Update on "Add int1 to int7 dtypes"

Summary:
Similar to #117208, we want to add int1 to int7 for edge use cases
for weight quantization (https://www.internalfb.com/diff/D62464487)

Test Plan:
python test/test_quantization.py -k test_uint4_int4_dtype

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]

Author: jerryzh168

.ci/docker/build.sh

@@ -291,7 +291,7 @@ case "$image" in
     PROTOBUF=yes
     DB=yes
     VISION=yes
-    ROCM_VERSION=6.0
+    ROCM_VERSION=6.1
     NINJA_VERSION=1.9.0
     CONDA_CMAKE=yes
     TRITON=yes
@@ -302,7 +302,7 @@ case "$image" in
     PROTOBUF=yes
     DB=yes
     VISION=yes
-    ROCM_VERSION=6.1
+    ROCM_VERSION=6.2
     NINJA_VERSION=1.9.0
     CONDA_CMAKE=yes
     TRITON=yes

.ci/docker/common/install_cusparselt.sh

@@ -5,7 +5,7 @@ set -ex
 # cuSPARSELt license: https://docs.nvidia.com/cuda/cusparselt/license.html
 mkdir tmp_cusparselt && cd tmp_cusparselt
 
-if [[ ${CUDA_VERSION:0:4} =~ ^12\.[2-4]$ ]]; then
+if [[ ${CUDA_VERSION:0:4} =~ ^12\.[2-6]$ ]]; then
     arch_path='sbsa'
     export TARGETARCH=${TARGETARCH:-$(uname -m)}
     if [ ${TARGETARCH} = 'amd64' ] || [ "${TARGETARCH}" = 'x86_64' ]; then

.ci/docker/common/install_miopen.sh

@@ -10,6 +10,21 @@ if [[ -z $ROCM_VERSION ]]; then
     exit 1;
 fi
 
+IS_UBUNTU=0
+ID=$(grep -oP '(?<=^ID=).+' /etc/os-release | tr -d '"')
+case "$ID" in
+  ubuntu)
+    IS_UBUNTU=1
+    ;;
+  centos)
+    IS_UBUNTU=0
+    ;;
+  *)
+    echo "Unable to determine OS..."
+    exit 1
+    ;;
+esac
+
 # To make version comparison easier, create an integer representation.
 save_IFS="$IFS"
 IFS=. ROCM_VERSION_ARRAY=(${ROCM_VERSION})
@@ -57,9 +72,11 @@ MIOPEN_CMAKE_COMMON_FLAGS="
 -DMIOPEN_BUILD_DRIVER=OFF
 "
 # Pull MIOpen repo and set DMIOPEN_EMBED_DB based on ROCm version
-if [[ $ROCM_INT -ge 60200 ]] && [[ $ROCM_INT -lt 60300 ]]; then
-    echo "ROCm 6.2 MIOpen does not need any patches, do not build from source"
+if [[ $ROCM_INT -ge 60300 ]]; then
+    echo "ROCm 6.3+ MIOpen does not need any patches, do not build from source"
     exit 0
+elif [[ $ROCM_INT -ge 60200 ]] && [[ $ROCM_INT -lt 60300 ]]; then
+    MIOPEN_BRANCH="release/rocm-rel-6.2-staging"
 elif [[ $ROCM_INT -ge 60100 ]] && [[ $ROCM_INT -lt 60200 ]]; then
     echo "ROCm 6.1 MIOpen does not need any patches, do not build from source"
     exit 0
@@ -93,12 +110,21 @@ else
     exit 1
 fi
 
-yum remove -y miopen-hip
+
+if [[ ${IS_UBUNTU} == 1 ]]; then
+  apt-get remove -y miopen-hip
+else
+  yum remove -y miopen-hip
+fi
 
 git clone https://github.com/ROCm/MIOpen -b ${MIOPEN_BRANCH}
 pushd MIOpen
 # remove .git to save disk space since CI runner was running out
 rm -rf .git
+# Don't build CK to save docker build time
+if [[ $ROCM_INT -ge 60200 ]]; then
+    sed -i '/composable_kernel/d' requirements.txt
+fi
 # Don't build MLIR to save docker build time
 # since we are disabling MLIR backend for MIOpen anyway
 if [[ $ROCM_INT -ge 50400 ]] && [[ $ROCM_INT -lt 50500 ]]; then
@@ -111,10 +137,15 @@ cmake -P install_deps.cmake --minimum
 
 # clean up since CI runner was running out of disk space
 rm -rf /tmp/*
-yum clean all
-rm -rf /var/cache/yum
-rm -rf /var/lib/yum/yumdb
-rm -rf /var/lib/yum/history
+if [[ ${IS_UBUNTU} == 1 ]]; then
+  apt-get autoclean && apt-get clean
+  rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+else
+  yum clean all
+  rm -rf /var/cache/yum
+  rm -rf /var/lib/yum/yumdb
+  rm -rf /var/lib/yum/history
+fi
 
 ## Build MIOpen
 mkdir -p build
@@ -131,7 +162,11 @@ make -j $(nproc) package
 # clean up since CI runner was running out of disk space
 rm -rf /usr/local/cget
 
-yum install -y miopen-*.rpm
+if [[ ${IS_UBUNTU} == 1 ]]; then
+  sudo dpkg -i miopen-hip*.deb
+else
+  yum install -y miopen-*.rpm
+fi
 
 popd
 rm -rf MIOpen

.ci/docker/ubuntu-rocm/Dockerfile

@@ -68,6 +68,8 @@ RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh
 RUN rm install_rocm_magma.sh
+ADD ./common/install_miopen.sh install_miopen.sh
+RUN bash ./install_miopen.sh ${ROCM_VERSION} && rm install_miopen.sh
 ENV ROCM_PATH /opt/rocm
 ENV PATH /opt/rocm/bin:$PATH
 ENV PATH /opt/rocm/hcc/bin:$PATH
@@ -121,5 +123,8 @@ RUN bash ./install_cache.sh && rm install_cache.sh
 ARG BUILD_ENVIRONMENT
 ENV BUILD_ENVIRONMENT ${BUILD_ENVIRONMENT}
 
+# Install LLVM dev version (Defined in the pytorch/builder github repository)
+COPY --from=pytorch/llvm:9.0.1 /opt/llvm /opt/llvm
+
 USER jenkins
 CMD ["bash"]

.ci/pytorch/build.sh

@@ -49,13 +49,8 @@ if [[ ${BUILD_ENVIRONMENT} == *"parallelnative"* ]]; then
 fi
 
 # Enable LLVM dependency for TensorExpr testing
-if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
-  export USE_LLVM=/opt/rocm/llvm
-  export LLVM_DIR=/opt/rocm/llvm/lib/cmake/llvm
-else
-  export USE_LLVM=/opt/llvm
-  export LLVM_DIR=/opt/llvm/lib/cmake/llvm
-fi
+export USE_LLVM=/opt/llvm
+export LLVM_DIR=/opt/llvm/lib/cmake/llvm
 
 if [[ "$BUILD_ENVIRONMENT" == *executorch* ]]; then
   # To build test_edge_op_registration
@@ -237,7 +232,7 @@ fi
 
 # Do not change workspace permissions for ROCm CI jobs
 # as it can leave workspace with bad permissions for cancelled jobs
-if [[ "$BUILD_ENVIRONMENT" != *rocm* ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* ]]; then
   # Workaround for dind-rootless userid mapping (https://github.com/pytorch/ci-infra/issues/96)
   WORKSPACE_ORIGINAL_OWNER_ID=$(stat -c '%u' "/var/lib/jenkins/workspace")
   cleanup_workspace() {
@@ -283,6 +278,7 @@ else
     # set only when building other architectures
     # or building non-XLA tests.
     if [[ "$BUILD_ENVIRONMENT" != *rocm*  &&
+          "$BUILD_ENVIRONMENT" != *s390x*   &&
           "$BUILD_ENVIRONMENT" != *xla* ]]; then
       if [[ "$BUILD_ENVIRONMENT" != *py3.8* ]]; then
         # Install numpy-2.0.2 for builds which are backward compatible with 1.X
@@ -345,11 +341,11 @@ else
     CUSTOM_OP_BUILD="${CUSTOM_TEST_ARTIFACT_BUILD_DIR}/custom-op-build"
     CUSTOM_OP_TEST="$PWD/test/custom_operator"
     python --version
-    SITE_PACKAGES="$(python -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())')"
+    SITE_PACKAGES="$(python -c 'import site; print(";".join([x for x in site.getsitepackages()] + [x + "/torch" for x in site.getsitepackages()]))')"
 
     mkdir -p "$CUSTOM_OP_BUILD"
     pushd "$CUSTOM_OP_BUILD"
-    cmake "$CUSTOM_OP_TEST" -DCMAKE_PREFIX_PATH="$SITE_PACKAGES/torch;$SITE_PACKAGES" -DPython_EXECUTABLE="$(which python)" \
+    cmake "$CUSTOM_OP_TEST" -DCMAKE_PREFIX_PATH="$SITE_PACKAGES" -DPython_EXECUTABLE="$(which python)" \
           -DCMAKE_MODULE_PATH="$CUSTOM_TEST_MODULE_PATH" -DUSE_ROCM="$CUSTOM_TEST_USE_ROCM"
     make VERBOSE=1
     popd
@@ -359,10 +355,10 @@ else
     JIT_HOOK_BUILD="${CUSTOM_TEST_ARTIFACT_BUILD_DIR}/jit-hook-build"
     JIT_HOOK_TEST="$PWD/test/jit_hooks"
     python --version
-    SITE_PACKAGES="$(python -c 'from distutils.sysconfig import get_python_lib; print(get_python_lib())')"
+    SITE_PACKAGES="$(python -c 'import site; print(";".join([x for x in site.getsitepackages()] + [x + "/torch" for x in site.getsitepackages()]))')"
     mkdir -p "$JIT_HOOK_BUILD"
     pushd "$JIT_HOOK_BUILD"
-    cmake "$JIT_HOOK_TEST" -DCMAKE_PREFIX_PATH="$SITE_PACKAGES/torch;$SITE_PACKAGES" -DPython_EXECUTABLE="$(which python)" \
+    cmake "$JIT_HOOK_TEST" -DCMAKE_PREFIX_PATH="$SITE_PACKAGES" -DPython_EXECUTABLE="$(which python)" \
           -DCMAKE_MODULE_PATH="$CUSTOM_TEST_MODULE_PATH" -DUSE_ROCM="$CUSTOM_TEST_USE_ROCM"
     make VERBOSE=1
     popd
@@ -374,7 +370,7 @@ else
     python --version
     mkdir -p "$CUSTOM_BACKEND_BUILD"
     pushd "$CUSTOM_BACKEND_BUILD"
-    cmake "$CUSTOM_BACKEND_TEST" -DCMAKE_PREFIX_PATH="$SITE_PACKAGES/torch;$SITE_PACKAGES" -DPython_EXECUTABLE="$(which python)" \
+    cmake "$CUSTOM_BACKEND_TEST" -DCMAKE_PREFIX_PATH="$SITE_PACKAGES" -DPython_EXECUTABLE="$(which python)" \
           -DCMAKE_MODULE_PATH="$CUSTOM_TEST_MODULE_PATH" -DUSE_ROCM="$CUSTOM_TEST_USE_ROCM"
     make VERBOSE=1
     popd
@@ -407,6 +403,6 @@ fi
 
 # snadampal: skipping it till sccache support added for aarch64
 # https://github.com/pytorch/pytorch/issues/121559
-if [[ "$BUILD_ENVIRONMENT" != *aarch64* ]]; then
+if [[ "$BUILD_ENVIRONMENT" != *aarch64* &&  "$BUILD_ENVIRONMENT" != *s390x* ]]; then
   print_sccache_stats
 fi

.ci/pytorch/create_test_cert.py

@@ -1,4 +1,4 @@
-from datetime import datetime, timedelta
+from datetime import datetime, timedelta, timezone
 from tempfile import mkdtemp
 
 from cryptography import x509
@@ -42,10 +42,10 @@ def create_cert(path, C, ST, L, O, key):
         .issuer_name(issuer)
         .public_key(key.public_key())
         .serial_number(x509.random_serial_number())
-        .not_valid_before(datetime.utcnow())
+        .not_valid_before(datetime.now(timezone.utc))
         .not_valid_after(
             # Our certificate will be valid for 10 days
-            datetime.utcnow()
+            datetime.now(timezone.utc)
             + timedelta(days=10)
         )
         .add_extension(
@@ -88,10 +88,10 @@ def sign_certificate_request(path, csr_cert, ca_cert, private_ca_key):
         .issuer_name(ca_cert.subject)
         .public_key(csr_cert.public_key())
         .serial_number(x509.random_serial_number())
-        .not_valid_before(datetime.utcnow())
+        .not_valid_before(datetime.now(timezone.utc))
         .not_valid_after(
             # Our certificate will be valid for 10 days
-            datetime.utcnow()
+            datetime.now(timezone.utc)
             + timedelta(days=10)
             # Sign our certificate with our private key
         )

.ci/pytorch/test.sh

@@ -1402,7 +1402,7 @@ test_linux_aarch64() {
        inductor/test_max_autotune inductor/test_memory_planning inductor/test_metrics inductor/test_multi_kernel inductor/test_pad_mm \
        inductor/test_pattern_matcher inductor/test_perf inductor/test_profiler inductor/test_select_algorithm inductor/test_smoke \
        inductor/test_split_cat_fx_passes inductor/test_standalone_compile inductor/test_torchinductor \
-       inductor/test_torchinductor_codegen_dynamic_shapes inductor/test_torchinductor_dynamic_shapes \
+       inductor/test_torchinductor_codegen_dynamic_shapes inductor/test_torchinductor_dynamic_shapes inductor/test_memory \
        --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 }
 

.github/merge_rules.yaml

@@ -544,6 +544,7 @@
   - anijain2305
   - bdhirsh
   - zou3519
+  - isuruf
   mandatory_checks_name:
   - EasyCLA
   - Lint