Re-enable ROCm runner for AMD GPU CI (#4854)

Summary:
Pull Request resolved: https://github.com/facebookresearch/faiss/pull/4854

This re-enables the AMD ROCm runner that was previously disabled in D86250489.

Changes from the original configuration:
- Updated runner from `faiss-amd-MI200` to `linux-amd-rocm-mi325-ubuntu-24` to match the currently available GitHub Actions runner
- Updated container image from Ubuntu 22.04 to Ubuntu 24.04 to align with the runner environment

Test change:
- seems like cuda and hip disagree about some small rounding errors, so AI updated the test.

Reviewed By: subhadeepkaran

Differential Revision: D94941142

fbshipit-source-id: d5158b7939e3b7327432aa89a9a0d2e5ed1ad190

.github/actions/build_cmake/action.yml

@@ -100,27 +100,43 @@ runs:
         sudo apt-get -qq update >/dev/null
         sudo apt-get -qq install -y kmod wget gpg >/dev/null
 
+        # Download, prepare, and install the package signing key
+        mkdir --parents --mode=0755 /etc/apt/keyrings
+        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+    - name: Add rocm repository
+      if: inputs.rocm == 'ON'
+      shell: bash
+      run: |
         # Get UBUNTU version name
         UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
 
         # Set ROCm version
         ROCM_VERSION="6.2"
 
-        # Download, prepare, and install the package signing key
-        mkdir --parents --mode=0755 /etc/apt/keyrings
-        wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+        rocm_baseurl="https://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
+        sudo mkdir -p /etc/apt/keyrings
+        wget -qO /tmp/rocm.gpg.key https://repo.radeon.com/rocm/rocm.gpg.key
+        echo "2de99e2354646a90d9903e2a669fc4e36b02c1bbff7075c481e12d7edab2c88b  /tmp/rocm.gpg.key" | sha256sum --check
 
-        # Add rocm repository
-        wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | sudo apt-key add -
-        rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
-        echo "deb [arch=amd64] ${rocm_baseurl} ${UBUNTU_VERSION_NAME} main" | sudo tee /etc/apt/sources.list.d/rocm.list
-        sudo apt-get -qq update --allow-insecure-repositories >/dev/null
-        sudo apt-get -qq install -y --allow-unauthenticated \
-            "rocm-dev${ROCM_VERSION}" "rocm-utils${ROCM_VERSION}" \
-            "rocm-libs${ROCM_VERSION}" >/dev/null
+        echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] ${rocm_baseurl} ${UBUNTU_VERSION_NAME} main" | sudo tee /etc/apt/sources.list.d/rocm.list
+
+        sudo apt-get -qq update >/dev/null
+        sudo apt-get -qq install -y \
+            "rocm-dev${ROCM_VERSION}" "rocm-utils${ROCM_VERSION}" "rocm-libs${ROCM_VERSION}" >/dev/null
+
+
+    - name: Pin BLAS/LAPACK versions
+      if: inputs.rocm == 'ON'
+      shell: bash
+      run: |
+        conda install -y \
+          "libblas=3.9.0=35_*" \
+          "libcblas=3.9.0=35_*" \
+          "liblapack=3.9.0=35_*"
 
         # Fake presence of MI200-class accelerators
-        echo "gfx90a" | sudo tee /opt/rocm/bin/target.lst
+        echo "gfx942" | sudo tee /opt/rocm/bin/target.lst
 
         # Cleanup
         sudo apt-get -qq autoclean >/dev/null
@@ -135,10 +151,14 @@ runs:
         sudo ln -s /lib/x86_64-linux-gnu/libc_nonshared.a /usr/lib64/libc_nonshared.a
         sudo ln -s /usr/lib/x86_64-linux-gnu/libpthread.so.0 /lib64/libpthread.so.0
         sudo ln -s $HOME/miniconda3/x86_64-conda-linux-gnu/sysroot/usr/lib64/libpthread_nonshared.a /usr/lib64/libpthread_nonshared.a
-    - name: Print GPU info
-      if: inputs.gpu == 'ON'
+    - name: Print NVIDIA GPU info
+      if: inputs.gpu == 'ON' && inputs.rocm != 'ON'
       shell: bash
       run: nvidia-smi
+    - name: Print AMD GPU info
+      if: inputs.gpu == 'ON' && inputs.rocm == 'ON'
+      shell: bash
+      run: rocm-smi
     - name: Build all targets
       shell: bash
       run: |

.github/workflows/build-pull-request.yml

@@ -96,6 +96,29 @@ jobs:
         uses: ./.github/actions/build_cmake
         with:
           gpu: ON
+  linux-x86_64-GPU-w-ROCm-cmake:
+    name: Linux x86_64 GPU w/ ROCm (cmake)
+    needs: linux-x86_64-cmake
+    runs-on: linux-amd-rocm-mi325-ubuntu-24
+    container:
+      image: ubuntu:24.04
+      options: --device=/dev/kfd --device=/dev/dri --ipc=host --shm-size 16G --group-add video --cap-add=SYS_PTRACE --cap-add=SYS_ADMIN
+    steps:
+      - name: Container setup
+        run: |
+            if [ -f /.dockerenv ]; then
+              apt-get update && apt-get install -y sudo && apt-get install -y git
+              git config --global --add safe.directory '*'
+            else
+              echo 'Skipping. Current job is not running inside a container.'
+            fi
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Build and Test (cmake)
+        uses: ./.github/actions/build_cmake
+        with:
+          gpu: ON
+          rocm: ON
   linux-x86_64-GPU-w-CUVS-cmake:
     name: Linux x86_64 GPU w/ cuVS (cmake)
     needs: linux-x86_64-cmake

faiss/gpu/test/TestGpuIndexIVFPQ.cpp

@@ -102,7 +102,11 @@ struct Options {
     }
 
     float getCompareEpsilon() const {
-        return 0.035f;
+        // With very low dimensionality (e.g., dim=4, codes=2 giving
+        // dimPerSubQuantizer=2), L2 distances can be very small
+        // (near-zero), causing relative error comparisons to be
+        // unstable despite tiny absolute differences.
+        return (dim <= 8) ? 0.15f : 0.035f;
     }
 
     float getPctMaxDiff1() const {