Merge branch 'pytorch:main' into pr_print_tosa_deserialized

Author: freddan80

CMakeLists.txt

@@ -77,6 +77,24 @@ if(NOT EXECUTORCH_ENABLE_LOGGING)
   add_definitions(-DET_LOG_ENABLED=0)
 endif()
 
+# Configure log level. Must be one of debug, info, error, fatal.
+set(EXECUTORCH_LOG_LEVEL "Info" CACHE STRING
+      "Build with the given ET_MIN_LOG_LEVEL value")
+string(TOLOWER "${EXECUTORCH_LOG_LEVEL}" LOG_LEVEL_LOWER)
+if(LOG_LEVEL_LOWER STREQUAL "debug")
+  add_definitions(-DET_MIN_LOG_LEVEL=Debug)
+elseif(LOG_LEVEL_LOWER STREQUAL "info")
+  add_definitions(-DET_MIN_LOG_LEVEL=Info)
+elseif(LOG_LEVEL_LOWER STREQUAL "error")
+  add_definitions(-DET_MIN_LOG_LEVEL=Error)
+elseif(LOG_LEVEL_LOWER STREQUAL "fatal")
+  add_definitions(-DET_MIN_LOG_LEVEL=Fatal)
+else()
+  message(SEND_ERROR
+      "Unknown log level \"${EXECUTORCH_LOG_LEVEL}\". Expected one of Debug, " +
+      "Info, Error, or Fatal.")
+endif()
+
 option(EXECUTORCH_ENABLE_PROGRAM_VERIFICATION
        "Build with ET_ENABLE_PROGRAM_VERIFICATION"
        ${_default_release_disabled_options})

backends/vulkan/runtime/graph/ops/glsl/binary_op.glsl

@@ -52,21 +52,21 @@ void main() {
   }
 
   ivec4 in_coord = out_coord_to_in_coord(coord, in_sizes.data);
-  vec4 in_texel = texelFetch(
+  ${VEC4_T[DTYPE]} in_texel = ${VEC4_T[DTYPE]}(texelFetch(
     image_in,
     COORD_TO_POS_${PACKING}(in_coord, in_sizes.data),
-    0);
+    0));
 
   ivec4 other_coord = out_coord_to_in_coord(coord, other_sizes.data);
-  vec4 other_texel = texelFetch(
+  ${VEC4_T[DTYPE]} other_texel = ${VEC4_T[DTYPE]}(texelFetch(
     image_other,
     COORD_TO_POS_${PACKING}(other_coord, other_sizes.data),
-    0);
+    0));
 
   // Detect broadcasting
   if (PACKED_DIM_${PACKING}(other_sizes.data) < PACKED_DIM_${PACKING}(in_sizes.data)) {
     other_texel = other_texel.xxxx;
   }
 
-  imageStore(image_out, pos, OP(in_texel, other_texel, alpha.data));
+  imageStore(image_out, pos, ${VEC4_T[DTYPE]}(OP(in_texel, other_texel, alpha.data)));
 }

backends/vulkan/runtime/graph/ops/glsl/binary_op.yaml

@@ -23,6 +23,8 @@ binary_op:
         SUFFIX: half
       - VALUE: float
         SUFFIX: float
+      - VALUE: int
+        SUFFIX: int
   shader_variants:
     - NAME: binary_add
     - NAME: binary_sub

build/Utils.cmake

@@ -34,6 +34,8 @@ function(executorch_print_configuration_summary)
   message(STATUS "  FLATC_EXECUTABLE              : ${FLATC_EXECUTABLE}")
   message(
     STATUS "  EXECUTORCH_ENABLE_LOGGING     : ${EXECUTORCH_ENABLE_LOGGING}")
+  message(
+    STATUS "  EXECUTORCH_LOG_LEVEL          : ${EXECUTORCH_LOG_LEVEL}")
   message(STATUS "  EXECUTORCH_ENABLE_PROGRAM_VERIFICATION : "
                  "${EXECUTORCH_ENABLE_PROGRAM_VERIFICATION}")
   message(
@@ -46,6 +48,8 @@ function(executorch_print_configuration_summary)
   )
   message(
     STATUS "  REGISTER_EXAMPLE_CUSTOM_OPS   : ${REGISTER_EXAMPLE_CUSTOM_OPS}")
+  message(STATUS "  EXECUTORCH_BUILD_EXTENSION_AOT_UTIL : "
+                 "${EXECUTORCH_BUILD_EXTENSION_AOT_UTIL}")
   message(STATUS "  EXECUTORCH_BUILD_EXTENSION_DATA_LOADER : "
                  "${EXECUTORCH_BUILD_EXTENSION_DATA_LOADER}")
   message(STATUS "  EXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL : "

examples/models/llama2/export_llama_lib.py

@@ -396,6 +396,7 @@ def build_args_parser() -> argparse.ArgumentParser:
     parser.add_argument("-X", "--xnnpack", action="store_true")
     parser.add_argument("-V", "--vulkan", action="store_true")
     parser.add_argument("--mps", action="store_true")
+    parser.add_argument("--coreml", action="store_true")
 
     parser.add_argument(
         "--expand_rope_table",
@@ -576,6 +577,39 @@ def _export_llama(modelname, args) -> str:  # noqa: C901
         partitioners[MPSPartitioner.__name__] = MPSPartitioner(compile_specs)
         modelname = f"mps_{modelname}"
 
+    if args.coreml:
+        assert (
+            args.use_kv_cache is True
+        ), "CoreML backend currently only supports static shape and use_kv_cache=True is the only way to support it at the moment"
+        try:
+            # pyre-ignore: Undefined import [21]: Could not find a module corresponding to import `executorch.backends.apple.coreml.partition.coreml_partitioner`.
+            import coremltools as ct
+
+            # pyre-ignore: Undefined import [21]: Could not find a module corresponding to import `executorch.backends.apple.coreml.compiler`
+            from executorch.backends.apple.coreml.compiler import CoreMLBackend
+
+            # pyre-ignore: Undefined import [21]: Could not find a module corresponding to import `executorch.backends.apple.coreml.partition.coreml_partitioner`
+            from executorch.backends.apple.coreml.partition.coreml_partitioner import (
+                CoreMLPartitioner,
+            )
+        except ImportError:
+            raise ImportError(
+                "Please install the CoreML backend follwing https://pytorch.org/executorch/main/build-run-coreml.html"
+            )
+
+        # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `apple`.
+        compile_specs = CoreMLBackend.generate_compile_specs(
+            compute_precision=ct.precision(ct.precision.FLOAT16.value),
+            compute_unit=ct.ComputeUnit[ct.ComputeUnit.ALL.name.upper()],
+            # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `apple`
+            model_type=CoreMLBackend.MODEL_TYPE.MODEL,
+        )
+        # pyre-ignore: Undefined attribute [16]: Module `executorch.backends` has no attribute `apple`
+        partitioners[CoreMLPartitioner.__name__] = CoreMLPartitioner(
+            skip_ops_for_coreml_delegation=None, compile_specs=compile_specs
+        )
+        modelname = f"coreml_{modelname}"
+
     if args.generate_etrecord:
         if not builder_exported_to_edge.edge_manager:
             raise ValueError("Unable to generate etrecord due to missing edge manager.")