fix(gpu): coprocessor bench

Author: andrei-stoian-zama

.github/workflows/coprocessor-benchmark-gpu.yml

@@ -3,6 +3,22 @@ name: coprocessor-benchmark-gpu
 
 on:
   workflow_dispatch:
+    inputs:
+      profile:
+        description: "Instance type"
+        required: true
+        type: choice
+        options:
+          - "l40 (n3-L40x1)"
+          - "4-l40 (n3-L40x4)"
+          - "single-h100 (n3-H100x1)"
+          - "2-h100 (n3-H100x2)"
+          - "4-h100 (n3-H100x4)"
+          - "multi-h100 (n3-H100x8)"
+          - "multi-h100-nvlink (n3-H100x8-NVLink)"
+          - "multi-h100-sxm5 (n3-H100x8-SXM5)"
+          - "multi-h100-sxm5_fallback (n3-H100x8-SXM5)"
+
   schedule:
     # Weekly tests @ 1AM
     - cron: "0 1 * * 6"
@@ -17,7 +33,7 @@ env:
   RUST_BACKTRACE: "full"
   RUST_MIN_STACK: "8388608"
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
-  PROFILE: "multi-h100-sxm5 (n3-H100x8-SXM5)"
+  PROFILE_SCHEDULED_RUN: "multi-h100-sxm5 (n3-H100x8-SXM5)"
   BENCHMARK_TYPE: "ALL"
   OPTIMIZATION_TARGET: "throughput"
   BATCH_SIZE: "5000"
@@ -40,15 +56,25 @@ jobs:
       - name: Parse profile
         id: parse_profile
         run: |
+          if [[ ${{ github.event_name == 'workflow_dispatch' }} == true ]]; then
+            PROFILE_RAW="${{ inputs.profile }}"
+          else
+            PROFILE_RAW="${PROFILE_SCHEDULED_RUN}"
+          fi
           # shellcheck disable=SC2001
-          PROFILE_VAL=$(echo "${PROFILE}" | sed 's|\(.*\)[[:space:]](.*)|\1|')
+          PROFILE_VAL=$(echo "${PROFILE_RAW}" | sed 's|\(.*\)[[:space:]](.*)|\1|')
           echo "profile=$PROFILE_VAL" >> "${GITHUB_OUTPUT}"
 
       - name: Parse hardware name
         id: parse_hardware_name
         run: |
+          if [[ ${{ github.event_name == 'workflow_dispatch' }} == true ]]; then
+            PROFILE_RAW="${{ inputs.profile }}"
+          else
+            PROFILE_RAW="${PROFILE}"
+          fi
           # shellcheck disable=SC2001
-          PROFILE_VAL=$(echo "${PROFILE}" | sed 's|.*[[:space:]](\(.*\))|\1|')
+          PROFILE_VAL=$(echo "${PROFILE_RAW}" | sed 's|.*[[:space:]](\(.*\))|\1|')
           echo "name=$PROFILE_VAL" >> "${GITHUB_OUTPUT}"
 
   setup-instance:
@@ -130,6 +156,13 @@ jobs:
           } >> "${GITHUB_ENV}"
         working-directory: tfhe-rs/
 
+      - name: Setup Hyperstack dependencies
+        uses: ./tfhe-rs/.github/actions/gpu_setup
+        with:
+          cuda-version: ${{ matrix.cuda }}
+          gcc-version: ${{ matrix.gcc }}
+          github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
+
       - name: Check fhEVM and TFHE-rs repos
         run: |
           pwd
@@ -140,21 +173,14 @@ jobs:
         run: git lfs checkout
         working-directory: fhevm/
 
-      - name: Setup Hyperstack dependencies
-        uses: ./fhevm/.github/actions/gpu_setup
-        with:
-          cuda-version: ${{ matrix.cuda }}
-          gcc-version: ${{ matrix.gcc }}
-          github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
-
       - name: Install rust
         uses: dtolnay/rust-toolchain@e97e2d8cc328f1b50210efc529dca0028893a2d9 # zizmor: ignore[stale-action-refs] this action doesn't create releases
         with:
           toolchain: nightly
 
       - name: Install cargo dependencies
         run: |
-          sudo apt-get install -y protobuf-compiler cmake pkg-config libssl-dev \
+          sudo apt-get install -y protobuf-compiler pkg-config libssl-dev \
                                   libclang-dev docker-compose-v2 docker.io acl
           sudo usermod -aG docker "$USER"
           newgrp docker
@@ -181,9 +207,16 @@ jobs:
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Login to Chainguard Registry
+        uses: docker/login-action@9780b0c442fbb1117ed29e0efdff1e18412f7567 # v3.3.0
+        with:
+          registry: cgr.dev
+          username: ${{ secrets.CGR_USERNAME }}
+          password: ${{ secrets.CGR_PASSWORD }}
+
       - name: Init database
         run: make init_db
-        working-directory: fhevm/coprocessor/fhevm-engine/coprocessor
+        working-directory: fhevm/coprocessor/fhevm-engine/tfhe-worker
 
       - name: Use Node.js
         uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0
@@ -203,8 +236,12 @@ jobs:
 
       - name: Profile erc20 no-cmux benchmark on GPU
         run: |
-          BENCHMARK_BATCH_SIZE="${BATCH_SIZE}" FHEVM_DF_SCHEDULE="${SCHEDULING_POLICY}" BENCHMARK_TYPE="LATENCY" OPTIMIZATION_TARGET="${OPTIMIZATION_TARGET}" make -e "profile_erc20_gpu"
-        working-directory: fhevm/coprocessor/fhevm-engine/coprocessor
+          BENCHMARK_BATCH_SIZE="${BATCH_SIZE}" \
+          FHEVM_DF_SCHEDULE="${SCHEDULING_POLICY}" \
+          BENCHMARK_TYPE="THROUGHPUT_200" \
+          OPTIMIZATION_TARGET="${OPTIMIZATION_TARGET}" \
+          make -e "profile_erc20_gpu"
+        working-directory: fhevm/coprocessor/fhevm-engine/tfhe-worker
 
       - name: Get nsys profile name
         id: nsys_profile_name
@@ -215,7 +252,7 @@ jobs:
           REPORT_NAME: ${{ steps.nsys_profile_name.outputs.profile }}
         run: |
           mv report1.nsys-rep ${{ env.REPORT_NAME }}
-        working-directory: fhevm/coprocessor/fhevm-engine/coprocessor
+        working-directory: fhevm/coprocessor/fhevm-engine/tfhe-worker
 
       - name: Upload profile artifact
         env:
@@ -228,12 +265,12 @@ jobs:
       - name: Run latency benchmark on GPU
         run: |
           BENCHMARK_BATCH_SIZE="${BATCH_SIZE}" FHEVM_DF_SCHEDULE="${SCHEDULING_POLICY}" BENCHMARK_TYPE="LATENCY" OPTIMIZATION_TARGET="${OPTIMIZATION_TARGET}" make -e "benchmark_${BENCHMARKS}_gpu"
-        working-directory: fhevm/coprocessor/fhevm-engine/coprocessor
+        working-directory: fhevm/coprocessor/fhevm-engine/tfhe-worker
 
       - name: Run throughput benchmarks on GPU
         run: |
           BENCHMARK_BATCH_SIZE="${BATCH_SIZE}" FHEVM_DF_SCHEDULE="${SCHEDULING_POLICY}" BENCHMARK_TYPE="THROUGHPUT_200" OPTIMIZATION_TARGET="${OPTIMIZATION_TARGET}" make -e "benchmark_${BENCHMARKS}_gpu"
-        working-directory: fhevm/coprocessor/fhevm-engine/coprocessor
+        working-directory: fhevm/coprocessor/fhevm-engine/tfhe-worker
 
       - name: Parse results
         run: |

.github/workflows/gpu_full_multi_gpu_tests.yml

@@ -86,7 +86,7 @@ jobs:
           slab-url: ${{ secrets.SLAB_BASE_URL }}
           job-secret: ${{ secrets.JOB_SECRET }}
           backend: hyperstack
-          profile: multi-gpu-test
+          profile: 4-l40
 
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance

.github/workflows/gpu_integer_long_run_tests.yml

@@ -43,7 +43,7 @@ jobs:
           slab-url: ${{ secrets.SLAB_BASE_URL }}
           job-secret: ${{ secrets.JOB_SECRET }}
           backend: hyperstack
-          profile: multi-gpu-test
+          profile: 4-l40
 
   cuda-tests:
     name: gpu_integer_long_run_tests/cuda-tests

Makefile

@@ -1004,6 +1004,11 @@ test_high_level_api_gpu: install_rs_build_toolchain install_cargo_nextest
 		--test-threads=4 --features=integer,internal-keycache,gpu,zk-pok -p tfhe \
 		-E "test(/high_level_api::.*gpu.*/)"
 
+.PHONY: test_high_level_build
+test_high_level_build: install_rs_build_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --no-run \
+		   --features=integer,gpu-debug -vv -p tfhe -- "$${TEST}" --test-threads=1 --nocapture
+
 test_high_level_api_hpu: install_rs_build_toolchain install_cargo_nextest
 ifeq ($(HPU_CONFIG), v80)
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) nextest run --cargo-profile $(CARGO_PROFILE) \

backends/tfhe-cuda-backend/cuda/CMakeLists.txt

@@ -86,6 +86,7 @@ if(CMAKE_BUILD_TYPE_LOWERCASE STREQUAL "debug")
   message("Compiling in Debug mode")
   add_definitions(-DDEBUG)
   set(OPTIMIZATION_FLAGS "${OPTIMIZATION_FLAGS} -O0 -G -g")
+  set(USE_NVTOOLS 1)
 else()
   # Release mode
   message("Compiling in Release mode")

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

@@ -2,6 +2,9 @@
 #include <cstdint>
 #include <cuda_runtime.h>
 #include <mutex>
+#ifdef USE_NVTOOLS
+#include <cuda_profiler_api.h>
+#endif
 
 uint32_t cuda_get_device() {
   int device;
@@ -83,6 +86,9 @@ void cuda_set_device(uint32_t gpu_index) {
   check_cuda_error(cudaSetDevice(gpu_index));
   // Mempools are initialized only once in all the GPUS available
   cuda_setup_mempool(gpu_index);
+#ifdef USE_NVTOOLS
+  check_cuda_error(cudaProfilerStart());
+#endif
 }
 
 cudaEvent_t cuda_create_event(uint32_t gpu_index) {

ci/slab.toml

@@ -83,18 +83,6 @@ image_name = "Ubuntu Server 22.04 LTS R570 CUDA 12.8"
 flavor_name = "n3-A100x8-NVLink"
 user = "ubuntu"
 
-[backend.hyperstack.multi-gpu-test]
-environment_name = "canada"
-image_name = "Ubuntu Server 22.04 LTS R570 CUDA 12.8"
-flavor_name = "n3-L40x4"
-user = "ubuntu"
-
-[backend.hyperstack.multi-gpu-test_fallback]
-environment_name = "canada"
-image_name = "Ubuntu Server 22.04 LTS R570 CUDA 12.8"
-flavor_name = "n3-RTX-A6000x2"
-user = "ubuntu"
-
 [backend.hyperstack.l40]
 environment_name = "canada"
 image_name = "Ubuntu Server 22.04 LTS R570 CUDA 12.8"
@@ -106,3 +94,9 @@ environment_name = "canada"
 image_name = "Ubuntu Server 22.04 LTS R570 CUDA 12.8"
 flavor_name = "n3-RTX-A6000x1"
 user = "ubuntu"
+
+[backend.hyperstack.4-l40]
+environment_name = "canada"
+image_name = "Ubuntu Server 22.04 LTS R570 CUDA 12.8"
+flavor_name = "n3-L40x4"
+user = "ubuntu"