points, const Scalar *d_scalars, uint32_t n,
-    uint32_t threads_per_block, uint32_t window_size, uint32_t bucket_count,
-    ProjectiveType *d_scratch, uint64_t &size_tracker,
-    bool gpu_memory_allocated) {
+void point_msm_pippenger_impl_async(cudaStream_t stream, uint32_t gpu_index,
+                                    ProjectiveType *h_result,
+                                    const AffineType *d_points,
+                                    const Scalar *d_scalars, uint32_t n,
+                                    uint32_t threads_per_block,
+                                    uint32_t window_size, uint32_t bucket_count,
+                                    ProjectiveType *d_scratch) {
   using ProjectivePoint = Projective<ProjectiveType>;
 
   if (n == 0) {
@@ -705,16 +706,13 @@ void point_msm_g1_pippenger_async(cudaStream_t stream, uint32_t gpu_index,
                                   G1Projective *h_result,
                                   const G1Affine *d_points,
                                   const Scalar *d_scalars, uint32_t n,
-                                  G1Projective *d_scratch,
-                                  uint64_t &size_tracker,
-                                  bool gpu_memory_allocated) {
+                                  G1Projective *d_scratch) {
   uint32_t window_size, bucket_count;
   get_g1_window_params(n, window_size, bucket_count);
 
   point_msm_pippenger_impl_async<G1Affine, G1Projective>(
       stream, gpu_index, h_result, d_points, d_scalars, n,
-      msm_threads_per_block<G1Affine>(n), window_size, bucket_count, d_scratch,
-      size_tracker, gpu_memory_allocated);
+      msm_threads_per_block<G1Affine>(n), window_size, bucket_count, d_scratch);
 }
 
 // MSM with BigInt scalars for G2 (projective coordinates internally)
@@ -724,14 +722,11 @@ void point_msm_g2_pippenger_async(cudaStream_t stream, uint32_t gpu_index,
                                   G2ProjectivePoint *h_result,
                                   const G2Point *d_points,
                                   const Scalar *d_scalars, uint32_t n,
-                                  G2ProjectivePoint *d_scratch,
-                                  uint64_t &size_tracker,
-                                  bool gpu_memory_allocated) {
+                                  G2ProjectivePoint *d_scratch) {
   uint32_t window_size, bucket_count;
   get_g2_window_params(n, window_size, bucket_count);
 
   point_msm_pippenger_impl_async<G2Point, G2ProjectivePoint>(
       stream, gpu_index, h_result, d_points, d_scalars, n,
-      msm_threads_per_block<G2Point>(n), window_size, bucket_count, d_scratch,
-      size_tracker, gpu_memory_allocated);
+      msm_threads_per_block<G2Point>(n), window_size, bucket_count, d_scratch);
 }

backends/zk-cuda-backend/cuda/src/primitives/fp.cu

@@ -187,37 +187,82 @@ __host__ __device__ void fp_copy(Fp &dst, const Fp &src) {
 // "Raw" means without modular reduction - performs a + b and returns carry.
 // This is an internal helper used by fp_add() which handles reduction.
 __host__ __device__ UNSIGNED_LIMB fp_add_raw(Fp &c, const Fp &a, const Fp &b) {
+#if defined(__CUDA_ARCH__) && LIMB_BITS_CONFIG == 64
+  // PTX carry-chain: add.cc sets the hardware carry flag, addc.cc propagates
+  // it. This replaces 2 software carry-detect comparisons per limb (~14 extra
+  // instructions across 7 limbs) with zero-cost hardware flag propagation.
+  uint64_t carry_out;
+  asm("add.cc.u64   %0,  %8,  %15;\n\t" // c[0] = a[0] + b[0], set CF
+      "addc.cc.u64  %1,  %9,  %16;\n\t" // c[1] = a[1] + b[1] + CF
+      "addc.cc.u64  %2,  %10, %17;\n\t" // c[2] = a[2] + b[2] + CF
+      "addc.cc.u64  %3,  %11, %18;\n\t" // c[3] = a[3] + b[3] + CF
+      "addc.cc.u64  %4,  %12, %19;\n\t" // c[4] = a[4] + b[4] + CF
+      "addc.cc.u64  %5,  %13, %20;\n\t" // c[5] = a[5] + b[5] + CF
+      "addc.cc.u64  %6,  %14, %21;\n\t" // c[6] = a[6] + b[6] + CF
+      "addc.u64     %7,  0,   0;\n\t"   // carry_out = 0 + 0 + CF
+      : "=l"(c.limb[0]), "=l"(c.limb[1]), "=l"(c.limb[2]), "=l"(c.limb[3]),
+        "=l"(c.limb[4]), "=l"(c.limb[5]), "=l"(c.limb[6]), "=l"(carry_out)
+      : "l"(a.limb[0]), "l"(a.limb[1]), "l"(a.limb[2]), "l"(a.limb[3]),
+        "l"(a.limb[4]), "l"(a.limb[5]), "l"(a.limb[6]), "l"(b.limb[0]),
+        "l"(b.limb[1]), "l"(b.limb[2]), "l"(b.limb[3]), "l"(b.limb[4]),
+        "l"(b.limb[5]), "l"(b.limb[6]));
+  return carry_out;
+#else
+  // Host path: portable software carry detection
   UNSIGNED_LIMB carry = 0;
 
   for (int i = 0; i < FP_LIMBS; i++) {
-    // Add with carry: c = a + b + carry
     UNSIGNED_LIMB sum = a.limb[i] + carry;
-    carry = (sum < a.limb[i]) ? 1 : 0; // Check for overflow
+    carry = (sum < a.limb[i]) ? 1 : 0;
     sum += b.limb[i];
-    carry += (sum < b.limb[i]) ? 1 : 0; // Check for overflow
+    carry += (sum < b.limb[i]) ? 1 : 0;
     c.limb[i] = sum;
   }
 
   return carry;
+#endif
 }
 
 // Subtraction with borrow propagation
 // "Raw" means without modular reduction - performs a - b and returns borrow.
 // This is an internal helper used by fp_sub() which handles reduction.
 __host__ __device__ UNSIGNED_LIMB fp_sub_raw(Fp &c, const Fp &a, const Fp &b) {
+#if defined(__CUDA_ARCH__) && LIMB_BITS_CONFIG == 64
+  // PTX borrow-chain: sub.cc sets the hardware borrow flag, subc.cc propagates
+  // it. Same benefit as fp_add_raw -- eliminates 2 comparisons per limb.
+  uint64_t borrow_out;
+  asm("sub.cc.u64   %0,  %8,  %15;\n\t" // c[0] = a[0] - b[0], set CF
+      "subc.cc.u64  %1,  %9,  %16;\n\t" // c[1] = a[1] - b[1] - CF
+      "subc.cc.u64  %2,  %10, %17;\n\t" // c[2] = a[2] - b[2] - CF
+      "subc.cc.u64  %3,  %11, %18;\n\t" // c[3] = a[3] - b[3] - CF
+      "subc.cc.u64  %4,  %12, %19;\n\t" // c[4] = a[4] - b[4] - CF
+      "subc.cc.u64  %5,  %13, %20;\n\t" // c[5] = a[5] - b[5] - CF
+      "subc.cc.u64  %6,  %14, %21;\n\t" // c[6] = a[6] - b[6] - CF
+      "subc.u64     %7,  0,   0;\n\t"   // borrow_out = 0 - 0 - CF
+      : "=l"(c.limb[0]), "=l"(c.limb[1]), "=l"(c.limb[2]), "=l"(c.limb[3]),
+        "=l"(c.limb[4]), "=l"(c.limb[5]), "=l"(c.limb[6]), "=l"(borrow_out)
+      : "l"(a.limb[0]), "l"(a.limb[1]), "l"(a.limb[2]), "l"(a.limb[3]),
+        "l"(a.limb[4]), "l"(a.limb[5]), "l"(a.limb[6]), "l"(b.limb[0]),
+        "l"(b.limb[1]), "l"(b.limb[2]), "l"(b.limb[3]), "l"(b.limb[4]),
+        "l"(b.limb[5]), "l"(b.limb[6]));
+  // subc.u64 with 0-0-CF produces 0 if no borrow, or 0xFFFFFFFFFFFFFFFF if
+  // borrow. Normalize to 0/1 for callers that check (borrow != 0) or add it.
+  return borrow_out & 1;
+#else
+  // Host path: portable software borrow detection
   UNSIGNED_LIMB borrow = 0;
 
   for (int i = 0; i < FP_LIMBS; i++) {
-    // Subtract with borrow: c = a - b - borrow
     UNSIGNED_LIMB diff = a.limb[i] - borrow;
-    borrow = (diff > a.limb[i]) ? 1 : 0; // Check for underflow
+    borrow = (diff > a.limb[i]) ? 1 : 0;
     UNSIGNED_LIMB old_diff = diff;
     diff -= b.limb[i];
-    borrow += (diff > old_diff) ? 1 : 0; // Check for underflow
+    borrow += (diff > old_diff) ? 1 : 0;
     c.limb[i] = diff;
   }
 
   return borrow;
+#endif
 }
 
 // Addition with modular reduction: c = (a + b) mod p
@@ -226,7 +271,27 @@ __host__ __device__ void fp_add(Fp &c, const Fp &a, const Fp &b) {
   Fp sum;
   UNSIGNED_LIMB carry = fp_add_raw(sum, a, b);
 
-  // If there's a carry or sum >= MODULUS, we need to reduce
+#if defined(__CUDA_ARCH__) && LIMB_BITS_CONFIG == 64
+  // Branchless reduction: always compute sum - p, then select based on
+  // whether reduction was needed. This avoids divergent branches that stall
+  // warps when some threads need reduction and others don't.
+  //
+  // Decision logic:
+  //   carry=1 -> sum overflowed 448 bits, definitely >= p -> use reduced
+  //   carry=0, borrow=0 -> sum >= p in 448 bits -> use reduced
+  //   carry=0, borrow=1 -> sum < p -> use original sum
+  // So: use_original = (!carry) & borrow
+  Fp reduced;
+  UNSIGNED_LIMB borrow = fp_sub_raw(reduced, sum, fp_modulus());
+  UNSIGNED_LIMB use_original = ((carry ^ 1) & borrow);
+  UNSIGNED_LIMB mask =
+      -use_original; // all-ones if keep sum, all-zeros if keep reduced
+
+  for (int i = 0; i < FP_LIMBS; i++) {
+    c.limb[i] = (sum.limb[i] & mask) | (reduced.limb[i] & ~mask);
+  }
+#else
+  // Host path: branching is fine on CPU (branch predictor handles it well)
   const Fp &p = fp_modulus();
   if (carry || fp_cmp(sum, p) != ComparisonType::Less) {
     Fp reduced;
@@ -235,6 +300,7 @@ __host__ __device__ void fp_add(Fp &c, const Fp &a, const Fp &b) {
   } else {
     fp_copy(c, sum);
   }
+#endif
 }
 
 // Subtraction with modular reduction: c = (a - b) mod p
@@ -243,13 +309,28 @@ __host__ __device__ void fp_sub(Fp &c, const Fp &a, const Fp &b) {
   Fp diff;
   UNSIGNED_LIMB borrow = fp_sub_raw(diff, a, b);
 
-  // If there was a borrow, we need to add MODULUS
+#if defined(__CUDA_ARCH__) && LIMB_BITS_CONFIG == 64
+  // Branchless correction: always compute diff + p, select based on borrow.
+  // Same rationale as fp_add -- avoids warp divergence.
+  //   borrow=1 -> a < b, need to add p -> use corrected
+  //   borrow=0 -> a >= b, result is valid -> use diff
+  Fp corrected;
+  fp_add_raw(corrected, diff, fp_modulus());
+  UNSIGNED_LIMB mask =
+      -borrow; // all-ones if borrow (use corrected), all-zeros if not
+
+  for (int i = 0; i < FP_LIMBS; i++) {
+    c.limb[i] = (corrected.limb[i] & mask) | (diff.limb[i] & ~mask);
+  }
+#else
+  // Host path: branching is fine on CPU
   const Fp &p = fp_modulus();
   if (borrow) {
     fp_add_raw(c, diff, p);
   } else {
     fp_copy(c, diff);
   }
+#endif
 }
 
 // Small-constant multiplication via addition chains.
@@ -453,23 +534,223 @@ __host__ __device__ void fp_mont_reduce(Fp &c, const UNSIGNED_LIMB *a) {
   }
 }
 
+// ============================================================================
+// PTX-accelerated CIOS Montgomery multiplication (device path)
+// ============================================================================
+// The CIOS algorithm for 7 x 64-bit limbs executes 98 multiply-accumulate
+// steps across 7 outer iterations. Each step computes:
+//   (carry, t[j]) = t[j] + a[j] * b_i + carry
+// which is a 64x64->128 multiply plus a three-operand addition with carry.
+//
+// The C++ path uses software carry detection: carry = (sum < old) ? 1 : 0.
+// The PTX path below uses hardware carry flags via the .cc suffix:
+//   - mul.lo.u64 / mul.hi.u64 : 64x64->128 wide multiply
+//   - add.cc.u64 / addc.u64   : addition chain with hardware carry flag
+//
+// Each multiply-accumulate step uses 6 PTX instructions instead of ~10+ in
+// the software-carry version. The 7 outer iterations are fully unrolled, and
+// the limb-shift loop (t[j] = t[j+1]) is eliminated by register renaming.
+//
+// REGISTER ALIASING NOTE: All PTX temporaries (_lo, _hi) are declared as
+// .reg inside the asm block. This prevents nvcc's register allocator from
+// aliasing them with C operands (t_j, carry), which was the root cause of
+// previous correctness bugs where "+l" outputs could share registers with
+// "l" inputs in the same asm statement.
+// ============================================================================
+
+#ifdef __CUDA_ARCH__
+#if LIMB_BITS_CONFIG == 64
+
+// Multiply-accumulate one limb: (carry_out, t_j) = t_j + a_j * b_i + carry_in
+//
+// All intermediates (_lo, _hi) are PTX .reg temporaries inside a { } scope
+// block to avoid: (1) nvcc register aliasing between C operands, and (2)
+// duplicate .reg definitions when the macro is expanded multiple times.
+// The 6-instruction sequence:
+//   mul.lo.u64  _lo, a_j, b_i      -- low 64 bits of product
+//   mul.hi.u64  _hi, a_j, b_i      -- high 64 bits of product
+//   add.cc.u64  t_j, t_j, _lo      -- t_j += _lo, set CF
+//   addc.u64    _hi, _hi, 0        -- _hi += CF
+//   add.cc.u64  t_j, t_j, carry    -- t_j += carry_in, set CF
+//   addc.u64    carry, _hi, 0      -- carry_out = _hi + CF
+#define LIMB_MACC(t_j, carry, a_j, b_i)                                        \
+  asm volatile("{\n\t"                                                         \
+               ".reg .u64 _lo, _hi;\n\t"                                       \
+               "mul.lo.u64  _lo, %2, %3;\n\t"                                  \
+               "mul.hi.u64  _hi, %2, %3;\n\t"                                  \
+               "add.cc.u64  %0, %0, _lo;\n\t"                                  \
+               "addc.u64    _hi, _hi, 0;\n\t"                                  \
+               "add.cc.u64  %0, %0, %1;\n\t"                                   \
+               "addc.u64    %1, _hi, 0;\n\t"                                   \
+               "}\n\t"                                                         \
+               : "+l"(t_j), "+l"(carry)                                        \
+               : "l"(a_j), "l"(b_i))
+
+// Single CIOS iteration: multiply-accumulate, reduce, and shift.
+//
+// Computes:
+//   1. t += a * b_i  (7 limb multiply-accumulate with carry chain)
+//   2. m = t[0] * p_prime  (Montgomery reduction factor)
+//   3. t += m * p  (reduction, zeros out t[0])
+//   4. Shift t right by one limb (via register renaming into r0..r7)
+//
+// The macro lets the compiler allocate registers across all 7 unrolled
+// iterations, avoiding spills to local memory.
+#define CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, \
+                           a5, a6, b_i, p0, p1, p2, p3, p4, p5, p6, p_prime,   \
+                           r0, r1, r2, r3, r4, r5, r6, r7)                     \
+  do {                                                                         \
+    uint64_t _carry = 0;                                                       \
+    /* Step 1: t += a * b_i */                                                 \
+    LIMB_MACC(t0, _carry, a0, b_i);                                            \
+    LIMB_MACC(t1, _carry, a1, b_i);                                            \
+    LIMB_MACC(t2, _carry, a2, b_i);                                            \
+    LIMB_MACC(t3, _carry, a3, b_i);                                            \
+    LIMB_MACC(t4, _carry, a4, b_i);                                            \
+    LIMB_MACC(t5, _carry, a5, b_i);                                            \
+    LIMB_MACC(t6, _carry, a6, b_i);                                            \
+    /* Accumulate final carry into overflow limb t7 */                         \
+    uint64_t _overflow;                                                        \
+    asm("add.cc.u64  %0, %0, %2;\n\t"                                          \
+        "addc.u64    %1, 0, 0;\n\t"                                            \
+        : "+l"(t7), "=l"(_overflow)                                            \
+        : "l"(_carry));                                                        \
+                                                                               \
+    /* Step 2: m = t0 * p_prime mod 2^64 */                                    \
+    uint64_t _m = t0 * p_prime;                                                \
+                                                                               \
+    /* Step 3: t += m * p (zeros out t0) */                                    \
+    _carry = 0;                                                                \
+    LIMB_MACC(t0, _carry, _m, p0);                                             \
+    LIMB_MACC(t1, _carry, _m, p1);                                             \
+    LIMB_MACC(t2, _carry, _m, p2);                                             \
+    LIMB_MACC(t3, _carry, _m, p3);                                             \
+    LIMB_MACC(t4, _carry, _m, p4);                                             \
+    LIMB_MACC(t5, _carry, _m, p5);                                             \
+    LIMB_MACC(t6, _carry, _m, p6);                                             \
+    /* Finalize overflow: t7 = t7 + _carry + _overflow */                      \
+    /* Plain adds (no carry chain) -- the CIOS invariant guarantees this */    \
+    /* sum fits in 64 bits so intermediate overflow does not matter. */        \
+    t7 += _carry;                                                              \
+    t7 += _overflow;                                                           \
+                                                                               \
+    /* Step 4: Shift right by one limb via register renaming */                \
+    /* t0 is now zero (by construction of m), discard it */                    \
+    r0 = t1;                                                                   \
+    r1 = t2;                                                                   \
+    r2 = t3;                                                                   \
+    r3 = t4;                                                                   \
+    r4 = t5;                                                                   \
+    r5 = t6;                                                                   \
+    r6 = t7;                                                                   \
+    r7 = 0;                                                                    \
+  } while (0)
+
+__device__ __noinline__ void fp_mont_mul_cios_ptx(Fp &c, const Fp &a,
+                                                  const Fp &b) {
+  const uint64_t p0 = DEVICE_MODULUS.limb[0];
+  const uint64_t p1 = DEVICE_MODULUS.limb[1];
+  const uint64_t p2 = DEVICE_MODULUS.limb[2];
+  const uint64_t p3 = DEVICE_MODULUS.limb[3];
+  const uint64_t p4 = DEVICE_MODULUS.limb[4];
+  const uint64_t p5 = DEVICE_MODULUS.limb[5];
+  const uint64_t p6 = DEVICE_MODULUS.limb[6];
+  const uint64_t pp = DEVICE_P_PRIME;
+
+  const uint64_t a0 = a.limb[0], a1 = a.limb[1], a2 = a.limb[2];
+  const uint64_t a3 = a.limb[3], a4 = a.limb[4], a5 = a.limb[5];
+  const uint64_t a6 = a.limb[6];
+
+  // Accumulator: 7 limbs + 1 overflow, initialized to zero
+  uint64_t t0 = 0, t1 = 0, t2 = 0, t3 = 0;
+  uint64_t t4 = 0, t5 = 0, t6 = 0, t7 = 0;
+
+  // 7 fully-unrolled CIOS iterations with register renaming for the shift.
+  // Each iteration processes one limb of b, accumulates a*b[i], reduces,
+  // and shifts. The output registers become the input for the next iteration.
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[0], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[1], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[2], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[3], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[4], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[5], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  CIOS_ITERATION_PTX(t0, t1, t2, t3, t4, t5, t6, t7, a0, a1, a2, a3, a4, a5, a6,
+                     b.limb[6], p0, p1, p2, p3, p4, p5, p6, pp, t0, t1, t2, t3,
+                     t4, t5, t6, t7);
+
+  // Final reduction: if t[0..7] >= p (extended to 8 limbs), subtract p.
+  // Compute (t[0..6] - p[0..6]) with borrow, then subtract borrow from t7.
+  // If t7 after subtraction is non-negative, the reduced result is valid;
+  // otherwise the original t[0..6] is already in [0, p).
+  uint64_t r0, r1, r2, r3, r4, r5, r6, mask;
+  asm("sub.cc.u64   %0, %8,  %15;\n\t" // r0 = t0 - p0
+      "subc.cc.u64  %1, %9,  %16;\n\t" // r1 = t1 - p1 - borrow
+      "subc.cc.u64  %2, %10, %17;\n\t" // r2 = t2 - p2 - borrow
+      "subc.cc.u64  %3, %11, %18;\n\t" // r3 = t3 - p3 - borrow
+      "subc.cc.u64  %4, %12, %19;\n\t" // r4 = t4 - p4 - borrow
+      "subc.cc.u64  %5, %13, %20;\n\t" // r5 = t5 - p5 - borrow
+      "subc.cc.u64  %6, %14, %21;\n\t" // r6 = t6 - p6 - borrow
+      "subc.u64     %7, %22, 0;\n\t"   // mask_src = t7 - 0 - borrow
+      "shr.s64      %7, %7, 63;\n\t" // mask = sign-extend: -1 if negative, 0 if
+                                     // >= 0
+      : "=l"(r0), "=l"(r1), "=l"(r2), "=l"(r3), "=l"(r4), "=l"(r5), "=l"(r6),
+        "=l"(mask)
+      : "l"(t0), "l"(t1), "l"(t2), "l"(t3), "l"(t4), "l"(t5), "l"(t6), "l"(p0),
+        "l"(p1), "l"(p2), "l"(p3), "l"(p4), "l"(p5), "l"(p6), "l"(t7));
+
+  // Branchless selection:
+  //   mask = 0  -> t >= p (use reduced r[0..6])
+  //   mask = -1 -> t < p  (keep original t[0..6])
+  c.limb[0] = (t0 & mask) | (r0 & ~mask);
+  c.limb[1] = (t1 & mask) | (r1 & ~mask);
+  c.limb[2] = (t2 & mask) | (r2 & ~mask);
+  c.limb[3] = (t3 & mask) | (r3 & ~mask);
+  c.limb[4] = (t4 & mask) | (r4 & ~mask);
+  c.limb[5] = (t5 & mask) | (r5 & ~mask);
+  c.limb[6] = (t6 & mask) | (r6 & ~mask);
+}
+
+#undef LIMB_MACC
+#undef CIOS_ITERATION_PTX
+
+#endif // LIMB_BITS_CONFIG == 64
+#endif // __CUDA_ARCH__
+
 // CIOS (Coarsely Integrated Operand Scanning) Montgomery multiplication
 // Fuses multiplication and reduction in a single pass for better efficiency.
 // Uses only FP_LIMBS+1 limbs of working space instead of 2*FP_LIMBS.
 // Both a and b are in Montgomery form, result is in Montgomery form.
 __host__ __device__ void fp_mont_mul_cios(Fp &c, const Fp &a, const Fp &b) {
+#if defined(__CUDA_ARCH__) && LIMB_BITS_CONFIG == 64
+  // Device path: fully unrolled PTX with hardware carry flags
+  fp_mont_mul_cios_ptx(c, a, b);
+#else
+  // Host path: portable C++ implementation
   const Fp &p = fp_modulus();
   UNSIGNED_LIMB p_prime = fp_p_prime();
 
   // Working array: only n+1 limbs needed (vs 2n for separate mul+reduce)
   UNSIGNED_LIMB t[FP_LIMBS + 1];
-#ifdef __CUDA_ARCH__
-  for (int i = 0; i < FP_LIMBS + 1; i++) {
-    t[i] = 0;
-  }
-#else
   memset(t, 0, (FP_LIMBS + 1) * sizeof(UNSIGNED_LIMB));
-#endif
 
   // Main CIOS loop: for each limb of b
   for (int i = 0; i < FP_LIMBS; i++) {
@@ -529,14 +810,7 @@ __host__ __device__ void fp_mont_mul_cios(Fp &c, const Fp &a, const Fp &b) {
   }
 
   // Copy result to output
-#ifdef __CUDA_ARCH__
-#pragma unroll
-  for (int i = 0; i < FP_LIMBS; i++) {
-    c.limb[i] = t[i];
-  }
-#else
   memcpy(&c.limb[0], t, FP_LIMBS * sizeof(UNSIGNED_LIMB));
-#endif
 
   // Final reduction: if result >= p or there's overflow, subtract p
   if (t[FP_LIMBS] != 0 || fp_cmp(c, p) != ComparisonType::Less) {
@@ -545,6 +819,7 @@ __host__ __device__ void fp_mont_mul_cios(Fp &c, const Fp &a, const Fp &b) {
     fp_copy(c, reduced);
   }
   // Result is in Montgomery form
+#endif
 }
 
 // Montgomery multiplication: c = (a * b * R_INV) mod p

backends/zk-cuda-backend/cuda/tests_and_benchmarks/benchmarks/CMakeLists.txt

@@ -23,7 +23,8 @@ set(ZK_SRC_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../../src)
 set(ZK_PRIMITIVES_DIR ${ZK_SRC_DIR}/primitives)
 
 # Build device library from tfhe-cuda-backend
-add_library(tfhe_device_bench STATIC ${TFHE_CUDA_BACKEND_DIR}/src/device.cu)
+add_library(tfhe_device_bench STATIC ${TFHE_CUDA_BACKEND_DIR}/src/device.cu
+                                     ${TFHE_CUDA_BACKEND_DIR}/src/utils/helper_profile.cu)
 set_target_properties(
   tfhe_device_bench
   PROPERTIES CUDA_SEPARABLE_COMPILATION ON

backends/zk-cuda-backend/cuda/tests_and_benchmarks/benchmarks/benchmark_msm.cu

@@ -140,14 +140,14 @@ static void BM_G1_MSM(benchmark::State &state) {
   // Warm-up iterations
   for (int i = 0; i < WARMUP_ITERATIONS; i++) {
     point_msm_g1_async(g_benchmark_stream, g_gpu_index, &h_result, d_points,
-                       d_scalars, n, d_scratch, size_tracker, true);
+                       d_scalars, n, d_scratch);
   }
   cuda_synchronize_stream(g_benchmark_stream, g_gpu_index);
 
   // Benchmark loop: only measure the MSM computation, no memory operations
   for (auto _ : state) {
     point_msm_g1_async(g_benchmark_stream, g_gpu_index, &h_result, d_points,
-                       d_scalars, n, d_scratch, size_tracker, true);
+                       d_scalars, n, d_scratch);
     benchmark::ClobberMemory();
   }
 
@@ -221,14 +221,14 @@ static void BM_G2_MSM(benchmark::State &state) {
   // Warm-up iterations
   for (int i = 0; i < WARMUP_ITERATIONS; i++) {
     point_msm_g2_async(g_benchmark_stream, g_gpu_index, &h_result, d_points,
-                       d_scalars, n, d_scratch, size_tracker, true);
+                       d_scalars, n, d_scratch);
   }
   cuda_synchronize_stream(g_benchmark_stream, g_gpu_index);
 
   // Benchmark loop: only measure the MSM computation, no memory operations
   for (auto _ : state) {
     point_msm_g2_async(g_benchmark_stream, g_gpu_index, &h_result, d_points,
-                       d_scalars, n, d_scratch, size_tracker, true);
+                       d_scalars, n, d_scratch);
     benchmark::ClobberMemory();
   }
 

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/CMakeLists.txt

@@ -20,7 +20,8 @@ set(ZK_SRC_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../../src)
 set(ZK_PRIMITIVES_DIR ${ZK_SRC_DIR}/primitives)
 
 # Build device library from tfhe-cuda-backend
-add_library(tfhe_device STATIC ${TFHE_CUDA_BACKEND_DIR}/src/device.cu)
+add_library(tfhe_device STATIC ${TFHE_CUDA_BACKEND_DIR}/src/device.cu
+                               ${TFHE_CUDA_BACKEND_DIR}/src/utils/helper_profile.cu)
 set_target_properties(
   tfhe_device
   PROPERTIES CUDA_SEPARABLE_COMPILATION ON

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/basic/basic_curve_ops.cu

@@ -13,8 +13,8 @@
 //   ./build/tests_and_benchmarks/tests/basic/basic_curve_ops
 
 #include "curve.h"
+#include "device.h"
 #include "fp.h"
-#include <cassert>
 #include <cstdio>
 #include <cstring>
 
@@ -24,7 +24,7 @@ int main() {
   // (non-Montgomery) form. Convert to Montgomery, then lift to projective for
   // host-side arithmetic.
   const G1Affine &gen_normal = g1_generator();
-  assert(!g1_is_infinity(gen_normal));
+  PANIC_IF_FALSE(!g1_is_infinity(gen_normal), "generator must not be infinity");
 
   G1Affine gen_affine = gen_normal;
   point_to_montgomery_inplace(gen_affine);
@@ -37,21 +37,21 @@ int main() {
 
   // G + (-G) = identity (Z = 0 in the projective convention)
   G1Projective identity = G + neg_G;
-  assert(fp_is_zero(identity.Z));
+  PANIC_IF_FALSE(fp_is_zero(identity.Z), "G + (-G) must be identity (Z = 0)");
   printf("Negation (-G) and G + (-G) = identity: OK\n");
 
   // ---- Addition: 2*G = G + G, 3*G = 2*G + G ----
   G1Projective two_G = G + G;
-  assert(!(two_G == G1Projective())); // not the identity
+  PANIC_IF_FALSE(!(two_G == G1Projective{}), "2*G must not be identity");
 
   G1Projective three_G = two_G + G;
-  assert(!(three_G == G1Projective()));
+  PANIC_IF_FALSE(!(three_G == G1Projective{}), "3*G must not be identity");
   printf("Addition (2*G, 3*G): OK\n");
 
   // ---- Compound assignment: G += G ----
   G1Projective acc = G;
   acc += G; // acc = 2*G
-  assert(acc == two_G);
+  PANIC_IF_FALSE(acc == two_G, "G += G must equal 2*G");
   printf("Compound assignment (+=): OK\n");
 
   // ---- Scalar multiplication: 3*G using Scalar type ----
@@ -61,19 +61,22 @@ int main() {
   scalar_3.limb[0] = 3;
 
   G1Projective three_G_via_scalar = G * scalar_3;
-  assert(!(three_G_via_scalar == G1Projective()));
+  PANIC_IF_FALSE(!(three_G_via_scalar == G1Projective{}),
+                 "3*G via scalar must not be identity");
 
   // Normalise both to Z = 1 (Montgomery) before comparing coordinates.
   normalize_projective_g1(three_G);
   normalize_projective_g1(three_G_via_scalar);
-  assert(three_G == three_G_via_scalar);
+  PANIC_IF_FALSE(three_G == three_G_via_scalar,
+                 "3*G via addition must equal 3*G via scalar multiply");
   printf("Scalar multiplication (3*G == G + G + G): OK\n");
 
   // ---- Projective -> affine conversion ----
   // projective_to_affine_g1 keeps coordinates in Montgomery form.
   G1Affine three_G_affine;
   projective_to_affine_g1(three_G_affine, three_G);
-  assert(!g1_is_infinity(three_G_affine));
+  PANIC_IF_FALSE(!g1_is_infinity(three_G_affine),
+                 "3*G in affine must not be infinity");
   printf("Projective -> affine conversion: OK\n");
 
   // ---- Convert to normal-form coordinates ----
@@ -82,7 +85,8 @@ int main() {
   G1Projective result = three_G_via_scalar;
   normalize_from_montgomery_g1(
       result); // coordinates now in normal (non-Montgomery) form
-  assert(!fp_is_zero(result.Z)); // Z = 1 (non-zero)
+  PANIC_IF_FALSE(!fp_is_zero(result.Z),
+                 "normalized result must have non-zero Z");
   printf("Conversion to normal-form projective: OK\n");
 
   printf("All G1 curve operations passed.\n");

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/basic/basic_fp_ops.cu

@@ -11,8 +11,8 @@
 //   cmake --build build --target basic_fp_ops
 //   ./build/tests_and_benchmarks/tests/basic/basic_fp_ops
 
+#include "device.h"
 #include "fp.h"
-#include <cassert>
 #include <cstdio>
 
 int main() {
@@ -25,16 +25,16 @@ int main() {
   fp_one(b); // b = 1
 
   c = a + b; // c = 2
-  assert(c.limb[0] == 2);
+  PANIC_IF_FALSE(c.limb[0] == 2, "1 + 1 must equal 2");
 
   c = c - a; // c = 1
-  assert(fp_is_one(c));
+  PANIC_IF_FALSE(fp_is_one(c), "2 - 1 must equal 1");
 
   // Compound assignment
   c += a; // c = 2
-  assert(c.limb[0] == 2);
+  PANIC_IF_FALSE(c.limb[0] == 2, "1 += 1 must equal 2");
   c -= b; // c = 1
-  assert(fp_is_one(c));
+  PANIC_IF_FALSE(fp_is_one(c), "2 -= 1 must equal 1");
 
   printf("Addition/subtraction: OK\n");
 
@@ -43,7 +43,7 @@ int main() {
   // form, but for add/sub/neg small normal-form values also work correctly.
   Fp neg_a = -a; // neg_a = -1 mod p
   Fp sum = a + neg_a;
-  assert(fp_is_zero(sum)); // 1 + (-1) = 0
+  PANIC_IF_FALSE(fp_is_zero(sum), "1 + (-1) must equal 0");
   printf("Negation: OK\n");
 
   // ---- Multiplication (Montgomery form required) ----
@@ -56,17 +56,17 @@ int main() {
 
   result_m = one_m * two_m; // result_m = 2 in Montgomery form
   fp_from_montgomery(result, result_m);
-  assert(result.limb[0] == 2);
+  PANIC_IF_FALSE(result.limb[0] == 2, "1 * 2 must equal 2");
 
   result_m = two_m * two_m; // result_m = 4 in Montgomery form
   fp_from_montgomery(result, result_m);
-  assert(result.limb[0] == 4);
+  PANIC_IF_FALSE(result.limb[0] == 4, "2 * 2 must equal 4");
 
   // Compound multiplication
   result_m = two_m;
   result_m *= two_m; // result_m = 4
   fp_from_montgomery(result, result_m);
-  assert(result.limb[0] == 4);
+  PANIC_IF_FALSE(result.limb[0] == 4, "2 *= 2 must equal 4");
 
   // Convert an arbitrary normal-form value to Montgomery before multiplying
   Fp five_normal, five_m, twenty_five_m, twenty_five;
@@ -76,7 +76,7 @@ int main() {
 
   fp_mont_mul(twenty_five_m, five_m, five_m); // 5 * 5 = 25
   fp_from_montgomery(twenty_five, twenty_five_m);
-  assert(twenty_five.limb[0] == 25);
+  PANIC_IF_FALSE(twenty_five.limb[0] == 25, "5 * 5 must equal 25");
 
   printf("Multiplication: OK\n");
 
@@ -88,7 +88,7 @@ int main() {
 
   Fp one_check;
   fp_div(one_check, five_normal, five_normal); // 5 / 5 = 1
-  assert(fp_is_one(one_check));
+  PANIC_IF_FALSE(fp_is_one(one_check), "5 / 5 must equal 1");
 
   // Verify: 5 * 5^{-1} == 1  (using fp_div as a cross-check)
   Fp product;
@@ -98,7 +98,7 @@ int main() {
   fp_zero(two_normal);
   two_normal.limb[0] = 2;
   fp_div(product, two_normal, two_normal); // 2 / 2 = 1
-  assert(fp_is_one(product));
+  PANIC_IF_FALSE(fp_is_one(product), "2 / 2 must equal 1");
 
   printf("Inversion/division: OK\n");
 

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/basic/basic_msm.cu

@@ -20,7 +20,6 @@
 #include "device.h"
 #include "fp.h"
 #include "msm.h"
-#include <cassert>
 #include <cstdio>
 #include <cstring>
 #include <vector>
@@ -33,7 +32,6 @@ int main() {
 
   const uint32_t gpu_index = 0;
   const uint32_t n = 4; // number of points / scalars
-  uint64_t size_tracker = 0;
 
   // ---- Prepare host-side points in Montgomery form ----
   // Use n doublings of the G1 generator: G, 2*G, 4*G, 8*G.
@@ -76,8 +74,7 @@ int main() {
 
   // ---- Run MSM (synchronous wrapper; result written directly to host) ----
   G1Projective h_result;
-  point_msm_g1(stream, gpu_index, &h_result, d_points, d_scalars, n, d_scratch,
-               size_tracker, true);
+  point_msm_g1(stream, gpu_index, &h_result, d_points, d_scalars, n, d_scratch);
 
   // ---- Verify against naive sequential computation on the host ----
   // Expected = sum over i of (scalar[i] * point[i]).
@@ -95,7 +92,8 @@ int main() {
   // Normalise to Z = 1 (Montgomery) before comparing projective coordinates.
   normalize_projective_g1(h_result);
   normalize_projective_g1(expected);
-  assert(h_result == expected);
+  PANIC_IF_FALSE(h_result == expected,
+                 "MSM result must match naive sequential computation");
   printf("MSM result matches naive sequential computation.\n");
 
   // ---- Cleanup ----

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/primitives/test_fp.cu

@@ -3,6 +3,7 @@
 #include "fp.h"
 #include "fp_helpers.h" // Include test-only batch operations and kernels
 #include <chrono>
+#include <cinttypes>
 #include <cstdint>
 #include <cstring>
 #include <cuda_runtime.h>
@@ -297,7 +298,7 @@ protected:
 
 // Test basic addition (on GPU)
 TEST_F(FpArithmeticTest, Addition) {
-  uint64_t size_tracker = 0;
+
   Fp a, b, c, c_cpu;
 
   // Test: 1 + 1 = 2
@@ -320,7 +321,7 @@ TEST_F(FpArithmeticTest, Addition) {
 
 // Test subtraction (on GPU)
 TEST_F(FpArithmeticTest, Subtraction) {
-  uint64_t size_tracker = 0;
+
   Fp a, b, c, a_cpu;
 
   // Test: 2 - 1 = 1
@@ -341,7 +342,7 @@ TEST_F(FpArithmeticTest, Subtraction) {
 
 // Test multiplication (on GPU)
 TEST_F(FpArithmeticTest, Multiplication) {
-  uint64_t size_tracker = 0;
+
   Fp five, three, result, expected;
 
   fp_zero(five);
@@ -370,7 +371,7 @@ TEST_F(FpArithmeticTest, Multiplication) {
 
 // Test negation (on GPU)
 TEST_F(FpArithmeticTest, Negation) {
-  uint64_t size_tracker = 0;
+
   Fp a, neg_a, result;
 
   fp_zero(a);
@@ -395,7 +396,7 @@ TEST_F(FpArithmeticTest, Negation) {
 
 // Test Montgomery conversion round-trip (on GPU)
 TEST_F(FpArithmeticTest, MontgomeryRoundTrip) {
-  uint64_t size_tracker = 0;
+
   Fp value, mont_form, back, mont_form_cpu, back_cpu;
 
   fp_zero(value);
@@ -421,7 +422,7 @@ TEST_F(FpArithmeticTest, MontgomeryRoundTrip) {
 
 // Test Montgomery multiplication (on GPU)
 TEST_F(FpArithmeticTest, MontgomeryMultiplication) {
-  uint64_t size_tracker = 0;
+
   Fp five, three, five_m, three_m, result_m, result, expected, result_cpu;
 
   fp_zero(five);
@@ -460,7 +461,7 @@ TEST_F(FpArithmeticTest, MontgomeryMultiplication) {
 
 // Test comparison operations (on GPU)
 TEST_F(FpArithmeticTest, Comparison) {
-  uint64_t size_tracker = 0;
+
   Fp five, three;
 
   fp_zero(five);
@@ -481,7 +482,7 @@ TEST_F(FpArithmeticTest, Comparison) {
 
 // Test zero and one (on GPU)
 TEST_F(FpArithmeticTest, ZeroAndOne) {
-  uint64_t size_tracker = 0;
+
   Fp zero, one;
 
   fp_zero(zero);
@@ -499,7 +500,7 @@ TEST_F(FpArithmeticTest, ZeroAndOne) {
 
 // Test copy (on GPU)
 TEST_F(FpArithmeticTest, Copy) {
-  uint64_t size_tracker = 0;
+
   Fp a, b, b_cpu;
 
   fp_zero(a);
@@ -522,7 +523,7 @@ TEST_F(FpArithmeticTest, Copy) {
 
 // Test conditional move (on GPU)
 TEST_F(FpArithmeticTest, ConditionalMove) {
-  uint64_t size_tracker = 0;
+
   Fp a, b, result, result_cpu;
 
   fp_zero(a);
@@ -563,7 +564,7 @@ TEST_F(FpArithmeticTest, ConditionalMove) {
 
 // Test multiplication by zero (on GPU)
 TEST_F(FpArithmeticTest, MultiplicationByZero) {
-  uint64_t size_tracker = 0;
+
   Fp a, zero, result, result_cpu;
 
   fp_zero(zero);
@@ -591,7 +592,7 @@ TEST_F(FpArithmeticTest, MultiplicationByZero) {
 
 // Test inversion (on GPU)
 TEST_F(FpArithmeticTest, Inversion) {
-  uint64_t size_tracker = 0;
+
   Fp a, a_inv, result, a_inv_cpu;
 
   fp_zero(a);
@@ -623,7 +624,7 @@ TEST_F(FpArithmeticTest, Inversion) {
 
 // Test inversion of one (on GPU)
 TEST_F(FpArithmeticTest, InversionOfOne) {
-  uint64_t size_tracker = 0;
+
   Fp one, one_inv, one_inv_cpu;
 
   fp_one(one);
@@ -645,7 +646,7 @@ TEST_F(FpArithmeticTest, InversionOfOne) {
 
 // Test division (on GPU)
 TEST_F(FpArithmeticTest, Division) {
-  uint64_t size_tracker = 0;
+
   Fp a, b, quotient, result;
 
   fp_zero(a);
@@ -678,7 +679,7 @@ TEST_F(FpArithmeticTest, Division) {
 
 // Test division by one (on GPU)
 TEST_F(FpArithmeticTest, DivisionByOne) {
-  uint64_t size_tracker = 0;
+
   Fp a, one, result;
 
   fp_one(one);
@@ -707,7 +708,7 @@ TEST_F(FpArithmeticTest, DivisionByOne) {
 
 // Test exponentiation with small exponent (on GPU)
 TEST_F(FpArithmeticTest, ExponentiationSmall) {
-  uint64_t size_tracker = 0;
+
   Fp base, result, expected, result_cpu;
 
   fp_zero(base);
@@ -734,7 +735,7 @@ TEST_F(FpArithmeticTest, ExponentiationSmall) {
 
 // Test exponentiation to power of one (on GPU)
 TEST_F(FpArithmeticTest, ExponentiationToPowerOfOne) {
-  uint64_t size_tracker = 0;
+
   Fp base, result, result_cpu;
 
   fp_zero(base);
@@ -758,7 +759,7 @@ TEST_F(FpArithmeticTest, ExponentiationToPowerOfOne) {
 
 // Test exponentiation to power of zero (on GPU)
 TEST_F(FpArithmeticTest, ExponentiationToPowerOfZero) {
-  uint64_t size_tracker = 0;
+
   Fp base, result, one, result_cpu;
 
   fp_zero(base);
@@ -782,7 +783,7 @@ TEST_F(FpArithmeticTest, ExponentiationToPowerOfZero) {
 
 // Test exponentiation with large exponent (Fermat's little theorem)
 TEST_F(FpArithmeticTest, ExponentiationFermat) {
-  uint64_t size_tracker = 0;
+
   Fp a, result;
 
   fp_zero(a);
@@ -798,7 +799,7 @@ TEST_F(FpArithmeticTest, ExponentiationFermat) {
 
 // Test exponentiation: a^(p-1) = 1 mod p
 TEST_F(FpArithmeticTest, ExponentiationFermatInverse) {
-  uint64_t size_tracker = 0;
+
   Fp a, result, one;
 
   fp_zero(a);
@@ -820,7 +821,7 @@ TEST_F(FpArithmeticTest, ExponentiationFermatInverse) {
 
 // Test square root (on GPU)
 TEST_F(FpArithmeticTest, SquareRoot) {
-  uint64_t size_tracker = 0;
+
   Fp a, square, sqrt_result, verify, square_cpu, sqrt_result_cpu, verify_cpu;
 
   // Test: sqrt(a^2) = a or -a
@@ -872,6 +873,12 @@ TEST_F(FpArithmeticTest, SquareRoot) {
     // Also test on CPU for comparison
     Fp neg_a_cpu = -a;
 
+    // Verify GPU negation matches CPU negation
+    EXPECT_EQ(fp_cmp_gpu(stream, gpu_index, &neg_a, &neg_a_cpu),
+              ComparisonType::Equal)
+        << "GPU negation should match CPU negation";
+    cuda_synchronize_stream(stream, gpu_index);
+
     bool matches_a = (fp_cmp_gpu(stream, gpu_index, &sqrt_result, &a) ==
                       ComparisonType::Equal);
     cuda_synchronize_stream(stream, gpu_index);
@@ -891,7 +898,7 @@ TEST_F(FpArithmeticTest, SquareRoot) {
 
 // Test square root of zero (on GPU)
 TEST_F(FpArithmeticTest, SquareRootOfZero) {
-  uint64_t size_tracker = 0;
+
   Fp zero, result, result_cpu;
   fp_zero(zero);
 
@@ -913,7 +920,7 @@ TEST_F(FpArithmeticTest, SquareRootOfZero) {
 
 // Test square root of one (on GPU)
 TEST_F(FpArithmeticTest, SquareRootOfOne) {
-  uint64_t size_tracker = 0;
+
   Fp one, result, result_cpu;
   fp_one(one);
 
@@ -935,7 +942,7 @@ TEST_F(FpArithmeticTest, SquareRootOfOne) {
 
 // Test quadratic residue check (on GPU)
 TEST_F(FpArithmeticTest, IsQuadraticResidue) {
-  uint64_t size_tracker = 0;
+
   Fp a, square, square_cpu, zero;
 
   fp_zero(a);
@@ -971,7 +978,7 @@ TEST_F(FpArithmeticTest, IsQuadraticResidue) {
 // device. For now, we test individual conversions on GPU and verify with GPU
 // comparisons
 TEST_F(FpArithmeticTest, BatchMontgomeryConversion) {
-  uint64_t size_tracker = 0;
+
   const int n = 10;
   Fp normal[n], montgomery[n], back[n];
 
@@ -1011,7 +1018,7 @@ TEST_F(FpArithmeticTest, BatchMontgomeryConversion) {
 
 // Test 1: Addition that doesn't overflow (on GPU)
 TEST_F(FpArithmeticTest, LargeAddition1) {
-  uint64_t size_tracker = 0;
+
   // a = large value
   Fp a = test_utils::make_fp(0x18e00013555855ULL, 0x2b772294629DAULL,
                              0x412736E1F11D66ULL, 0x87BAD325DD638ULL,
@@ -1044,7 +1051,7 @@ TEST_F(FpArithmeticTest, LargeAddition1) {
 
 // Test 2: Addition that triggers reduction (sum > p) (on GPU)
 TEST_F(FpArithmeticTest, LargeAddition2WithReduction) {
-  uint64_t size_tracker = 0;
+
   // Use two large numbers that will trigger reduction
   // a + b should wrap around modulus
   Fp a = test_utils::make_fp(0x311c0026aab0aaaaULL, 0x56ee4528c573b5ccULL,
@@ -1076,7 +1083,7 @@ TEST_F(FpArithmeticTest, LargeAddition2WithReduction) {
 
 // Test 3: Subtraction without borrow (on GPU)
 TEST_F(FpArithmeticTest, LargeSubtraction1) {
-  uint64_t size_tracker = 0;
+
   // a = large value
   Fp a = test_utils::make_fp(0x18e00013555855ULL, 0x2b772294629DAULL,
                              0x412736E1F11D66ULL, 0x87BAD325DD638ULL,
@@ -1106,11 +1113,15 @@ TEST_F(FpArithmeticTest, LargeSubtraction1) {
             ComparisonType::Equal)
       << "GPU result should match CPU result";
   cuda_synchronize_stream(stream, gpu_index);
+  EXPECT_EQ(fp_cmp_gpu(stream, gpu_index, &verify, &verify_cpu),
+            ComparisonType::Equal)
+      << "GPU subtraction roundtrip should match CPU roundtrip";
+  cuda_synchronize_stream(stream, gpu_index);
 }
 
 // Test 4: Subtraction with borrow (a < b) (on GPU)
 TEST_F(FpArithmeticTest, LargeSubtraction2WithBorrow) {
-  uint64_t size_tracker = 0;
+
   // a = 50
   Fp a = test_utils::make_fp(0x32ULL, 0x0ULL, 0x0ULL, 0x0ULL, 0x0ULL, 0x0ULL,
                              0x0ULL);
@@ -1144,7 +1155,7 @@ TEST_F(FpArithmeticTest, LargeSubtraction2WithBorrow) {
 
 // Test 5: Multiplication of large values (triggers reduction) (on GPU)
 TEST_F(FpArithmeticTest, LargeMultiplication1) {
-  uint64_t size_tracker = 0;
+
   // a = 2^200 (bit 200 set)
   Fp a;
   fp_zero(a);
@@ -1190,7 +1201,7 @@ TEST_F(FpArithmeticTest, LargeMultiplication1) {
 
 // Test 6: (p-1) * (p-1) = 1 (mod p) (on GPU)
 TEST_F(FpArithmeticTest, LargeMultiplication2ModulusMinus1) {
-  uint64_t size_tracker = 0;
+
   // a = p - 1
   Fp a = test_utils::make_fp(0x311c0026aab0aaaaULL, 0x56ee4528c573b5ccULL,
                              0x824e6dc3e23acdeeULL, 0xf75a64bbac71602ULL,
@@ -1228,7 +1239,7 @@ TEST_F(FpArithmeticTest, LargeMultiplication2ModulusMinus1) {
 
 // Test 7: Multiplication with 2: a * 2 = a + a (on GPU)
 TEST_F(FpArithmeticTest, LargeMultiplication3Half) {
-  uint64_t size_tracker = 0;
+
   // a = large value
   Fp a = test_utils::make_fp(0x18e00013555855ULL, 0x2b772294629DAE6ULL,
                              0x412736E1F11D66F7ULL, 0x7BAD325DD638B01ULL,
@@ -1264,11 +1275,15 @@ TEST_F(FpArithmeticTest, LargeMultiplication3Half) {
             ComparisonType::Equal)
       << "GPU result should match CPU result";
   cuda_synchronize_stream(stream, gpu_index);
+  EXPECT_EQ(fp_cmp_gpu(stream, gpu_index, &expected, &expected_cpu),
+            ComparisonType::Equal)
+      << "GPU addition should match CPU addition";
+  cuda_synchronize_stream(stream, gpu_index);
 }
 
 // Test 8: Large number squared (on GPU)
 TEST_F(FpArithmeticTest, LargeMultiplication4Square) {
-  uint64_t size_tracker = 0;
+
   // a = large value
   Fp a = test_utils::make_fp(0x123456789ABCDEFULL, 0xFEDCBA9876543210ULL,
                              0x0ULL, 0x0ULL, 0x0ULL, 0x0ULL, 0x0ULL);
@@ -1305,7 +1320,7 @@ TEST_F(FpArithmeticTest, LargeMultiplication4Square) {
 
 // Test 9: Addition chain near modulus (on GPU)
 TEST_F(FpArithmeticTest, LargeAddition3Chain) {
-  uint64_t size_tracker = 0;
+
   // Start with p-1
   Fp a = test_utils::make_fp(0x311c0026aab0aaaaULL, 0x56ee4528c573b5ccULL,
                              0x824e6dc3e23acdeeULL, 0x0f75a64bbac71602ULL,
@@ -1332,7 +1347,7 @@ TEST_F(FpArithmeticTest, LargeAddition3Chain) {
 
 // Test 10: Complex multiplication with reduction (on GPU)
 TEST_F(FpArithmeticTest, LargeMultiplication5Complex) {
-  uint64_t size_tracker = 0;
+
   // a = large prime-like number
   Fp a = test_utils::make_fp(0x123456789ABCDEFULL, 0xFEDCBA9876543210ULL,
                              0x0123456789ABCDEFULL, 0xFEDCBA9876543210ULL,
@@ -1386,7 +1401,7 @@ TEST_F(FpArithmeticTest, LargeMultiplication5Complex) {
 
 // Test addition associativity: (a + b) + c = a + (b + c) (on GPU)
 TEST_F(FpPropertyTest, AdditionAssociativity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1412,7 +1427,7 @@ TEST_F(FpPropertyTest, AdditionAssociativity) {
 
 // Test multiplication associativity: (a * b) * c = a * (b * c) (on GPU)
 TEST_F(FpPropertyTest, MultiplicationAssociativity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) { // Fewer iterations due to multiplication cost
     Fp a = random_value();
     Fp b = random_value();
@@ -1438,7 +1453,7 @@ TEST_F(FpPropertyTest, MultiplicationAssociativity) {
 
 // Test distributivity: a * (b + c) = a*b + a*c (on GPU)
 TEST_F(FpPropertyTest, MultiplicationDistributivity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1465,7 +1480,7 @@ TEST_F(FpPropertyTest, MultiplicationDistributivity) {
 
 // Test addition commutativity with random values (on GPU)
 TEST_F(FpPropertyTest, AdditionCommutativityRandom) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1483,7 +1498,7 @@ TEST_F(FpPropertyTest, AdditionCommutativityRandom) {
 
 // Test multiplication commutativity with random values (on GPU)
 TEST_F(FpPropertyTest, MultiplicationCommutativityRandom) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1501,7 +1516,7 @@ TEST_F(FpPropertyTest, MultiplicationCommutativityRandom) {
 
 // Test additive identity: a + 0 = a (on GPU)
 TEST_F(FpPropertyTest, AdditiveIdentity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp result;
@@ -1517,7 +1532,7 @@ TEST_F(FpPropertyTest, AdditiveIdentity) {
 
 // Test multiplicative identity: a * 1 = a (on GPU)
 TEST_F(FpPropertyTest, MultiplicativeIdentity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp result;
@@ -1533,7 +1548,7 @@ TEST_F(FpPropertyTest, MultiplicativeIdentity) {
 
 // Test additive inverse: a + (-a) = 0 (on GPU)
 TEST_F(FpPropertyTest, AdditiveInverse) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp neg_a, result;
@@ -1550,7 +1565,7 @@ TEST_F(FpPropertyTest, AdditiveInverse) {
 
 // Test double negation: -(-a) = a (on GPU)
 TEST_F(FpPropertyTest, DoubleNegation) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp neg_a, neg_neg_a;
@@ -1568,7 +1583,7 @@ TEST_F(FpPropertyTest, DoubleNegation) {
 
 // Test subtraction as addition of negation: a - b = a + (-b) (on GPU)
 TEST_F(FpPropertyTest, SubtractionAsNegation) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1590,7 +1605,7 @@ TEST_F(FpPropertyTest, SubtractionAsNegation) {
 
 // Test Montgomery form round-trip with random values (on GPU)
 TEST_F(FpPropertyTest, MontgomeryRoundTripRandom) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp mont_form, back;
@@ -1607,7 +1622,7 @@ TEST_F(FpPropertyTest, MontgomeryRoundTripRandom) {
 
 // Test multiplicative inverse: a * a^(-1) = 1 (on GPU)
 TEST_F(FpPropertyTest, MultiplicativeInverse) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     // Skip zero (on GPU)
@@ -1630,7 +1645,7 @@ TEST_F(FpPropertyTest, MultiplicativeInverse) {
 
 // Test division: (a / b) * b = a (on GPU)
 TEST_F(FpPropertyTest, DivisionProperty) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1654,7 +1669,7 @@ TEST_F(FpPropertyTest, DivisionProperty) {
 
 // Test division as multiplication by inverse: a / b = a * b^(-1) (on GPU)
 TEST_F(FpPropertyTest, DivisionAsInverse) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     Fp b = random_value();
@@ -1682,7 +1697,7 @@ TEST_F(FpPropertyTest, DivisionAsInverse) {
 
 // Test exponentiation: (a^e1)^e2 = a^(e1*e2) for small exponents (on GPU)
 TEST_F(FpPropertyTest, ExponentiationPowerOfPower) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 20; i++) { // Fewer iterations due to cost
     Fp a = random_value();
     // Skip zero (on GPU)
@@ -1716,7 +1731,7 @@ TEST_F(FpPropertyTest, ExponentiationPowerOfPower) {
 
 // Test exponentiation: a^e1 * a^e2 = a^(e1+e2) (on GPU)
 TEST_F(FpPropertyTest, ExponentiationProduct) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 20; i++) { // Fewer iterations due to cost
     Fp a = random_value();
     // Skip zero (on GPU)
@@ -1751,7 +1766,7 @@ TEST_F(FpPropertyTest, ExponentiationProduct) {
 
 // Test inversion of inversion: (a^(-1))^(-1) = a (on GPU)
 TEST_F(FpPropertyTest, DoubleInversion) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     // Skip zero (on GPU)
@@ -1775,7 +1790,7 @@ TEST_F(FpPropertyTest, DoubleInversion) {
 
 // Test square root property: sqrt(a^2) = a (for random a) (on GPU)
 TEST_F(FpPropertyTest, SquareRootProperty) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp a = random_value();
     Fp square, sqrt_result, verify;
@@ -1819,7 +1834,7 @@ TEST_F(FpPropertyTest, SquareRootProperty) {
 // Test quadratic residue property: squares are always quadratic residues (on
 // GPU)
 TEST_F(FpPropertyTest, QuadraticResidueProperty) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp a = random_value();
     Fp square;
@@ -1841,7 +1856,7 @@ TEST_F(FpPropertyTest, QuadraticResidueProperty) {
 
 // Test operations with p-1 (on GPU)
 TEST_F(FpEdgeCaseTest, OperationsWithModulusMinusOne) {
-  uint64_t size_tracker = 0;
+
   // (p-1) + 1 = 0 (on GPU)
   Fp result;
   fp_add_gpu(stream, gpu_index, &result, &modulus_minus_one, &one);
@@ -1868,7 +1883,7 @@ TEST_F(FpEdgeCaseTest, OperationsWithModulusMinusOne) {
 
 // Test operations with p-2 (on GPU)
 TEST_F(FpEdgeCaseTest, OperationsWithModulusMinusTwo) {
-  uint64_t size_tracker = 0;
+
   // (p-2) + 1 = p-1 (on GPU)
   Fp result;
   fp_add_gpu(stream, gpu_index, &result, &modulus_minus_two, &one);
@@ -1888,7 +1903,7 @@ TEST_F(FpEdgeCaseTest, OperationsWithModulusMinusTwo) {
 
 // Test operations with very small values (on GPU)
 TEST_F(FpEdgeCaseTest, VerySmallValues) {
-  uint64_t size_tracker = 0;
+
   Fp zero_val, one_val, two_val, three_val;
   fp_zero(zero_val);
   fp_one(one_val);
@@ -1930,7 +1945,7 @@ TEST_F(FpEdgeCaseTest, VerySmallValues) {
 
 // Test operations with max limb values (on GPU)
 TEST_F(FpEdgeCaseTest, MaxLimbValues) {
-  uint64_t size_tracker = 0;
+
   // Test that max_limb_value is valid
   EXPECT_TRUE(test_utils::is_valid_fp(max_limb_value))
       << "max_limb_value should be < p";
@@ -1953,7 +1968,7 @@ TEST_F(FpEdgeCaseTest, MaxLimbValues) {
 
 // Test operations with alternating bit patterns (on GPU)
 TEST_F(FpEdgeCaseTest, AlternatingBitPatterns) {
-  uint64_t size_tracker = 0;
+
   // Test that alternating_bits is valid
   EXPECT_TRUE(test_utils::is_valid_fp(alternating_bits))
       << "alternating_bits should be < p";
@@ -1976,7 +1991,7 @@ TEST_F(FpEdgeCaseTest, AlternatingBitPatterns) {
 
 // Test edge case: zero operations (on GPU)
 TEST_F(FpEdgeCaseTest, ZeroOperations) {
-  uint64_t size_tracker = 0;
+
   // 0 + 0 = 0 (on GPU)
   Fp result;
   fp_add_gpu(stream, gpu_index, &result, &zero, &zero);
@@ -2006,7 +2021,7 @@ TEST_F(FpEdgeCaseTest, ZeroOperations) {
 
 // Test edge case: one operations (on GPU)
 TEST_F(FpEdgeCaseTest, OneOperations) {
-  uint64_t size_tracker = 0;
+
   // 1 + 1 = 2 (on GPU)
   Fp result;
   fp_add_gpu(stream, gpu_index, &result, &one, &one);
@@ -2033,7 +2048,7 @@ TEST_F(FpEdgeCaseTest, OneOperations) {
 
 // Test fp_one_montgomery (on GPU)
 TEST_F(FpEdgeCaseTest, OneMontgomery) {
-  uint64_t size_tracker = 0;
+
   Fp one_mont, one_normal;
   fp_one(one_normal);
   fp_one_montgomery(one_mont);
@@ -2049,7 +2064,7 @@ TEST_F(FpEdgeCaseTest, OneMontgomery) {
 
 // Test repeated operations (stress test) (on GPU)
 TEST_F(FpEdgeCaseTest, RepeatedOperations) {
-  uint64_t size_tracker = 0;
+
   Fp a = test_utils::random_fp(rng);
   Fp result = a;
 
@@ -2091,7 +2106,7 @@ TEST_F(FpEdgeCaseTest, RepeatedOperations) {
 
 // Test CUDA kernel: array addition
 TEST_F(FpCudaKernelTest, CudaKernelArrayAdd) {
-  uint64_t size_tracker = 0;
+
   const int n = 1000;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2126,7 +2141,7 @@ TEST_F(FpCudaKernelTest, CudaKernelArrayAdd) {
 
 // Test CUDA kernel: array multiplication
 TEST_F(FpCudaKernelTest, CudaKernelArrayMul) {
-  uint64_t size_tracker = 0;
+
   const int n = 1000;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2166,7 +2181,7 @@ TEST_F(FpCudaKernelTest, CudaKernelArrayMul) {
 
 // Test CUDA kernel: array addition with edge cases
 TEST_F(FpCudaKernelTest, CudaKernelArrayAddEdgeCases) {
-  uint64_t size_tracker = 0;
+
   const int n = 100;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2216,7 +2231,7 @@ TEST_F(FpCudaKernelTest, CudaKernelArrayAddEdgeCases) {
 
 // Test CUDA kernel: array multiplication with edge cases
 TEST_F(FpCudaKernelTest, CudaKernelArrayMulEdgeCases) {
-  uint64_t size_tracker = 0;
+
   const int n = 100;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2271,7 +2286,7 @@ TEST_F(FpCudaKernelTest, CudaKernelArrayMulEdgeCases) {
 
 // Test CUDA kernel: large array
 TEST_F(FpCudaKernelTest, CudaKernelLargeArray) {
-  uint64_t size_tracker = 0;
+
   const int n = 10000;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2312,7 +2327,7 @@ TEST_F(FpCudaKernelTest, CudaKernelLargeArray) {
 // Test CUDA kernel: boundary conditions for launch configuration
 // Tests that the "if (idx < n)" check works correctly at block boundaries
 TEST_F(FpCudaKernelTest, CudaKernelBoundaryConditions) {
-  uint64_t size_tracker = 0;
+
   // Test sizes that stress the launch configuration
   // threadsPerBlock = 256, so test around block boundaries
   std::vector<int> test_sizes = {1,   255, 256,  257,  511,
@@ -2352,7 +2367,7 @@ TEST_F(FpCudaKernelTest, CudaKernelBoundaryConditions) {
 
 // Test CUDA kernel: verify kernel actually launches (not just CPU fallback)
 TEST_F(FpCudaKernelTest, CudaKernelActuallyLaunches) {
-  uint64_t size_tracker = 0;
+
   const int n = 1000;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2384,7 +2399,7 @@ TEST_F(FpCudaKernelTest, CudaKernelActuallyLaunches) {
 
 // Test CUDA kernel: verify device constant memory is accessible
 TEST_F(FpCudaKernelTest, CudaKernelDeviceConstants) {
-  uint64_t size_tracker = 0;
+
   // This test verifies that DEVICE_MODULUS is properly initialized
   // by running a kernel that uses it (multiplication uses Montgomery which
   // needs modulus)
@@ -2428,7 +2443,7 @@ TEST_F(FpCudaKernelTest, CudaKernelDeviceConstants) {
 
 // Test CUDA kernel: empty array (edge case)
 TEST_F(FpCudaKernelTest, CudaKernelEmptyArray) {
-  uint64_t size_tracker = 0;
+
   const int n = 0;
   Fp *h_a = nullptr;
   Fp *h_b = nullptr;
@@ -2445,7 +2460,7 @@ TEST_F(FpCudaKernelTest, CudaKernelEmptyArray) {
 
 // Test CUDA kernel: single element
 TEST_F(FpCudaKernelTest, CudaKernelSingleElement) {
-  uint64_t size_tracker = 0;
+
   const int n = 1;
   Fp *h_a = new Fp[n];
   Fp *h_b = new Fp[n];
@@ -2471,44 +2486,53 @@ TEST_F(FpCudaKernelTest, CudaKernelSingleElement) {
 // ============================================================================
 
 // Test to print generator values (for hardcoding)
+// PRIx64 format specifiers require 64-bit limbs
+#if LIMB_BITS_CONFIG == 64
 TEST_F(FpArithmeticTest, PrintGenerators) {
-  uint64_t size_tracker = 0;
+
   const G1Affine &g1 = g1_generator();
   const G2Affine &g2 = g2_generator();
 
   printf("\n=== G1 Generator (Montgomery form) ===\n");
-  printf("x: {0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, "
-         "0x%llxULL, 0x%llxULL}\n",
+  printf("x: {0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL}\n",
          g1.x.limb[0], g1.x.limb[1], g1.x.limb[2], g1.x.limb[3], g1.x.limb[4],
          g1.x.limb[5], g1.x.limb[6]);
-  printf("y: {0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, "
-         "0x%llxULL, 0x%llxULL}\n",
+  printf("y: {0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL}\n",
          g1.y.limb[0], g1.y.limb[1], g1.y.limb[2], g1.y.limb[3], g1.y.limb[4],
          g1.y.limb[5], g1.y.limb[6]);
 
   printf("\n=== G2 Generator (Montgomery form) ===\n");
-  printf("x.c0: {0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, "
-         "0x%llxULL, 0x%llxULL}\n",
+  printf("x.c0: {0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL}\n",
          g2.x.c0.limb[0], g2.x.c0.limb[1], g2.x.c0.limb[2], g2.x.c0.limb[3],
          g2.x.c0.limb[4], g2.x.c0.limb[5], g2.x.c0.limb[6]);
-  printf("x.c1: {0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, "
-         "0x%llxULL, 0x%llxULL}\n",
+  printf("x.c1: {0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL}\n",
          g2.x.c1.limb[0], g2.x.c1.limb[1], g2.x.c1.limb[2], g2.x.c1.limb[3],
          g2.x.c1.limb[4], g2.x.c1.limb[5], g2.x.c1.limb[6]);
-  printf("y.c0: {0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, "
-         "0x%llxULL, 0x%llxULL}\n",
+  printf("y.c0: {0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL}\n",
          g2.y.c0.limb[0], g2.y.c0.limb[1], g2.y.c0.limb[2], g2.y.c0.limb[3],
          g2.y.c0.limb[4], g2.y.c0.limb[5], g2.y.c0.limb[6]);
-  printf("y.c1: {0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, 0x%llxULL, "
-         "0x%llxULL, 0x%llxULL}\n",
+  printf("y.c1: {0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, 0x%" PRIx64 "ULL, "
+         "0x%" PRIx64 "ULL}\n",
          g2.y.c1.limb[0], g2.y.c1.limb[1], g2.y.c1.limb[2], g2.y.c1.limb[3],
          g2.y.c1.limb[4], g2.y.c1.limb[5], g2.y.c1.limb[6]);
   printf("\n");
 }
+#endif
 
 // Test is_on_curve_g1 with point at infinity
 TEST_F(FpArithmeticTest, CurveG1PointAtInfinity) {
-  uint64_t size_tracker = 0;
+
   G1Affine point;
   g1_point_at_infinity(point);
 
@@ -2520,7 +2544,7 @@ TEST_F(FpArithmeticTest, CurveG1PointAtInfinity) {
 // We'll create a point by starting with a valid y and computing x
 // Or use a known valid point
 TEST_F(FpArithmeticTest, CurveG1ValidPoint) {
-  uint64_t size_tracker = 0;
+
   G1Affine point;
   point.infinity = false;
 
@@ -2593,7 +2617,7 @@ TEST_F(FpArithmeticTest, CurveG1ValidPoint) {
 
 // Test is_on_curve_g1 with invalid point
 TEST_F(FpArithmeticTest, CurveG1InvalidPoint) {
-  uint64_t size_tracker = 0;
+
   G1Affine point;
   point.infinity = false;
 
@@ -2606,7 +2630,7 @@ TEST_F(FpArithmeticTest, CurveG1InvalidPoint) {
 
 // Test that negating y preserves curve validity (on GPU)
 TEST_F(FpArithmeticTest, CurveG1FieldOperationsConsistency) {
-  uint64_t size_tracker = 0;
+
   G1Affine point;
   point.infinity = false;
 
@@ -2650,7 +2674,7 @@ TEST_F(FpArithmeticTest, CurveG1FieldOperationsConsistency) {
 
 // Test is_on_curve_g2 with point at infinity
 TEST_F(FpArithmeticTest, CurveG2PointAtInfinity) {
-  uint64_t size_tracker = 0;
+
   G2Affine point;
   g2_point_at_infinity(point);
 

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/primitives/test_fp2.cu

@@ -152,7 +152,7 @@ protected:
 
 // Test basic addition (on GPU)
 TEST_F(Fp2ArithmeticTest, Addition) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, b, c, c_cpu;
 
   // Test: (1 + 0*i) + (1 + 0*i) = (2 + 0*i)
@@ -175,7 +175,7 @@ TEST_F(Fp2ArithmeticTest, Addition) {
 
 // Test subtraction (on GPU)
 TEST_F(Fp2ArithmeticTest, Subtraction) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, b, c, a_cpu;
 
   // Test: (2 + 0*i) - (1 + 0*i) = (1 + 0*i)
@@ -197,7 +197,7 @@ TEST_F(Fp2ArithmeticTest, Subtraction) {
 
 // Test multiplication (on GPU)
 TEST_F(Fp2ArithmeticTest, Multiplication) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, b, result, expected, result_cpu;
 
   // Test: (1 + 1*i) * (1 + 1*i) = (0 + 2*i)
@@ -224,7 +224,7 @@ TEST_F(Fp2ArithmeticTest, Multiplication) {
 
 // Test i * i = -1 (on GPU)
 TEST_F(Fp2ArithmeticTest, I_Squared) {
-  uint64_t size_tracker = 0;
+
   Fp2 i_val, result, expected, result_cpu;
 
   // i = 0 + 1*i
@@ -250,7 +250,7 @@ TEST_F(Fp2ArithmeticTest, I_Squared) {
 
 // Test negation (on GPU)
 TEST_F(Fp2ArithmeticTest, Negation) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, neg_a, result, neg_a_cpu, result_cpu;
 
   a = test_utils_fp2::make_fp2_simple(5, 3);
@@ -273,7 +273,7 @@ TEST_F(Fp2ArithmeticTest, Negation) {
 
 // Test conjugation (on GPU)
 TEST_F(Fp2ArithmeticTest, Conjugation) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, conj, result, conj_cpu, result_cpu;
 
   a = test_utils_fp2::make_fp2_simple(5, 3);
@@ -304,7 +304,7 @@ TEST_F(Fp2ArithmeticTest, Conjugation) {
 
 // Test squaring (on GPU)
 TEST_F(Fp2ArithmeticTest, Squaring) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, square, square_cpu;
 
   // Test: (1 + 1*i)^2 = 2*i
@@ -327,7 +327,7 @@ TEST_F(Fp2ArithmeticTest, Squaring) {
 
 // Test zero and one (on GPU)
 TEST_F(Fp2ArithmeticTest, ZeroAndOne) {
-  uint64_t size_tracker = 0;
+
   Fp2 zero_val, one_val;
 
   fp2_zero(zero_val);
@@ -349,7 +349,7 @@ TEST_F(Fp2ArithmeticTest, ZeroAndOne) {
 
 // Test copy (on GPU)
 TEST_F(Fp2ArithmeticTest, Copy) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, b, b_cpu;
 
   a = test_utils_fp2::make_fp2_simple(42, 123);
@@ -370,7 +370,7 @@ TEST_F(Fp2ArithmeticTest, Copy) {
 
 // Test conditional move (on GPU)
 TEST_F(Fp2ArithmeticTest, ConditionalMove) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, b, result, result_cpu;
 
   a = test_utils_fp2::make_fp2_simple(10, 20);
@@ -411,7 +411,7 @@ TEST_F(Fp2ArithmeticTest, ConditionalMove) {
 
 // Test multiplication by zero (on GPU)
 TEST_F(Fp2ArithmeticTest, MultiplicationByZero) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, zero_val, result, result_cpu;
 
   fp2_zero(zero_val);
@@ -432,7 +432,7 @@ TEST_F(Fp2ArithmeticTest, MultiplicationByZero) {
 
 // Test inversion (on GPU)
 TEST_F(Fp2ArithmeticTest, Inversion) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, a_inv, result, a_inv_cpu, result_cpu;
 
   a = test_utils_fp2::make_fp2_simple(5, 3);
@@ -456,7 +456,7 @@ TEST_F(Fp2ArithmeticTest, Inversion) {
 
 // Test division (on GPU)
 TEST_F(Fp2ArithmeticTest, Division) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, b, quotient, result, quotient_cpu, result_cpu;
 
   a = test_utils_fp2::make_fp2_simple(10, 6);
@@ -482,7 +482,7 @@ TEST_F(Fp2ArithmeticTest, Division) {
 
 // Test multiply by i (on GPU)
 TEST_F(Fp2ArithmeticTest, MultiplyByI) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, result, result_cpu;
 
   // Test: (a + b*i) * i = -b + a*i
@@ -509,7 +509,7 @@ TEST_F(Fp2ArithmeticTest, MultiplyByI) {
 
 // Test Frobenius map (on GPU)
 TEST_F(Fp2ArithmeticTest, Frobenius) {
-  uint64_t size_tracker = 0;
+
   Fp2 a, frob, conj, frob_cpu, conj_cpu;
 
   a = test_utils_fp2::make_fp2_simple(5, 3);
@@ -541,7 +541,7 @@ TEST_F(Fp2ArithmeticTest, Frobenius) {
 
 // Test addition associativity: (a + b) + c = a + (b + c) (on GPU)
 TEST_F(Fp2PropertyTest, AdditionAssociativity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp2 a = random_value();
     Fp2 b = random_value();
@@ -567,7 +567,7 @@ TEST_F(Fp2PropertyTest, AdditionAssociativity) {
 
 // Test multiplication associativity: (a * b) * c = a * (b * c) (on GPU)
 TEST_F(Fp2PropertyTest, MultiplicationAssociativity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp2 a = random_value();
     Fp2 b = random_value();
@@ -593,7 +593,7 @@ TEST_F(Fp2PropertyTest, MultiplicationAssociativity) {
 
 // Test distributivity: a * (b + c) = a*b + a*c (on GPU)
 TEST_F(Fp2PropertyTest, MultiplicationDistributivity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp2 a = random_value();
     Fp2 b = random_value();
@@ -620,7 +620,7 @@ TEST_F(Fp2PropertyTest, MultiplicationDistributivity) {
 
 // Test addition commutativity (on GPU)
 TEST_F(Fp2PropertyTest, AdditionCommutativity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp2 a = random_value();
     Fp2 b = random_value();
@@ -638,7 +638,7 @@ TEST_F(Fp2PropertyTest, AdditionCommutativity) {
 
 // Test multiplication commutativity (on GPU)
 TEST_F(Fp2PropertyTest, MultiplicationCommutativity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp2 a = random_value();
     Fp2 b = random_value();
@@ -656,7 +656,7 @@ TEST_F(Fp2PropertyTest, MultiplicationCommutativity) {
 
 // Test additive identity: a + 0 = a (on GPU)
 TEST_F(Fp2PropertyTest, AdditiveIdentity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp2 a = random_value();
     Fp2 result;
@@ -672,7 +672,7 @@ TEST_F(Fp2PropertyTest, AdditiveIdentity) {
 
 // Test multiplicative identity: a * 1 = a (on GPU)
 TEST_F(Fp2PropertyTest, MultiplicativeIdentity) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp2 a = random_value();
     Fp2 result;
@@ -688,7 +688,7 @@ TEST_F(Fp2PropertyTest, MultiplicativeIdentity) {
 
 // Test additive inverse: a + (-a) = 0 (on GPU)
 TEST_F(Fp2PropertyTest, AdditiveInverse) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 100; i++) {
     Fp2 a = random_value();
     Fp2 neg_a, result;
@@ -705,7 +705,7 @@ TEST_F(Fp2PropertyTest, AdditiveInverse) {
 
 // Test multiplicative inverse: a * a^(-1) = 1 (on GPU)
 TEST_F(Fp2PropertyTest, MultiplicativeInverse) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp2 a = random_value();
     // Skip zero
@@ -726,7 +726,7 @@ TEST_F(Fp2PropertyTest, MultiplicativeInverse) {
 
 // Test square vs multiply by self: a^2 = a * a (on GPU)
 TEST_F(Fp2PropertyTest, SquareVsMultiply) {
-  uint64_t size_tracker = 0;
+
   for (int i = 0; i < 50; i++) {
     Fp2 a = random_value();
 
@@ -747,7 +747,7 @@ TEST_F(Fp2PropertyTest, SquareVsMultiply) {
 
 // Test CUDA kernel: array addition
 TEST_F(Fp2CudaKernelTest, CudaKernelArrayAdd) {
-  uint64_t size_tracker = 0;
+
   const int n = 1000;
   Fp2 *h_a = new Fp2[n];
   Fp2 *h_b = new Fp2[n];
@@ -784,7 +784,7 @@ TEST_F(Fp2CudaKernelTest, CudaKernelArrayAdd) {
 
 // Test CUDA kernel: array multiplication
 TEST_F(Fp2CudaKernelTest, CudaKernelArrayMul) {
-  uint64_t size_tracker = 0;
+
   const int n = 1000;
   Fp2 *h_a = new Fp2[n];
   Fp2 *h_b = new Fp2[n];
@@ -825,7 +825,7 @@ TEST_F(Fp2CudaKernelTest, CudaKernelArrayMul) {
 
 // Test is_on_curve_g2 with point at infinity
 TEST_F(Fp2ArithmeticTest, CurveG2PointAtInfinity) {
-  uint64_t size_tracker = 0;
+
   G2Affine point;
   g2_point_at_infinity(point);
 
@@ -835,7 +835,7 @@ TEST_F(Fp2ArithmeticTest, CurveG2PointAtInfinity) {
 
 // Test is_on_curve_g2 with valid point construction
 TEST_F(Fp2ArithmeticTest, CurveG2ValidPointCheck) {
-  uint64_t size_tracker = 0;
+
   G2Affine point;
   point.infinity = false;
 
@@ -860,7 +860,7 @@ TEST_F(Fp2ArithmeticTest, CurveG2ValidPointCheck) {
 
 // Test that field operations maintain curve validity for G2
 TEST_F(Fp2ArithmeticTest, CurveG2FieldOperationsConsistency) {
-  uint64_t size_tracker = 0;
+
   // Create a point (we'll test the consistency check works)
   G2Affine point;
   point.infinity = false;

backends/zk-cuda-backend/cuda/tests_and_benchmarks/tests/test_msm.cu

@@ -24,8 +24,7 @@ static void test_point_msm_g1(cudaStream_t stream, uint32_t gpu_index,
   auto *d_scratch =
       static_cast<G1Projective *>(cuda_malloc_with_size_tracking_async(
           scratch_bytes, stream, gpu_index, size_tracker, true));
-  point_msm_g1(stream, gpu_index, h_result, d_points, d_scalars, n, d_scratch,
-               size_tracker, true);
+  point_msm_g1(stream, gpu_index, h_result, d_points, d_scalars, n, d_scratch);
   cuda_drop_with_size_tracking_async(d_scratch, stream, gpu_index, true);
 }
 
@@ -37,8 +36,7 @@ static void test_point_msm_g2(cudaStream_t stream, uint32_t gpu_index,
   auto *d_scratch =
       static_cast<G2Projective *>(cuda_malloc_with_size_tracking_async(
           scratch_bytes, stream, gpu_index, size_tracker, true));
-  point_msm_g2(stream, gpu_index, h_result, d_points, d_scalars, n, d_scratch,
-               size_tracker, true);
+  point_msm_g2(stream, gpu_index, h_result, d_points, d_scalars, n, d_scratch);
   cuda_drop_with_size_tracking_async(d_scratch, stream, gpu_index, true);
 }
 

backends/zk-cuda-backend/src/bindings.rs

@@ -147,8 +147,6 @@ unsafe extern "C" {
         d_scalars: *const Scalar,
         n: u32,
         d_scratch: *mut G1ProjectivePoint,
-        gpu_memory_allocated: bool,
-        size_tracker: *mut u64,
     );
 }
 unsafe extern "C" {
@@ -160,8 +158,6 @@ unsafe extern "C" {
         d_scalars: *const Scalar,
         n: u32,
         d_scratch: *mut G2ProjectivePoint,
-        gpu_memory_allocated: bool,
-        size_tracker: *mut u64,
     );
 }
 unsafe extern "C" {

backends/zk-cuda-backend/src/c_wrapper.cu

@@ -11,6 +11,8 @@
 #include <stddef.h>
 #include <cstring>
 
+#include "../../tfhe-cuda-backend/cuda/src/utils/helper_profile.cuh"
+
 // C++ helper functions (not exported, used internally)
 // These can call template functions since they have C++ linkage
 static void convert_g1_points_to_montgomery(cudaStream_t stream, uint32_t gpu_index, G1Affine* d_points, uint32_t n) {
@@ -64,8 +66,7 @@ bool g2_is_infinity_wrapper(const G2Affine* point) {
 }
 
 // Unmanaged MSM wrapper for G1 (points/scalars/scratch on device, result on host)
-// Points MUST be in Montgomery form. Caller provides scratch buffer and
-// controls allocation tracking via gpu_memory_allocated.
+// Points MUST be in Montgomery form. Caller provides scratch buffer.
 // Zero internal allocations — this is a thin validation + dispatch layer.
 void g1_msm_unmanaged_wrapper_async(
     cudaStream_t stream,
@@ -74,12 +75,9 @@ void g1_msm_unmanaged_wrapper_async(
     const G1Affine* d_points,
     const Scalar* d_scalars,
     uint32_t n,
-    G1Projective* d_scratch,
-    bool gpu_memory_allocated,
-    uint64_t* size_tracker
+    G1Projective* d_scratch
 ) {
-    PANIC_IF_FALSE(size_tracker != nullptr, "G1 MSM error: size_tracker is null");
-    uint64_t& size_tracker_ref = *size_tracker;
+    PUSH_RANGE("G1 MSM UNMANAGED");
     PANIC_IF_FALSE(n > 0, "G1 MSM error: n must be positive, got %u", n);
     PANIC_IF_FALSE(stream != nullptr, "G1 MSM error: stream is null");
     PANIC_IF_FALSE(h_result != nullptr, "G1 MSM error: h_result is null");
@@ -91,13 +89,13 @@ void g1_msm_unmanaged_wrapper_async(
                    cuda_get_number_of_gpus());
 
     point_msm_g1_async(stream, gpu_index, h_result, d_points, d_scalars, n,
-                       d_scratch, size_tracker_ref, gpu_memory_allocated);
+                       d_scratch);
     check_cuda_error(cudaGetLastError());
+    POP_RANGE();
 }
 
 // Unmanaged MSM wrapper for G2 (points/scalars/scratch on device, result on host)
-// Points MUST be in Montgomery form. Caller provides scratch buffer and
-// controls allocation tracking via gpu_memory_allocated.
+// Points MUST be in Montgomery form. Caller provides scratch buffer.
 // Zero internal allocations — this is a thin validation + dispatch layer.
 void g2_msm_unmanaged_wrapper_async(
     cudaStream_t stream,
@@ -106,12 +104,9 @@ void g2_msm_unmanaged_wrapper_async(
     const G2Affine* d_points,
     const Scalar* d_scalars,
     uint32_t n,
-    G2Projective* d_scratch,
-    bool gpu_memory_allocated,
-    uint64_t* size_tracker
+    G2Projective* d_scratch
 ) {
-    PANIC_IF_FALSE(size_tracker != nullptr, "G2 MSM error: size_tracker is null");
-    uint64_t& size_tracker_ref = *size_tracker;
+    PUSH_RANGE("G2 MSM UNMANAGED");
     PANIC_IF_FALSE(n > 0, "G2 MSM error: n must be positive, got %u", n);
     PANIC_IF_FALSE(stream != nullptr, "G2 MSM error: stream is null");
     PANIC_IF_FALSE(h_result != nullptr, "G2 MSM error: h_result is null");
@@ -123,8 +118,9 @@ void g2_msm_unmanaged_wrapper_async(
                    cuda_get_number_of_gpus());
 
     point_msm_g2_async(stream, gpu_index, h_result, d_points, d_scalars, n,
-                       d_scratch, size_tracker_ref, gpu_memory_allocated);
+                       d_scratch);
     check_cuda_error(cudaGetLastError());
+    POP_RANGE();
 }
 
 // Scratch size query wrappers (needed for bindgen `.*_wrapper` allowlist)
@@ -148,7 +144,8 @@ void g1_msm_managed_wrapper(
     bool points_in_montgomery,
     uint64_t* size_tracker
 ) {
-    uint64_t& size_tracker_ref = *size_tracker;
+    PUSH_RANGE("G1 MSM MANAGED");
+    uint64_t& size_tracker_local = *size_tracker;
     PANIC_IF_FALSE(n > 0, "G1 MSM error: n must be positive, got %u", n);
     PANIC_IF_FALSE(result != nullptr, "G1 MSM error: result is null");
     PANIC_IF_FALSE(stream != nullptr, "G1 MSM error: stream is null");
@@ -165,8 +162,8 @@ void g1_msm_managed_wrapper(
     size_t scalars_bytes = safe_mul_sizeof<Scalar>(static_cast<size_t>(n));
 
     // TODO: We should migrate to _unmanaged_ methods and have scratch/cleanup functions as tfhe-cuda-backend
-    auto* d_points = static_cast<G1Affine*>(cuda_malloc_with_size_tracking_async(points_bytes, stream, gpu_index, size_tracker_ref, true));
-    auto* d_scalars = static_cast<Scalar*>(cuda_malloc_with_size_tracking_async(scalars_bytes, stream, gpu_index, size_tracker_ref, true));
+    auto* d_points = static_cast<G1Affine*>(cuda_malloc_with_size_tracking_async(points_bytes, stream, gpu_index, size_tracker_local, true));
+    auto* d_scalars = static_cast<Scalar*>(cuda_malloc_with_size_tracking_async(scalars_bytes, stream, gpu_index, size_tracker_local, true));
 
     // Always copy points to GPU first
     cuda_memcpy_with_size_tracking_async_to_gpu(d_points, points, points_bytes, stream, gpu_index, true);
@@ -181,14 +178,14 @@ void g1_msm_managed_wrapper(
     // Allocate scratch buffer sized to match the pippenger internal partitioning
     size_t scratch_bytes = pippenger_scratch_size_g1(n, gpu_index);
     auto* d_scratch = static_cast<G1Projective*>(cuda_malloc_with_size_tracking_async(
-        scratch_bytes, stream, gpu_index, size_tracker_ref, true));
+        scratch_bytes, stream, gpu_index, size_tracker_local, true));
 
     PANIC_IF_FALSE(d_points && d_scalars && d_scratch,
                    "G1 MSM error: device memory allocation failed");
 
     // Result written directly to host pointer -- no device round-trip needed
     point_msm_g1_async(stream, gpu_index, result, d_points, d_scalars, n,
-                       d_scratch, size_tracker_ref, true);
+                       d_scratch);
     check_cuda_error(cudaGetLastError());
 
     cuda_drop_with_size_tracking_async(d_scratch, stream, gpu_index, true);
@@ -197,6 +194,7 @@ void g1_msm_managed_wrapper(
 
     // Sync for the async frees above.
     cuda_synchronize_stream(stream, gpu_index);
+    POP_RANGE();
 }
 
 
@@ -212,7 +210,8 @@ void g2_msm_managed_wrapper(
     bool points_in_montgomery,
     uint64_t* size_tracker
 ) {
-    uint64_t& size_tracker_ref = *size_tracker;
+    PUSH_RANGE("G2 MSM MANAGED");
+    uint64_t& size_tracker_local = *size_tracker;
     PANIC_IF_FALSE(n > 0, "G2 MSM error: n must be positive, got %u", n);
     PANIC_IF_FALSE(result != nullptr, "G2 MSM error: result is null");
     PANIC_IF_FALSE(stream != nullptr, "G2 MSM error: stream is null");
@@ -229,8 +228,8 @@ void g2_msm_managed_wrapper(
     size_t scalars_bytes = safe_mul_sizeof<Scalar>(static_cast<size_t>(n));
 
     // TODO: We should migrate to _unmanaged_ methods and have scratch/cleanup functions as tfhe-cuda-backend
-    auto* d_points = static_cast<G2Affine*>(cuda_malloc_with_size_tracking_async(points_bytes, stream, gpu_index, size_tracker_ref, true));
-    auto* d_scalars = static_cast<Scalar*>(cuda_malloc_with_size_tracking_async(scalars_bytes, stream, gpu_index, size_tracker_ref, true));
+    auto* d_points = static_cast<G2Affine*>(cuda_malloc_with_size_tracking_async(points_bytes, stream, gpu_index, size_tracker_local, true));
+    auto* d_scalars = static_cast<Scalar*>(cuda_malloc_with_size_tracking_async(scalars_bytes, stream, gpu_index, size_tracker_local, true));
 
     cuda_memcpy_with_size_tracking_async_to_gpu(d_points, points, points_bytes, stream, gpu_index, true);
     cuda_memcpy_with_size_tracking_async_to_gpu(d_scalars, scalars, scalars_bytes, stream, gpu_index, true);
@@ -243,14 +242,14 @@ void g2_msm_managed_wrapper(
     // Allocate scratch buffer sized to match the pippenger internal partitioning
     size_t scratch_bytes = pippenger_scratch_size_g2(n, gpu_index);
     auto* d_scratch = static_cast<G2Projective*>(cuda_malloc_with_size_tracking_async(
-        scratch_bytes, stream, gpu_index, size_tracker_ref, true));
+        scratch_bytes, stream, gpu_index, size_tracker_local, true));
 
     PANIC_IF_FALSE(d_points && d_scalars && d_scratch,
                    "G2 MSM error: device memory allocation failed");
 
     // Result written directly to host pointer -- no device round-trip needed
     point_msm_g2_async(stream, gpu_index, result, d_points, d_scalars, n,
-                       d_scratch, size_tracker_ref, true);
+                       d_scratch);
     check_cuda_error(cudaGetLastError());
 
     cuda_drop_with_size_tracking_async(d_scratch, stream, gpu_index, true);
@@ -259,6 +258,7 @@ void g2_msm_managed_wrapper(
 
     // Sync for the async frees above.
     cuda_synchronize_stream(stream, gpu_index);
+    POP_RANGE();
 }
 
 void g1_from_montgomery_wrapper(G1Affine* result, const G1Affine* point) {