Merge branch 'main' into dev/daiki/fabric_pkt_hdr_optim

tenstorrent · Feb 18, 2025 · f7b8c96 · f7b8c96
2 parents a163057 + ed210e7
commit f7b8c96
Show file tree

Hide file tree

Showing 85 changed files with 2,271 additions and 2,596 deletions.
diff --git a/.github/workflows/build-artifact.yaml b/.github/workflows/build-artifact.yaml
@@ -54,14 +54,16 @@ on:
       #ci-test-docker-image:
       #  description: "Docker tag for the CI Test Docker image for testing TT-Metalium et al"
       #  value: ${{ jobs.build-docker-image.outputs.ci-test-tag }}
+      packages-artifact-name:
+        description: "Name to give download-artifact to get the packages"
+        value: ${{ jobs.build-artifact.outputs.packages-artifact-name }}
       build-artifact-name:
         description: "Name of the published build artifact"
         value: ${{ jobs.build-artifact.outputs.build_artifact_name }}
       wheel-artifact-name:
         description: "Name of the published wheel artifact"
         value: ${{ jobs.build-artifact.outputs.wheel_artifact_name }}
 
-
   workflow_dispatch:
     inputs:
       build-type:
@@ -114,6 +116,7 @@ jobs:
       - build
       - in-service
     outputs:
+      packages-artifact-name: ${{ steps.set-artifact-name.outputs.name }}
       build_artifact_name: ${{ steps.set_build_artifact_name.outputs.build_artifact_name }}
       wheel_artifact_name: ${{ steps.set_wheel_artifact_name.outputs.wheel_artifact_name }}
     container:
@@ -149,11 +152,22 @@ jobs:
             exit 1
           fi
 
+      - name: Set artifact name
+        id: set-artifact-name
+        run: |
+          TOOLCHAIN="${{ inputs.toolchain }}"
+          TOOLCHAIN_CLEANED=$(echo "$TOOLCHAIN" | sed -E 's/^cmake\///; s/-toolchain\.cmake$//')
+          ARTIFACT_NAME="packages-${{ inputs.distro }}-${{ inputs.version }}-${{ inputs.architecture }}-${{ inputs.build-type }}-${TOOLCHAIN_CLEANED}${{ (inputs.tracy && '_profiler') || '' }}"
+
+          echo "name=$ARTIFACT_NAME" >> "$GITHUB_OUTPUT"
+          echo "ARTIFACT_NAME=$ARTIFACT_NAME" >> "$GITHUB_ENV"
+
       - name: ⬇️ Checkout
         uses: actions/checkout@v4
         with:
-          fetch-depth: 0
           submodules: recursive
+          fetch-depth: 500 # Need enough history for `git describe`
+          fetch-tags: true # Need tags for `git describe`
           path: docker-job # Here be dragons; keep it scoped to our desired volume, yet must be under github.workspace and be sure to clean up at the end
 
       - name: Sanity check
@@ -194,9 +208,9 @@ jobs:
           nice -19 cmake --build build --target install
 
       - name: 📦 Package
-        if: false # Packaging coming later
         run: |
-          nice -19 cmake --build $build_dir --target package
+          nice -n 19 cmake --build build --target package
+          ls -1sh build/*.deb build/*.ddeb
 
       - name: 🐍 Build wheel
         if: ${{ inputs.build-wheel }}
@@ -210,6 +224,16 @@ jobs:
           ccache -s >> $GITHUB_STEP_SUMMARY
           echo '```' >> $GITHUB_STEP_SUMMARY
 
+      - name: ☁️ Upload packages
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ env.ARTIFACT_NAME }}
+          path: |
+            /work/build/*.deb
+            /work/build/*.ddeb
+          compression-level: 0
+          if-no-files-found: error
+
       - name: Set wheel artifact name
         id: set_wheel_artifact_name
         run: |

diff --git a/.github/workflows/cpp-post-commit.yaml b/.github/workflows/cpp-post-commit.yaml
@@ -62,11 +62,9 @@ jobs:
           {name: eth, cmd: "./build/test/tt_metal/unit_tests_eth_${{ inputs.arch }}"},
           {name: llk, cmd: "./build/test/tt_metal/unit_tests_llk"},
           {name: stl, cmd: "./build/test/tt_metal/unit_tests_stl"},
-          {name: distributed, cmd: "./build/test/tt_metal/distributed/distributed_unit_tests_${{ inputs.arch }} --gtest_filter=MeshDeviceSuite.*"},
-
+          {name: distributed, cmd: "./build/test/tt_metal/distributed/distributed_unit_tests_${{ inputs.arch }}"},
           {name: lightmetal, cmd: "./build/test/tt_metal/unit_tests_lightmetal"},
           {name: dispatch multicmd queue, cmd: "TT_METAL_GTEST_NUM_HW_CQS=2 ./build/test/tt_metal/unit_tests_dispatch_${{ inputs.arch }} --gtest_filter=MultiCommandQueue*Fixture.*"},
-
           {name: ttnn cpp unit tests, cmd: ./build/test/ttnn/unit_tests_ttnn},
           {name: ttnn ccl cpp unit tests, cmd: ./build/test/ttnn/unit_tests_ttnn_ccl},
           {name: ttnn tensor cpp unit tests, cmd: ./build/test/ttnn/unit_tests_ttnn_tensor},

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -18,6 +18,7 @@ project(
     DESCRIPTION "Tenstorrent Metalium"
     HOMEPAGE_URL "https://github.com/tenstorrent/tt-metal"
     LANGUAGES
+        C # Some of the jit-build files are plain C
         CXX
 )
 message(STATUS "Metalium version: ${PROJECT_VERSION}")
@@ -242,6 +243,11 @@ else()
     add_compile_definitions(TT_ENABLE_LIGHT_METAL_TRACE=0)
 endif()
 
+include(GNUInstallDirs)
+# GNUInstallDirs takes PROJECT_DIR verbatim, but directories should always be lowercase
+string(TOLOWER ${PROJECT_NAME} PROJECT_NAME_LOWER)
+string(REPLACE ${PROJECT_NAME} ${PROJECT_NAME_LOWER} CMAKE_INSTALL_DOCDIR ${CMAKE_INSTALL_DOCDIR})
+
 if(ENABLE_CODE_TIMERS)
     add_compile_definitions(TT_ENABLE_CODE_TIMERS)
 endif()
@@ -265,7 +271,6 @@ endif()
 # For top level install: cmake --build build --target install  or  make/ninja install -C build
 ############################################################################################################################
 # Install for build artifacts that will upload build/lib
-include(GNUInstallDirs)
 
 install(
     TARGETS
@@ -274,7 +279,7 @@ install(
         DESTINATION ${CMAKE_INSTALL_LIBDIR}
     LIBRARY
         DESTINATION ${CMAKE_INSTALL_LIBDIR}
-        COMPONENT dev
+        COMPONENT tar
 )
 install(
     TARGETS
@@ -283,7 +288,7 @@ install(
         DESTINATION ${CMAKE_INSTALL_LIBDIR}
     LIBRARY
         DESTINATION ${CMAKE_INSTALL_LIBDIR}
-        COMPONENT dev
+        COMPONENT tar
 )
 install(
     TARGETS
@@ -292,7 +297,7 @@ install(
         DESTINATION ${CMAKE_INSTALL_LIBDIR}
     LIBRARY
         DESTINATION ${CMAKE_INSTALL_LIBDIR}
-        COMPONENT dev
+        COMPONENT tar
 )
 if(WITH_PYTHON_BINDINGS)
     # Install .so into src files for pybinds implementation
@@ -306,6 +311,15 @@ if(WITH_PYTHON_BINDINGS)
     )
 endif()
 
+# FIXME(17578): figure out what bits we actually need to ship and omit the rest
+install(
+    DIRECTORY
+        runtime
+    DESTINATION ${CMAKE_INSTALL_LIBEXECDIR}/tt-metalium
+    USE_SOURCE_PERMISSIONS
+    COMPONENT jit-build
+)
+
 # Custom clean target for `built` folder for when new kernel changes are pulled
 add_custom_target(
     clean-built

diff --git a/CODEOWNERS b/CODEOWNERS
@@ -176,10 +176,12 @@ models/demos/t3000/mixtral8x7b @yieldthought @mtairum @uaydonat
 models/demos/tg/llama3_70b @cglagovichTT @uaydonat @johanna-rock-tt @djordje-tt @kpaigwar
 models/demos/tg/falcon7b @skhorasganiTT @djordje-tt @uaydonat
 models/demos/grayskull @uaydonat
-models/demos/yolov4 @dvartaniansTT @tenstorrent/metalium-developers-convolutions
-models/demos/wormhole/yolov4 @dvartaniansTT @tenstorrent/metalium-developers-convolutions
+models/demos/yolov4 @dvartaniansTT @mbahnasTT @tenstorrent/metalium-developers-convolutions
+models/demos/wormhole/yolov4 @dvartaniansTT @mbahnasTT @tenstorrent/metalium-developers-convolutions
 models/demos/**/*resnet*  @tt-aho @tenstorrent/metalium-developers-convolutions
 models/experimental/functional_unet @esmalTT @uaydonat @tenstorrent/metalium-developers-convolutions
+models/experimental/functional_vit @mbahnasTT @uaydonat
+models/demos/segformer @mbahnasTT @uaydonat @tenstorrent/metalium-developers-convolutions
 models/perf/ @uaydonat
 models/perf/perf_report.py @yieldthought @uaydonat
 models/perf/benchmarking_utils.py @skhorasganiTT

diff --git a/cmake/packaging.cmake b/cmake/packaging.cmake
@@ -1,5 +1,56 @@
-set(CPACK_GENERATOR "DEB")
-set(CPACK_DEBIAN_PACKAGE_MAINTAINER "[email protected]")
-#set(CPACK_DEBIAN_PACKAGE_DEPENDS "")
+set(CPACK_GENERATOR DEB)
+set(CPACK_PACKAGE_CONTACT "[email protected]")
+set(CMAKE_PROJECT_HOMEPAGE_URL "https://tenstorrent.com")
+set(CPACK_PACKAGE_NAME tt)
+
+set(CPACK_COMPONENT_METALIUM_DESCRIPTION "TT-Metalium runtime library")
+set(CPACK_DEBIAN_METALIUM_PACKAGE_SECTION "libs")
+
+set(CPACK_DEB_COMPONENT_INSTALL YES)
+set(CPACK_DEBIAN_PACKAGE_VERSION "${VERSION_DEB}")
+set(CPACK_DEBIAN_FILE_NAME DEB-DEFAULT)
+
+set(CPACK_DEBIAN_PACKAGE_CONTROL_STRICT_PERMISSION TRUE)
+# set(CPACK_DEBIAN_DEBUGINFO_PACKAGE TRUE)
+set(CPACK_DEBIAN_METALIUM_DEBUGINFO_PACKAGE TRUE)
+set(CPACK_DEBIAN_JIT-BUILD_DEBUGINFO_PACKAGE FALSE) # Some binaries don't have a Build ID; we cannot split dbgsyms
+
+set(CPACK_INSTALL_DEFAULT_DIRECTORY_PERMISSIONS
+    OWNER_READ
+    OWNER_WRITE
+    OWNER_EXECUTE
+    GROUP_READ
+    GROUP_EXECUTE
+    WORLD_READ
+    WORLD_EXECUTE
+)
+
+set(CPACK_DEBIAN_ENABLE_COMPONENT_DEPENDS TRUE)
+set(CPACK_DEBIAN_PACKAGE_SHLIBDEPS FALSE)
+
+get_cmake_property(CPACK_COMPONENTS_ALL COMPONENTS)
+list(
+    REMOVE_ITEM
+    CPACK_COMPONENTS_ALL
+    umd-dev # FIXME: -dev packages will come later
+    tt_pybinds # Wow this one is big!
+    tar # TODO: Remove that tarball entirely
+    # Deps that define install targets that we can't (or haven't) disabled
+    msgpack-cxx
+    Headers
+    Library
+    Unspecified # TODO: audit if there's anything we need to ship here
+)
+
+# Logically we should ship jit-build with metalium-runtime, but jit-build fails to split dbgsyms for now (lacking a Build ID on the binaries)
+cpack_add_component(jit-build GROUP metalium-jit)
+
+cpack_add_component(metalium-runtime GROUP metalium)
+cpack_add_component(umd-runtime GROUP metalium)
+cpack_add_component(dev GROUP metalium) # FIXME: delete this line when we bump UMD submodule
+cpack_add_component_group(metalium)
+
+cpack_add_component(gtest GROUP metalium-validation)
+cpack_add_component_group(metalium-validation)
 
 include(CPack)
diff --git a/cmake/version.cmake b/cmake/version.cmake
@@ -74,20 +74,25 @@ function(ParseGitDescribe)
     endif()
 
     set(VERSION_FULL "${VERSION_NUMERIC}")
+    set(VERSION_DEB "${VERSION_NUMERIC}")
     if(VERSION_STATUS)
         string(APPEND VERSION_FULL "-${VERSION_STATUS}")
+        string(APPEND VERSION_DEB "~${VERSION_STATUS}") # Debian versioning uses a ~ for "less than blank"
     endif()
     if(VERSION_COMMIT_COUNT)
         string(APPEND VERSION_FULL "+${VERSION_COMMIT_COUNT}.${VERSION_HASH}")
+        string(APPEND VERSION_DEB "+${VERSION_COMMIT_COUNT}.${VERSION_HASH}")
     endif()
     if(VERSION_DIRTY)
         string(APPEND VERSION_FULL "+m")
+        string(APPEND VERSION_DEB "+m")
     endif()
 
     message(STATUS "Version: ${VERSION_FULL}")
 
     # Output variables
     set(VERSION_FULL "${VERSION_FULL}" PARENT_SCOPE)
+    set(VERSION_DEB "${VERSION_DEB}" PARENT_SCOPE)
     set(VERSION_NUMERIC "${VERSION_NUMERIC}" PARENT_SCOPE)
     set(VERSION_HASH "${VERSION_HASH}" PARENT_SCOPE)
 endfunction()
diff --git a/dockerfile/Dockerfile b/dockerfile/Dockerfile
@@ -202,6 +202,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     clang-tidy-17 \
     curl \
     dialog \
+    file \
     graphviz \
     jq \
     pandoc \

diff --git a/models/demos/llama3/tt/llama_attention.py b/models/demos/llama3/tt/llama_attention.py
@@ -206,7 +206,7 @@ def __init__(
         pt_wo = self.state_dict[f"{wo_str}.weight"].transpose(-1, -2).unsqueeze(0).unsqueeze(0)
 
         wo_mem_config = configuration.create_dram_sharded_mem_config(
-            configuration.dim // configuration.num_devices, configuration.dim
+            (configuration.n_heads * configuration.head_dim) // configuration.num_devices, configuration.dim
         )
 
         self.wo = ttnn.as_tensor(

diff --git a/tests/scripts/run_tt_eager.py b/tests/scripts/run_tt_eager.py
@@ -33,7 +33,6 @@
     TestEntry("tt_eager/tests/ops/test_eltwise_binary_op", "ops/test_eltwise_binary_op"),
     TestEntry("tt_eager/tests/ops/test_bcast_op", "ops/test_bcast_op"),
     TestEntry("tt_eager/tests/ops/test_sliding_window_ops", "ops/test_sliding_window_ops"),
-    TestEntry("tt_eager/tests/ops/test_conv_prepare_weights_and_biases", "ops/test_conv_prepare_weights_and_biases"),
     TestEntry("tt_eager/tests/ops/test_bmm_op", "ops/test_bmm_op"),
     void_for_bh(void_for_whb0(TestEntry("tt_eager/tests/ops/test_eltwise_unary_op", "ops/test_eltwise_unary_op"))),
     TestEntry("tt_eager/tests/ops/test_layernorm_op", "ops/test_layernorm_op"),

diff --git a/tests/sweep_framework/sweeps/conv2d/short/conv2d_short_sweep.py b/tests/sweep_framework/sweeps/conv2d/short/conv2d_short_sweep.py
@@ -1622,8 +1622,6 @@ def test_conv2d_localrun(device, input_spec):
     [1, 768, 3, 224, 224, 32, 32, 32, 32, 0, 0, 1, 1, 1, False],  # 1460
     [1, 768, 3, 224, 224, 32, 32, 32, 32, 0, 0, 1, 1, 1, True],  # 1461
     [1, 768, 3, 384, 512, 32, 32, 32, 32, 0, 0, 1, 1, 1, True],  # 1464
-    [1, 64, 3, 720, 1280, 7, 7, 2, 2, 3, 3, 1, 1, 1, False],  # 1471
-    [1, 64, 3, 800, 1088, 7, 7, 2, 2, 3, 3, 1, 1, 1, False],  # 1472
     [1, 1, 64, 480, 640, 3, 3, 1, 1, 1, 1, 1, 1, 1, True],  # 1495
     [1, 64, 64, 480, 640, 3, 3, 1, 1, 1, 1, 1, 1, 1, True],  # 1496
 ]

diff --git a/tests/tt_eager/CMakeLists.txt b/tests/tt_eager/CMakeLists.txt
@@ -13,7 +13,6 @@ set(TT_EAGER_TESTS_OPS
     ops/test_sfpu.cpp
     ops/test_sliding_window_ops.cpp
     ops/test_fold_op.cpp
-    ops/test_conv_prepare_weights_and_biases.cpp
 )
 
 set(TT_EAGER_TESTS_TENSORS