Integrate internal changes

This PR moves the following internal changes to OSS,

commit 2986cac97888b5a9c1cd9064e4728cb38ca9dc45
Author: Sagar Shelke <[email protected]>

    [executor] Add complex type support to `ScalarValue`

    Previously, ScalarValue which represents scalar runtime value did not
    support complex type. This MR adds support for complex type by making storage
    union of real and complex data instaed of just real.

    MLIR tests are added via constant subgraph execution.

commit cf83a0d318b8035695d0b9fd24d578733632e253
Author: Christopher Bate <[email protected]>

    [compiler] Enable more `stablehlo.dot_general` to TensorRT using `tensorrt.einsum`

    Previously, we relied on canonicalization of `stablehlo.dot_general` to
    put all such contraction operations into a form that could be converted to
    `tensorrt.matrix_multiply`. Based on recent experiments, this can actually
    produce very inefficient TensorRT programs due to the number of reshapes
    and transpositions that must be inserted to coerce general
    `stablehlo.dot_general` into batched matrix multiplications. This change
    enables conversion of `stablehlo.dot_general` to `tensorrt.einsum`, and
    the pass and patterns now contain configurable parameters to control whether
    `tensorrt.einsum` is used as the primary method or only for fallback when
    conversion to `tensorrt.matrix_multiply` is not possible.

    A follow on change will revamp the Stablehlo preprocessing that we
    perform on 'stablehlo.dot_general' to avoid creating inefficient patterns
    and enable wider use of this pattern.

commit 528651ed1cd36c36376180c1c2232526ce972fef
Author: Christopher Bate <[email protected]>

    [compiler] Fix stablehlo-to-scf scalarization heuristics

    Fixes an issue where float tensors in the 'before' region of converted
    while loops where scalarized. The transform should only scalarize operands
    which are likely to be for-style induction variables.

commit 1d52e0a9e30dc104178c4761c1a24153abc7ea90
Author: Christopher Bate <[email protected]>

    [compiler] NFC: Drop dead code from StablehloToExecutableTask

commit f1c8d8c7cd860aedfe339d76ef7fb953baf9bd55
Author: Chris Bate <[email protected]>

    [compiler] Add `plan-promote-host-tensors-to-host-pinned` pass

    Adds a simple pass to promote "host" tensors to "host-pinned" tensors
    in common cases where we know a tensor will be transferred between host
    and device spaces. This pass runs after `plan-optimize-memory-spaces`
    since the former is sensitive to mismatching host spaces for patterns
    related to moving tranfers out of loops.

commit c27d56ea7a9661395e17fa895c610a79a92fa0c2
Author: Sagar Shelke <[email protected]>

    [executor] Handle elided dense resource elements attr during translation

    Translation to executable (which is flatbuffer) uses MLIR attr serialization to serialize
    `ElementsAttr`. However, this doesn't work when attr is elided dense resource and results
    in segfault. This MR handles this situation by replacing elided resource with `DenseElementsAttr`
    of all `one`s (`true` in case of boolean).

    IR with elided resource is usally seen only during testing of passes and not useful for
    e2e functional execution. Testing of `ExecuteConstantFoldableSubgraphs` pass is such case.
    Thus,  MLIR test cases for this pass are added.

commit 920a84e648833764563d3dc1de544a8f1b9f027e
Author: Chris Bate <[email protected]>

    [tensorrt] Fix TRT layer name generation function

    The TRT layer naming had some faulty logic that could cause the layer
    name to grow very large in the process to create a unique name. Fix
    the issue and use a static counter to reduce time spent in the loop.

commit ff0c5fa4bf5321ad0ce18579598c49f4b552fb37
Author: Christopher Bate <[email protected]>

    Further fixes to LIT configs

    Previously, we were setting `lit_config.parallelism_group` instead of
    `config.parallelism_group`. Apparently, the previous method does nothing,
    only `config.parallelism_group` has any effect.

commit d65c220b712c262992dbdf5a87fa3220a06bfb21
Author: Chris Bate <[email protected]>

    Update LIG test parallelism configs

    In more recent versions of TensorRT (10.11+ at least), the builder is taking a
    much larger amount of host memory. This can cause OOM when running the LIT test
    suites under their existing configurations.

    This change updates all LIT configs:

    - Make sure to use `%pick-one-gpu` in the LIT command line to ensure we
      stall if there are not enough GPU or host resources available. Add
      a hard limit that there must be at least 5GB of host memory available.

    - Update configurations to reduce the amount of estimated parallelism by
      increasing host memory requirements and reducing the amount of host
      memory to 50% for the purposes of the parallelism calculation.

    - Force all tests to use a common parallelism group unless otherwise
      specified in the test config.

commit 1f996f607640d81bf7137a4ed874b20c2a16cca2
Author: Christopher Bate <[email protected]>

    [compiler] Fix failure case in stablehlo-to-scf

    Fixes a failure case due to one of the recently introduced rewrites in
    `stablehlo-to-scf`.

commit 2779b632465fc3e840f5ce987f6233e824fe2ed3
Author: Christopher Bate <[email protected]>

    [compiler] Further improvements to plan bufferization pipeline

    - Split `plan-assign-memory-spaces` into three passes:
      - `plan-assign-memory-spaces`
      - `plan-optimize-memory-spaces`
      - `plan-materialize-explicit-transfers`
    - The last one is the only new code: `plan-materialize-explicit-transfers`
      converts `tensor.cast` ops that change the memory space encoding into
      explicit `bufferization.alloc_tensor` +
      `bufferization.materialize_in_destination` operations.
    - Improve handling of `bufferization.alloc_tensor` and optimization of
      `scf.for` iteration args in `plan-assign-memory-spaces`.
    - Improve handling of `tensor.reshape` in `plan-assign-memory-spaces`.
    - Fix handling of `tensor.reshape` when rewriting functions to be in
      DPS style in `plan-alloc-tensors`.

    This change also updates the LLVM dependencies in order to cherry-pick
    fix to the `tensor.reshape` bufferization interface that I merged
    upstream (llvm/llvm-project#128590).

    In addition, fix APInt assertions in `plan-execute-constant-foldable-subgraphs`.

commit 312170d8cbcd4c1fcf9cefdd628583e3dbdcc4f5
Author: Chris Bate <[email protected]>

    [compiler] Enable While-to-For conversion in Stablehlo-to-Scf pass

    Stablehlo only has one type of loop construct, `stablehlo.while`. The
    `while` loop can represent "for"-style loops as well, but if we only
    have `scf.while` loops after conversion to SCF, then we miss out on
    lot of potential optimizations which are rooted on `scf.for`.
    Experiments show that complicated JAX programs like the
    PhysicalIntelligence Pi0 model can benefit from converting `scf.while`
    to `scf.for` where possible. This improves opportunities for
    constant folding and makes analysis much easier to gauge the benefit
    of transforms like unrolling.

    This change adds some patterns to the Stablehlo-to-Scf pass to enable
    While-to-For conversion after the Stablehlo-to-Scf conversion. This
    transformation is combined with the Stablehlo-to-Scf conversion because
    the While-to-For patterns require first scalarizing block arguments of
    the While operation. The heuristics for which block arguments should
    be scalarized are implemented as control callbacks for the scalarization
    patterns. These callbacks need Stablehlo-specific logic, so it makes sense
    to test the combined conversion as a single pass. From the pass users'
    perspective, it gives the appearence of going directly from `stablehlo.while`
    to `scf.for`.

    The test cases are updated to cover the new patterns.

commit 425d19e749104354b5ea9e76e7509d029f9eac59
Author: Chris Bate <[email protected]>

    [compiler] Fix assign-memory-spaces pass to respect function-level constraints

    Fixes an issue where the `plan.memory_space` attribute on a function was
    not being respected when converting function signatures.

    MR: initialdl/mlir-tensorrt!2146

commit b612d5a22e7e3c4f08bf80fd504df5193b370bd3
Author: Chris Bate <[email protected]>

    [compiler] Update scf.while detensorization to increase flexibility

    In order to incorporate the upstream "uplift scf.while to scf.for"
    transformation as part of the `stablehlo-to-scf` conversion, we need to
    detensorize the operands of `scf.while` that are likely to correspond to
    the loop induction variable. This change refactors our existing 'scf.while'
    detensorization transformation to give more flexibility and control. The
    TensorKindAnalysis is no longer required in order to use the pattern(s).
    Detensorization of `after` and `before` arguments of `scf.while` are now
    controlled separately.

commit 3e21bf465b90e1eaaad872da40c305b70253cce0
Author: Chris Bate <[email protected]>

    [compiler] Improve handling of memory space constraints in the Plan dialect

    This commit improves the handling of memory space constraints in the
    Plan dialect. Constraints are now specified using a common attribute
    'plan.memory_space' that can be applied to functions or individual
    arguments/results. In addition, patterns in `plan-alloc-tensors`
    and `plan-assign-memory-spaces` are updated to avoid introducing
    unnecessary transfers between memory spaces.

commit 36a3b4a77242685e473817cb692a4010f690c0b3
Author: Chris Bate <[email protected]>

    [compiler] Add plan-buffer-results-to-out-params pass

    This change adds a new Plan dialect pass `plan-buffer-results-to-out-params`.
    This pass is based on the upstream Bufferization pass
    `buffer-results-to-out-params`, but it can handle a wider number of
    cases (such as promoting dynamic allocations) and uses alias analysis
    utilities to guard against failure cases that the upstream pass currently
    cannot handle. These improvements should eventually be upstreamed back
    to the Bufferization dialect.

commit 9e7127ca1e61be72b032a54d270a3da0d75639b2
Author: Chris Bate <[email protected]>

    [compiler] Update func conversion in host-to-emitc

    In the EmitC conversion/translation process, you can use `func.func`
    or `emitc.func` to define functions. Previously, we converted all
    `func.func` to `emitc.func`. However, `emitc.func` does not have a
    path for supporting multiple return values. Therefore, prefer use
    of type conversions on `func.func` instead of converting the entire
    op to `emitc.func`. Add tests to verify that we can support multiple
    return values.

commit 934db1f78ef3e7bedb67f1252b41ded7419010f8
Author: Chris Bate <[email protected]>

    [compiler] Fix two host-to-emitc bugs

    This change fixes two bugs exposed by new 'host-to-emitc' conversion
    testing:

    - The `!emitc.size_t` type does not have DataLayout information specified
      upstream. Therefore, to ensure that the type can be queried using
      DataLayout, we add a DataLayoutTypeInterface external model to the type.
      All queries are simply mapped to queries to the `index` type.

    - The upstream `func.call` conversion has a bug where it does not
      correctly convert the result types of the call operation, which
      can lead to a type mismatch for any type that does not have an
      identity conversion.

    Additional tests are added to `host-to-emitc`. Eventually the fixes for
    both these issues should be moved upstream.

commit 9d27f08ee4429f4ffbb72023babc193c7724a700
Author: Chris Bate <[email protected]>

    [common] Add Linalg-to-loops (on tensors) implementation and conversion pass

    Adds a ToLoopsOpInterface implementation and for Linalg operations. In addition,
    a conversion pass is added that converts ToLoopOpInterface operations to loops.

commit 3a419f120808eafc31f45516977ed6169b809ab9
Author: Chris Bate <[email protected]>

    NFC: Move ToLoopsOpInterface to 'mlir-tensorrt-common'

    Moves the ToLoopsOpInterface to the 'mlir-tensorrt-common' project. This
    is in preperation for enabling the ToLoopsOpInterface on LinalgOp (lowering
    while still using Tensor types) to replace the `convert-stablehlo-arith-to-scalar`
    pipeline.

    MR: initialdl/mlir-tensorrt!2137

commit 442bea12b763dd36fce864695f63896912438d87
Author: Christopher Bate <[email protected]>

    NFC: Fix formatting across several files

commit b2a65bc3e806aaa95d932af512cfa4750a9cbe4e
Author: Chris Bate <[email protected]>

    [executor] Introduce RuntimeSession "features" to control loading of runtime modules

    Previously, the RuntimeSession would always load all available runtime
    modules. This causes some inefficiences. For example, in certain integration
    tests for the Executor runtime, we don't use CUDA at all. However, because
    CUDA is still initialized by default, we would still require a GPU to
    be present just to run the integration test. Furthermore, some experimental
    modules (e.g. Lua cublas module) are not ready for "production" use and
    are only really invoked inside special integration tests.

    This change inroduces a notion of "features" to the RuntimeSession and
    RuntimeSessionOptions. A feature is just a string that identifies a
    particular runtime component. The particular semantic of a "feature" depends
    on the the actual runtime implementation. For example, for the
    LuaRuntimeSession, the feature names correspond to the available Lua
    "modules" (a module is just a group of C++ Lua extension functions),
    e.g. "core", "cuda", "tensorrt", etc.

    The RuntimeSessionOptions gains methods for enabling/disabling features.
    Certain features cause others to be added to the set automatically, e.g.
    "tensorrt" and "nccl" both require "cuda" to be added.

    The API is piped through all the way to the Python bindings to allow
    control of loaded modules at all levels. To preserve existing behavior,
    RuntimeSessions created from Python will load all available modules by
    default, but the `executor-runner|mlir-tensorrt-runner` tools now require
    features to be explicitly specified.

commit b90f8f345b2941e958f3a1cc5bcac21daebe783b
Author: Christopher Bate <[email protected]>

    NFC: Fix include guard for 'mlir-executor/Support/Status.h'

commit cdbe1f560483047291a30115a043a60bdce34d99
Author: Sagar Shelke <[email protected]>

    [compiler/lib] Add stablehlo composite to call pass to pre-processing pipeline

    This MR adds `StablehloLegalizeCompositeToCallPass` to the pre-processing
    pipeline.

    MLIR test is added.

commit 6ea3ab77aa2909cee11d08aa24543f247e8a24bf
Author: Chris Bate <[email protected]>

    [compiler] Add "default memory space" to ClusterKindAttrInterface

    Adds a new method to the ClusterKindAttrInterface so that backends can
    control the default tensor encoding (#plan.memory_space<..>) assigned by
    the `plan.assign-memory-spaces` pass at a function-scope level. In
    addition, we also allow an attribute to override the default space
    at function argument/results. This override mechnanism was previously
    lacking and will help resolve a long-standing issue where users cannot
    control the memory space of arguments/results reliably.

commit 0ea59238f5c280ab3ffbc340bb9aee7ed7bfbebb
Author: Christopher Bate <[email protected]>

    [compiler] Fix some issues related to pipeline extension mechanism

    The StablehloToExecutableTensorRTExtension had both 'disable' and
    an inherited 'disabled' member variable. Delete the inherited one
    such it should not have been introduced and was not bound to any
    option. Further, remove unused 'extensions' vector from
    CompilationTaskOptionsBase.

commit 372476d77fcaa399460965ab7bfc052f0e44c99f
Author: Christopher Bate <[email protected]>

    [executor] Fix ptrtoint and inttoptr op translation to Lua

    Previously, we could generate conflicting function types (due to pointer
    address space) when converting `executor.ptrtoint` and `executor.inttoptr`
    ops to opaque calls. Instead, defer the conversion to function call until
    the actual Lua translation point. At that point we can generate a
    function name without having to consider the pointer address space.

commit 75d18534fa67b452dd2253d6981bda6954bf1056
Author: Chris Bate <[email protected]>

    Introduce 'MLIRTensorRTCommmon' sub-project

    Certain targets need to be used across multiple sub-projects. For example,
    the 'TensorRTDynamicLoader' target is used in all sub-projects. In addition,
    the sub-projects need to be independently buildable. This change introduces
    another sub-project under the 'common' directory where shared code can be
    placed. This allows us to use `find_package` to declare the dependency, and
    downstream consumers to meet the requirement using any number of
    techniques to fullfill the 'find_package' call.

commit d7d8104087cf272bdd08f6330f27734754f0d71d
Author: Chris Bate <[email protected]>

    [compiler] Harden `stablehlo.constant` to `arith.constant` conversion

    There is a utility pass that runs in the stablehlo-to-executable pipeline
    that converts `stablehlo.constant` to `arith.constant`. This pass
    can temporarily create invalid IR due to `arith.constant` not supporting
    signful integer types. If the "verify-each" option is off,
    then the issue will not be caught since it happens to be self-correcting.
    However, the issue can still cause verification failures while debugging.
    This change fixes the issue by adding a `builtin.unrealized_conversion_cast`
    operation to bridge the type change between signless-and-signfull integer
    types.

commit a500de82a7bd70d6bfe32234719b4daa7cf32a8a
Author: Chris Bate <[email protected]>

    Integrate LLVM at f137c3d592e96330e450a8fd63ef7e8877fc1908

commit cd56aa6a511e2091fcd86106f20d27ff3673db75
Author: Christopher Bate <[email protected]>

    Fix build with BUILD_SHARED_LIBS=ON

    The new InferTensorValueRangeInterface was used without correctly
    specifying the library dependency the PlanIR and StablehloExtIR
    libraries.

commit cf1aff0ad0997947ab87485cfeec4595cb0285d7
Author: Sagar Shelke <[email protected]>

    [compiler] Maintain output order in TensorRT engine.

    For TensorRT engine conversion, first step in lowering a
    cluster containing TensorRT ops is created inline group op.
    Operands to the yield op (i.e. terminator) of inline group op
    are values from the cluster that are used outside the cluster.
    These values are collected by getting uses of each op (with
    `op->getUses()`) and checking if they are outside the cluster.
    However, this use order is not deterministic and sometimes
    it is desired to get yield results in a certian order.

    This MR makes the following changes,
    1. Add a function callback option named `ReorderRegionOpYieldValues`
    to `mlir::createRegionOpFromCluster` method. This callback function
    has signature `std::function<void(SetVector<Value> &yieldValues,
    SmallVectorImpl<Type> &yieldTypes)>` which takes cluster values used
    outside the cluster (in SetVector) and their types. By default this is
    set to nullptr.
    2. TensorRTToExecutable task is used in cases where a single `func.func`
    represents a single TensorRT engine. In this case, `ReorderRegionOpYieldValues`
    callback is implemented to make sure inline group op yield value order
    is same as func.func return values order.

    Valid MLIR test is added.

GitOrigin-RevId: 630a69d8e14506db43cfefe4be2c790f9352da4f

Author: Copybara Bot

mlir-tensorrt/CMakeLists.txt

@@ -190,6 +190,8 @@ if(MLIR_TRT_ENABLE_PYTHON)
   mlir_tensorrt_find_dlpack()
 endif()
 
+find_package(MLIRTensorRTCommon REQUIRED)
+
 #--------------------------------------------------
 # Diagnostics
 #--------------------------------------------------

mlir-tensorrt/DependencyProvider.cmake

@@ -17,7 +17,7 @@ if("${MLIR_TRT_USE_LLVM}" STREQUAL "prebuilt")
   set(MTRT_BUILD_LLVM_FROM_SOURCE OFF)
 endif()
 
-set(MLIR_TRT_LLVM_COMMIT "729416e586fba71b4f63d71b1b5c765aefbf200b")
+set(MLIR_TRT_LLVM_COMMIT "f137c3d592e96330e450a8fd63ef7e8877fc1908")
 
 set(mlir_patch_dir "${CMAKE_CURRENT_LIST_DIR}/build_tools/patches/mlir")
 
@@ -123,6 +123,54 @@ nv_register_package(
   ]]
 )
 
+#-------------------------------------------------------------------------------------
+# MLIRTensorRTCommon
+#
+# MLIRTensorRTCommon is a sub-project that contains components used across the
+# other sub-projects like MLIRExecutor and MLIRTensorRTDialect.
+#-------------------------------------------------------------------------------------
+
+nv_register_package(
+  NAME MLIRTensorRTCommon
+  SOURCE_DIR "${CMAKE_SOURCE_DIR}/common"
+)
+
+# -----------------------------------------------------------------------------
+# NVTX
+# -----------------------------------------------------------------------------
+
+nv_register_package(
+  NAME NVTX
+  GIT_REPOSITORY https://github.com/NVIDIA/NVTX.git
+  GIT_TAG v3.1.0
+  GIT_SHALLOW TRUE
+  SOURCE_SUBDIR c
+  EXCLUDE_FROM_ALL TRUE
+  DOWNLOAD_ONLY TRUE
+  POST_ADD_HOOK [[
+    if(NOT TARGET nvtx3-cpp)
+      add_library(nvtx3-cpp INTERFACE IMPORTED)
+      target_include_directories(nvtx3-cpp INTERFACE
+        "$<BUILD_INTERFACE:${NVTX_SOURCE_DIR}/c/include>")
+      # Ignore some warnings due to NVTX3 code style.
+      target_compile_options(nvtx3-cpp INTERFACE
+        -Wno-missing-braces)
+    endif()
+  ]]
+)
+
+#-------------------------------------------------------------------------------------
+# MLIRTensorRTCommon
+#
+# MLIRTensorRTCommon is a sub-project that contains components used across the
+# other sub-projects like MLIRExecutor and MLIRTensorRTDialect.
+#-------------------------------------------------------------------------------------
+
+nv_register_package(
+  NAME MLIRTensorRTCommon
+  SOURCE_DIR "${CMAKE_SOURCE_DIR}/common"
+)
+
 #-------------------------------------------------------------------------------------
 # MLIR-Executor
 #
@@ -164,7 +212,7 @@ nv_register_package(
   NAME torch_mlir
   GIT_REPOSITORY https://github.com/llvm/torch-mlir.git
   GIT_TAG 0bb263e99415d43255350d29263097b4980303bf
-  PATCHES 
+  PATCHES
     "build_tools/patches/torch_mlir/0001-cmake-Allow-finding-Stablehlo-via-find_package.patch"
     "build_tools/patches/torch_mlir/0002-Make-compatible-with-more-recent-Stablehlo-version.patch"
     "build_tools/patches/torch_mlir/0003-Fix-some-configuration-paths-in-LIT-cfg.patch"
@@ -202,7 +250,7 @@ macro(mtrt_provide_dependency method dep_name)
   endif()
 
   if("${dep_name}" MATCHES
-     "^(MLIRExecutor|MLIRTensorRTDialect|Stablehlo|torch_mlir)$")
+     "^(MLIRExecutor|MLIRTensorRTDialect|Stablehlo|torch_mlir|NVTX|MLIRTensorRTCommon)$")
     nv_add_package("${dep_name}")
     set("${dep_name}_FOUND" TRUE)
   endif()
@@ -230,6 +278,7 @@ macro(mtrt_provide_dependency method dep_name)
       find_package(LLVM ${ARGN} BYPASS_PROVIDER)
     endif()
   endif()
+
 endmacro()
 
 cmake_language(

mlir-tensorrt/build_tools/cmake/Dependencies.cmake

@@ -6,10 +6,16 @@ include(${CMAKE_CURRENT_LIST_DIR}/TensorRTDownloadURL.cmake)
 # expected version.
 #-------------------------------------------------------------------------------------
 macro(get_tensorrt_version nvinfer_version_file out_var)
-  file(STRINGS "${nvinfer_version_file}" VERSION_STRINGS REGEX "#define NV_TENSORRT_.*")
+  file(STRINGS "${nvinfer_version_file}" VERSION_STRINGS REGEX "#define (TRT_.+|NV_TENSORRT_.+) [0-9]+")
   foreach(TYPE MAJOR MINOR PATCH BUILD)
-    string(REGEX MATCH "NV_TENSORRT_${TYPE} [0-9]+" TRT_TYPE_STRING ${VERSION_STRINGS})
-    string(REGEX MATCH "[0-9]+" TRT_${TYPE} ${TRT_TYPE_STRING})
+    string(REGEX MATCH "(TRT_${TYPE}_ENTERPRISE|NV_TENSORRT_${TYPE}) [0-9]+" TRT_TYPE_STRING ${VERSION_STRINGS})
+    if("${TRT_TYPE_STRING}" STREQUAL "")
+      message(FATAL_ERROR "Failed to extract TensorRT ${TYPE} version from ${nvinfer_version_file}")
+    endif()
+    string(REGEX MATCH "[0-9]+" "TRT_${TYPE}" "${TRT_TYPE_STRING}")
+    if("TRT_${TYPE}" STREQUAL "")
+      message(FATAL_ERROR "Failed to extract TensorRT ${TYPE} version from ${nvinfer_version_file}")
+    endif()
   endforeach(TYPE)
   set("${out_var}" "${TRT_MAJOR}.${TRT_MINOR}.${TRT_PATCH}.${TRT_BUILD}")
 endmacro()
@@ -50,7 +56,7 @@ macro(configure_tensorrt_python_plugin_header)
   if(ARG_INSTALL_DIR)
     find_file(
       trt_python_plugin_header
-      NAMES plugin.h
+      NAMES NvInferPythonPlugin.h plugin.h
       HINTS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/python/include/impl
       PATHS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/python/include/impl
       REQUIRED
@@ -60,7 +66,7 @@ macro(configure_tensorrt_python_plugin_header)
   else()
     find_path(
       trt_python_plugin_header
-      NAMES plugin.h
+      NAMES NvInferPythonPlugin.h plugin.h
       REQUIRED
       NO_CACHE
     )
@@ -173,36 +179,6 @@ function(find_tensorrt)
     )
 endfunction()
 
-macro(configure_tensorrt_python_plugin_header)
-  if(ARG_INSTALL_DIR)
-    find_file(
-      trt_python_plugin_header
-      NAMES plugin.h
-      HINTS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/python/include/impl
-      PATHS ${ARG_INSTALL_DIR} ${ARG_INSTALL_DIR}/python/include/impl
-      REQUIRED
-      NO_CMAKE_PATH NO_DEFAULT_PATH
-      NO_CACHE
-    )
-  else()
-    find_path(
-      trt_python_plugin_header
-      NAMES plugin.h
-      REQUIRED
-      NO_CACHE
-    )
-  endif()
-  file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/include/nvinfer")
-  file(COPY_FILE "${trt_python_plugin_header}"
-    "${CMAKE_BINARY_DIR}/include/nvinfer/trt_plugin_python.h"
-    ONLY_IF_DIFFERENT
-    RESULT copy_result
-  )
-  if(copy_result)
-    message(FATAL_ERROR "failed to copy TensorRT QDP plugin header: ${copy_result}")
-  endif()
-endmacro()
-
 #-------------------------------------------------------------------------------------
 # Download and add DLPack to the build (header only)
 #-------------------------------------------------------------------------------------

mlir-tensorrt/common/CMakeLists.txt

@@ -0,0 +1,38 @@
+cmake_minimum_required(VERSION 3.25)
+project(mlir-tensorrt-common LANGUAGES CXX)
+
+# Depdendencies
+find_package(LLVM REQUIRED CONFIG)
+find_package(MLIR REQUIRED CONFIG)
+include(HandleLLVMOptions)
+include_directories(${LLVM_INCLUDE_DIRS})
+include_directories(${MLIR_INCLUDE_DIRS})
+
+if(MLIR_TRT_TARGET_TENSORRT)
+  find_package(TensorRT REQUIRED)
+endif()
+
+find_package(CUDAToolkit REQUIRED)
+
+set(MLIR_TENSORRT_COMMON_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
+set(MLIR_TENSORRT_COMMON_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
+
+include_directories(include ${CMAKE_CURRENT_BINARY_DIR}/include)
+
+add_library(MLIRTensorRTCommonIncludes INTERFACE)
+target_include_directories(MLIRTensorRTCommonIncludes INTERFACE
+  "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>"
+  "$<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/include>"
+)
+
+add_subdirectory(include/mlir-tensorrt-common)
+add_subdirectory(lib)
+
+install(TARGETS MLIRTensorRTCommonIncludes
+  EXPORT MLIRTensorRTCommonTargets
+  RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
+  LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
+  INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
+)
+

mlir-tensorrt/common/include/mlir-tensorrt-common/CMakeLists.txt

@@ -1,6 +1,6 @@
-//===- LuaRegistration.h ----------------------------------------*- C++ -*-===//
+//===- Passes.h -------------------------------------------------*- C++ -*-===//
 //
-// SPDX-FileCopyrightText: Copyright 2024 NVIDIA CORPORATION & AFFILIATES.
+// SPDX-FileCopyrightText: Copyright 2025 NVIDIA CORPORATION & AFFILIATES.
 // All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 //
@@ -18,22 +18,22 @@
 //
 //===----------------------------------------------------------------------===//
 ///
-/// Registration for the Lua runtime methods.
+/// This file contains the declarations for the common conversion passes.
 ///
 //===----------------------------------------------------------------------===//
+#ifndef MLIR_TENSORRT_COMMON_CONVERSION_PASSES
+#define MLIR_TENSORRT_COMMON_CONVERSION_PASSES
 
-#include "mlir-executor/Runtime/API/API.h"
+#include "mlir/Pass/Pass.h"
+#include <memory>
 
-struct lua_State;
-
-namespace mlirtrt::runtime {
-/// Register various external functions with the given Lua state using a
-/// directly specified device number, total device count, and a pre-determined
-/// NCCL uuid.
-void registerLuaRuntimeMethods(lua_State *state,
-                               const RuntimeSessionOptions &options,
-                               PinnedMemoryAllocator *pinnedMemoryAllocator,
-                               AllocTracker *allocTracker,
-                               ResourceTracker *resourceTracker);
+//===----------------------------------------------------------------------===//
+// Add Tablegen'd pass declarations and registration methods.
+//===----------------------------------------------------------------------===//
+namespace mlir {
+#define GEN_PASS_DECL
+#define GEN_PASS_REGISTRATION
+#include "mlir-tensorrt-common/Conversion/Passes.h.inc"
+} // namespace mlir
 
-} // namespace mlirtrt::runtime
+#endif // MLIR_TENSORRT_COMMON_CONVERSION_PASSES

mlir-tensorrt/executor/include/mlir-executor/Runtime/Backend/Lua/LuaRegistration.h renamed to mlir-tensorrt/common/include/mlir-tensorrt-common/Conversion/Passes.h

@@ -0,0 +1,18 @@
+#ifndef MLIR_TENSORRT_COMMON_CONVERSION_PASSES
+#define MLIR_TENSORRT_COMMON_CONVERSION_PASSES
+
+include "mlir/Pass/PassBase.td"
+
+def ConvertToLoops : Pass<"convert-to-loops"> {
+  let summary = "Convert a LoopLikeOpInterface to loops";
+  let description = [{
+    This pass converts a LoopLikeOpInterface to loops.
+  }];
+
+  let dependentDialects = [
+    "::mlir::tensor::TensorDialect",
+    "::mlir::scf::SCFDialect",
+  ];
+}
+
+#endif // MLIR_TENSORRT_COMMON_CONVERSION_PASSES

mlir-tensorrt/common/include/mlir-tensorrt-common/Conversion/Passes.td

@@ -0,0 +1,36 @@
+//===- DataLayoutImpl.h -----------------------------------------*- C++ -*-===//
+//
+// SPDX-FileCopyrightText: Copyright 2025 NVIDIA CORPORATION & AFFILIATES.
+// All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+//===----------------------------------------------------------------------===//
+///
+/// This file contains the declarations for the DataLayout extensions to
+/// the EmitC dialect.
+/// TODO: These interfaces should be upstreamed to the EmitC dialect so that
+/// external models are not required.
+///
+//===----------------------------------------------------------------------===//
+#ifndef MLIR_TENSORRT_COMMON_DIALECT_EMITCEXT_IR_DATALAYOUTIMPL_H
+#define MLIR_TENSORRT_COMMON_DIALECT_EMITCEXT_IR_DATALAYOUTIMPL_H
+
+#include "mlir/IR/DialectRegistry.h"
+
+namespace mlir::emitc_ext {
+void registerDataLayoutInterfaceExternalModels(DialectRegistry &registry);
+}
+
+#endif // MLIR_TENSORRT_COMMON_DIALECT_EMITCEXT_IR_DATALAYOUTIMPL_H

mlir-tensorrt/common/include/mlir-tensorrt-common/Dialect/EmitCExt/IR/DataLayoutImpl.h

@@ -0,0 +1,53 @@
+//===- ToLoopsOpInterfaceImpl.h ---------------------------------*- C++ -*-===//
+//
+// SPDX-FileCopyrightText: Copyright 2025 NVIDIA CORPORATION & AFFILIATES.
+// All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+//===----------------------------------------------------------------------===//
+///
+/// This file contains the declarations for the ToLoopsOpInterface extensions to
+/// the Linalg dialect.
+///
+//===----------------------------------------------------------------------===//
+#ifndef MLIR_TENSORRT_COMMON_DIALECT_LINALGEXT_TRANSFORMS_TOLOOPSOPINTERFACEIMPL
+#define MLIR_TENSORRT_COMMON_DIALECT_LINALGEXT_TRANSFORMS_TOLOOPSOPINTERFACEIMPL
+
+#include "mlir-tensorrt-common/Interfaces/ToLoopsOpInterface.h"
+#include "mlir/IR/DialectRegistry.h"
+
+namespace mlir::linalg {
+class LinalgOp;
+}
+
+namespace mlir::scf {
+class ForOp;
+}
+
+namespace mlir::linalg_ext {
+
+/// Register the ToLoopsOpInterface external models for GenericOp. For other
+/// kinds of operations that are LinalgOps, we don't register an external model
+/// because there are so many; instead use the below function to perform
+/// conversion.
+void registerToLoopsOpInterfaceExternalModels(DialectRegistry &registry);
+
+/// Convert a LinalgOp (on tensors) to SCF loops.
+FailureOr<SmallVector<Operation *>>
+convertLinalgOpToLoops(RewriterBase &rewriter, linalg::LinalgOp op);
+
+} // namespace mlir::linalg_ext
+
+#endif // MLIR_TENSORRT_COMMON_DIALECT_LINALGEXT_TRANSFORMS_TOLOOPSOPINTERFACEIMPL