Merge branch 'main' into etiotto/coalesce_for_block_ptr

Author: etiotto

.github/pins/pytorch-upstream.txt

@@ -1 +1 @@
-0a2685160140656e3e53818611dd2c65c4397be5
+8321eec009c8c79145ebccd51fdfc336e5f8b848

.github/workflows/build-test-reusable.yml

@@ -159,7 +159,6 @@ jobs:
           echo "TRANSFORMERS_VERSION=$TRANSFORMERS_VERSION" | tee -a $GITHUB_ENV
 
       - name: Install transformers
-        if: ${{ inputs.python_version != '3.12' }}
         uses: ./.github/actions/install-dependency
         with:
           package: transformers
@@ -168,7 +167,6 @@ jobs:
           try-tag-prefix: v
 
       - name: Run E2E test
-        if: ${{ inputs.python_version != '3.12' }}
         run: |
           # Set WORKSPACE for inductor_xpu_test.sh to make sure it creates "inductor_log" outside of pytorch cloned directory
           export WORKSPACE=$GITHUB_WORKSPACE

.github/workflows/llvm-build.yml

@@ -107,7 +107,6 @@ jobs:
         -DLLVM_INSTALL_UTILS=ON
         -DLLVM_TARGETS_TO_BUILD="host;NVPTX;AMDGPU"
         -DLLVM_ENABLE_TERMINFO=OFF
-        -DLLVM_ABI_BREAKING_CHECKS=FORCE_OFF
         llvm-project/llvm
 
         ninja -C llvm-project/build check-mlir install
@@ -131,7 +130,6 @@ jobs:
         -DLLVM_INSTALL_UTILS=ON
         -DLLVM_TARGETS_TO_BUILD="host;NVPTX;AMDGPU"
         -DLLVM_ENABLE_TERMINFO=OFF
-        -DLLVM_ABI_BREAKING_CHECKS=FORCE_OFF
         llvm-project/llvm
 
         ninja -C llvm-project/build check-mlir install

.github/workflows/llvm-build/almalinux.Dockerfile

@@ -33,7 +33,6 @@ RUN cmake -GNinja -Bbuild \
   -DLLVM_ENABLE_PROJECTS=mlir \
   -DLLVM_ENABLE_TERMINFO=OFF \
   -DLLVM_INSTALL_UTILS=ON \
-  -DLLVM_ABI_BREAKING_CHECKS=FORCE_OFF \
   -DLLVM_TARGETS_TO_BUILD="host;NVPTX;AMDGPU" \
   /source/llvm-project/llvm
 

.github/workflows/triton-benchmarks.yml

@@ -167,7 +167,7 @@ jobs:
           source ../../scripts/capture-hw-details.sh
 
           python ../../scripts/build_report.py $REPORTS/matmul-performance-bt.csv $REPORTS/gemm-bt-triton-report.csv --benchmark gemm-bt --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
-          python ../../scripts/build_report.py $REPORTS/matmul-performance-bt.csv $REPORTS/gemm-bt-triton-report.csv --benchmark gemm-bt --compiler onednn --param_cols "B,M,K,N" --tflops_col onednn-TFlops --hbm_col "onednn-GB/s" --tag $TAG
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-bt.csv $REPORTS/gemm-bt-onednn-report.csv --benchmark gemm-bt --compiler onednn --param_cols "B,M,K,N" --tflops_col onednn-TFlops --hbm_col "onednn-GB/s" --tag $TAG
 
       - name: Run Triton GEMM (A^t@B) kernel benchmark
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}
@@ -178,7 +178,7 @@ jobs:
           source ../../scripts/capture-hw-details.sh
 
           python ../../scripts/build_report.py $REPORTS/matmul-performance-at.csv $REPORTS/gemm-at-triton-report.csv --benchmark gemm-at --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG
-          python ../../scripts/build_report.py $REPORTS/matmul-performance-at.csv $REPORTS/gemm-at-triton-report.csv --benchmark gemm-at --compiler onednn --param_cols "B,M,K,N" --tflops_col onednn-TFlops --hbm_col "onednn-GB/s" --tag $TAG
+          python ../../scripts/build_report.py $REPORTS/matmul-performance-at.csv $REPORTS/gemm-at-onednn-report.csv --benchmark gemm-at --compiler onednn --param_cols "B,M,K,N" --tflops_col onednn-TFlops --hbm_col "onednn-GB/s" --tag $TAG
 
       - name: Run Triton GEMM (stream-k) kernel benchmark
         if: ${{ steps.install.outcome == 'success' && !cancelled() }}

.gitignore

@@ -6,11 +6,19 @@ build-*/
 python/build/
 python/dist/
 python/triton*.egg-info/
+python/*.whl
 
 python/triton/_C/*.pyd
 python/triton/_C/*.so
 python/triton/_C/*.dylib
 
+benchmarks/dist
+benchmarks/*.egg-info/
+benchmarks/**/*.so
+
+# Logs
+inductor_log/
+
 # Backends copied from submodules
 python/triton/backends/
 !python/triton/backends/__init__.py

.pre-commit-config.yaml

@@ -22,7 +22,7 @@ repos:
       - id: ruff
         files: '^python/.*'
         args: ["--fix", "--line-length", "120"]
-        stages: [commit, push, manual]
+        stages: [pre-commit, pre-push, manual]
         exclude: |
           (?x)(
             ^python/triton/runtime/.*|
@@ -35,14 +35,14 @@ repos:
     hooks:
       - id: yapf
         args: ["-p", "-i"]
-        stages: [commit, push, manual]
+        stages: [pre-commit, pre-push, manual]
         exclude: "python/test/unit/language/test_line_info.py"
 
   - repo: https://github.com/pre-commit/mirrors-clang-format
     rev: v16.0.6
     hooks:
       - id: clang-format
-        stages: [commit, push, manual]
+        stages: [pre-commit, pre-push, manual]
 
   # Expand YAML anchors in files used by github workflows, because github can't
   # do this itself.  This lets us use anchors, which avoids code duplication.
@@ -69,15 +69,15 @@ repos:
     - id: bandit
       files: '^(benchmarks|scripts|third_party/intel)/.*\.py$'
       args: ["-c", "bandit.yaml", "-s", "B404,B603,B607"]
-      stages: [commit, push, manual]
+      stages: [pre-commit, pre-push, manual]
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
     rev: v0.1.3
     hooks:
       - id: ruff
         files: '^(benchmarks|third_party/intel|scripts)/.*'
         args: ["--fix", "--line-length", "120"]
-        stages: [commit, push, manual]
+        stages: [pre-commit, pre-push, manual]
 
   - repo: https://github.com/pycqa/pylint
     rev: v3.2.6
@@ -105,7 +105,7 @@ repos:
           - --disable=too-many-locals
           - --disable=too-many-statements
           - --disable=too-many-arguments
-        stages: [commit, push, manual]
+        stages: [pre-commit, pre-push, manual]
 
       - id: pylint
         name: pylint for benchmarks
@@ -136,7 +136,7 @@ repos:
           - --disable=too-many-statements
           - --disable=too-many-arguments
           - --disable=fixme
-        stages: [commit, push, manual]
+        stages: [pre-commit, pre-push, manual]
 
 
 exclude: |

CMakeLists.txt

@@ -211,6 +211,7 @@ if(TRITON_BUILD_PYTHON_MODULE)
     MLIRSCFToControlFlow
     MLIRIndexToLLVM
     MLIRGPUToROCDLTransforms
+    MLIRUBToLLVM
 
     # LLVM
     LLVMPasses

benchmarks/CMakeLists.txt

@@ -10,9 +10,11 @@ if(NOT WIN32)
     list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 endif()
 
-find_package(Python3 COMPONENTS Interpreter)
+find_package(Python3 REQUIRED
+  COMPONENTS Development.Module)
 find_package(Torch REQUIRED)
 find_library(TORCH_PYTHON_LIBRARY torch_python PATH "${TORCH_INSTALL_PREFIX}/lib")
+find_package(XeTLALibrary REQUIRED)
 
 if(USE_IPEX)
   string(APPEND CMAKE_CXX_FLAGS " -DUSE_IPEX")

benchmarks/cmake/FindXeTLALibrary.cmake

@@ -3,13 +3,15 @@
 include(FetchContent)
 
 if (NOT XeTLALibrary_FOUND)
+    # TODO: switch ot FetchContent_MakeAvailable once XeTLA supports it
+    cmake_policy(SET CMP0169 OLD)
 
     set(XeTLALibrary_SOURCE_DIR
             "${CMAKE_CURRENT_BINARY_DIR}/XeTLALibrary")
     message(STATUS "XeTLALibrary is not specified. Will try to download
                   XeTLA library from https://github.com/intel/xetla into
                   ${XeTLALibrary_SOURCE_DIR}")
-    file(READ xetla-library.conf XeTLALibrary_TAG)
+    file(READ xetla_kernel/xetla-library.conf XeTLALibrary_TAG)
     # Strip the potential trailing newline from tag
     string(STRIP "${XeTLALibrary_TAG}" XeTLALibrary_TAG)
     FetchContent_Declare(xetla-library