2024-12-14 nightly release (3fcf0bd)

.gitignore

@@ -1,5 +1,6 @@
 .hypothesis
 buck-out/
+buck2-bin/
 cmake-out*
 .DS_Store
 cmake-android-out/

CMakeLists.txt

@@ -257,6 +257,17 @@ if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
   set(EXECUTORCH_BUILD_KERNELS_OPTIMIZED ON)
 endif()
 
+if(NOT DEFINED FXDIV_SOURCE_DIR)
+  set(ORIGINAL_CMAKE_POSITION_INDEPENDENT_CODE_FLAG
+    ${CMAKE_POSITION_INDEPENDENT_CODE}
+  )
+  set(FXDIV_SOURCE_DIR "backends/xnnpack/third-party/FXdiv")
+  add_subdirectory("${FXDIV_SOURCE_DIR}")
+  set(CMAKE_POSITION_INDEPENDENT_CODE
+      ${ORIGINAL_CMAKE_POSITION_INDEPENDENT_CODE_FLAG}
+  )
+endif()
+
 if(EXECUTORCH_BUILD_CPUINFO)
   # --- cpuinfo
   set(ORIGINAL_CMAKE_POSITION_INDEPENDENT_CODE_FLAG

backends/arm/tosa_quant_utils.py

@@ -71,7 +71,7 @@ def quantize_value(x, qargs: QuantArgs, dtype=np.int8):
 
 
 def dequantize_value(qx, qargs: QuantArgs):
-    return (qx - qargs.zp) * qargs.scale
+    return (np.int64(qx) - qargs.zp) * qargs.scale
 
 
 def qargs_from_qnode(node: torch.fx.Node):

backends/vulkan/_passes/tag_memory_meta_pass.py

@@ -23,8 +23,6 @@
 
 from executorch.exir.pass_base import ExportPass, PassResult
 
-from torch._subclasses.fake_tensor import FakeTensor
-
 from torch.fx.passes.tools_common import NodeList
 from torch.fx.passes.utils.fuser_utils import topo_sort
 
@@ -138,9 +136,7 @@ def propose_node_storage(
                 return storage
 
         for arg in node.args:
-            if isinstance(arg, torch.fx.Node) and isinstance(
-                arg.meta["val"], FakeTensor
-            ):
+            if isinstance(arg, torch.fx.Node) and utils.is_tensor_node(arg):
                 storage = utils.get_node_storage_type(arg)
                 if storage is not None and storage in valid_storage_types:
                     return storage
@@ -178,9 +174,7 @@ def propose_node_layout(
                 return layout
 
         for arg in node.args:
-            if isinstance(arg, torch.fx.Node) and isinstance(
-                arg.meta["val"], FakeTensor
-            ):
+            if isinstance(arg, torch.fx.Node) and utils.is_tensor_node(arg):
                 layout = utils.get_node_memory_layout(arg)
                 if layout is not None and layout in valid_layouts:
                     return layout
@@ -202,14 +196,19 @@ def should_annotate(self, node) -> bool:
         if not isinstance(node, torch.fx.Node):
             return False
 
-        if not isinstance(node.meta["val"], FakeTensor):
+        if not utils.is_tensor_node(node):
             return False
 
         # Storage type and memory layout for tensorref will be determined at runtime
         # so there's no use in setting those attributes ahead of time.
         if node.meta.get("vkdg_tensorref", False):
             return False
 
+        # Skip annotating output node. The output tensors should be annotated by the
+        # time the output node is observed.
+        if node.op == "output":
+            return False
+
         return True
 
     def should_delay_annotation(self, node: torch.fx.Node) -> bool:

backends/vulkan/docs/android_demo.md

@@ -81,7 +81,8 @@ First, build and install ExecuTorch libraries, then build the LLaMA runner
 binary using the Android NDK toolchain.
 
 ```shell
-(rm -rf cmake-android-out && \
+./install_requirements.sh --clean
+(mkdir cmake-android-out && \
   cmake . -DCMAKE_INSTALL_PREFIX=cmake-android-out \
     -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
     -DANDROID_ABI=$ANDROID_ABI \