Update on "[executorch][flat_tensor] flat tensor header"

Parse a FlatTensor header. The FlatTensor header contains:
- flatbuffer offset
- flatbuffer size
- segment base offset
- segment data size (total size of all segments)

Mostly taken from extended_header.cpp/.h

Differential Revision: [D67064570](https://our.internmc.facebook.com/intern/diff/D67064570/)

[ghstack-poisoned]

Author: lucylq

.ci/scripts/utils.sh

@@ -17,17 +17,17 @@ retry () {
 }
 
 clean_executorch_install_folders() {
-  ./install_requirements.sh --clean
+  ./install_executorch.sh --clean
 }
 
 install_executorch() {
   which pip
   # Install executorch, this assumes that Executorch is checked out in the
   # current directory.
   if [[ "${1:-}" == "use-pt-pinned-commit" ]]; then
-    ./install_requirements.sh --pybind xnnpack --use-pt-pinned-commit
+    ./install_executorch.sh --pybind xnnpack --use-pt-pinned-commit
   else
-    ./install_requirements.sh --pybind xnnpack
+    ./install_executorch.sh --pybind xnnpack
   fi
   # Just print out the list of packages for debugging
   pip list

.github/workflows/apple.yml

@@ -9,7 +9,7 @@ on:
     paths:
       - .ci/scripts/setup-ios.sh
       - .github/workflows/apple.yml
-      - install_requirements.sh
+      - install_executorch.sh
       - backends/apple/**
       - build/build_apple_frameworks.sh
       - build/build_apple_llm_demo.sh

.github/workflows/pull.yml

@@ -200,7 +200,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
 
         # install pybind
-        bash install_requirements.sh --pybind xnnpack
+        bash install_executorch.sh --pybind xnnpack
 
         # install Llava requirements
         bash examples/models/llama/install_requirements.sh
@@ -436,7 +436,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
 
         # install pybind
-        bash install_requirements.sh --pybind xnnpack
+        bash install_executorch.sh --pybind xnnpack
 
         # install phi-3-mini requirements
         bash examples/models/phi-3-mini/install_requirements.sh
@@ -463,7 +463,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
 
         # install pybind
-        bash install_requirements.sh --pybind xnnpack
+        bash install_executorch.sh --pybind xnnpack
 
         # install llama requirements
         bash examples/models/llama/install_requirements.sh
@@ -490,7 +490,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
 
         # install pybind
-        bash install_requirements.sh --pybind xnnpack
+        bash install_executorch.sh --pybind xnnpack
 
         # install llama requirements
         bash examples/models/llama/install_requirements.sh
@@ -517,7 +517,7 @@ jobs:
         PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
 
         # install pybind
-        bash install_requirements.sh --pybind xnnpack
+        bash install_executorch.sh --pybind xnnpack
 
         # install llama requirements
         bash examples/models/llama/install_requirements.sh

backends/apple/coreml/test/test_coreml_partitioner.py

@@ -16,8 +16,7 @@
 
 
 class TestCoreMLPartitioner(unittest.TestCase):
-    # TODO(T182928844): Delegate dim order op to backend.
-    edge_compile_config = executorch.exir.EdgeCompileConfig(_skip_dim_order=True)
+    edge_compile_config = executorch.exir.EdgeCompileConfig()
 
     def test_add_sub_skip_mm(self):
         class Model(torch.nn.Module):

backends/apple/mps/setup.md

@@ -97,7 +97,7 @@ I 00:00:00.122615 executorch:mps_executor_runner.mm:501] Model verified successf
 ### [Optional] Run the generated model directly using pybind
 1. Make sure `pybind` MPS support was installed:
 ```bash
-./install_requirements.sh --pybind mps
+./install_executorch.sh --pybind mps
 ```
 2. Run the `mps_example` script to trace the model and run it directly from python:
 ```bash

backends/arm/_passes/annotate_channels_last_dim_order_pass.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 # All rights reserved.
 #
 # This source code is licensed under the BSD-style license found in the
@@ -15,7 +15,7 @@
     get_node_arg,
     insert_q_dq_pair,
 )
-from executorch.backends.arm.tosa_quant_utils import dq_op, q_op, register_passable_op
+from executorch.backends.arm.tosa_quant_utils import dq_op, q_op
 from executorch.backends.arm.tosa_utils import is_consumer_node_depthwise_conv2d
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass, PassResult
@@ -43,9 +43,6 @@ def _transpose_impl(*args, **kwargs):
     return args[0]
 
 
-register_passable_op(torch.ops.passthrough_to_tosa._transpose)
-
-
 class AnnotateChannelsLastDimOrder(ExportPass):
     """
     Annotates each node with a tosa_dim_order. tosa_dim_order can be seen as a channels-last dim-order

backends/arm/arm_backend.py

@@ -50,6 +50,7 @@ def __init__(self):
         self.output_format = None
         self.path_for_intermediates = None
         self.tosa_version = None
+        self.tosa_spec = None
         self.input_order = None
 
     def ethosu_compile_spec(

backends/arm/operators/__init__.py

@@ -13,7 +13,6 @@
     op_bmm,
     op_cat,
     op_conv2d,
-    op_dequant,
     op_exp,
     op_full,
     op_get_item,
@@ -24,7 +23,6 @@
     op_min,
     op_mul,
     op_permute,
-    op_quant,
     op_reciprocal,
     op_relu,
     op_repeat,

backends/arm/operators/op_dequant.py

@@ -1,4 +1,4 @@
-# Copyright 2023-2024 Arm Limited and/or its affiliates.
+# Copyright 2023-2025 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -19,7 +19,6 @@
 )
 from executorch.backends.arm.tosa_mapping import TosaArg
 
-from executorch.backends.arm.tosa_quant_utils import quantize_value
 from serializer.tosa_serializer import TosaOp
 
 
@@ -44,8 +43,8 @@ def define_node(
             input_qparams = get_input_qparams(node)  # pyre-ignore[16]
             qargs = input_qparams[0]
             # Convert to quantized representation
-            clamp_min_qs = quantize_value(inputs[1].number, qargs)
-            clamp_max_qs = quantize_value(inputs[2].number, qargs)
+            clamp_min_qs = qargs.quantize_value(inputs[1].number).item()
+            clamp_max_qs = qargs.quantize_value(inputs[2].number).item()
             # Set fp values to 0.0 since they are not used
             clamp_min_fp = 0.0
             clamp_max_fp = 0.0

backends/arm/operators/op_hardtanh.py

@@ -1,11 +1,10 @@
-# Copyright 2024 Arm Limited and/or its affiliates.
+# Copyright 2024-2025 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
 # pyre-unsafe
 
-import executorch.backends.arm.tosa_quant_utils as tqutils
 import serializer.tosa_serializer as ts
 import torch.fx
 
@@ -43,9 +42,8 @@ def define_node(
         clamp_max_qs = 0
         if inputs[0].dtype == ts.DType.INT8:
             out_qargs = get_output_qparams(node)  # pyre-ignore[16]
-            clamp_min_qs = tqutils.quantize_value(0, out_qargs[0])
-            clamp_max_qs = tqutils.quantize_value(float("inf"), out_qargs[0])
-
+            clamp_min_qs = out_qargs[0].quantize_value(0).item()
+            clamp_max_qs = out_qargs[0].quantize_value(float("inf")).item()
         else:
             clamp_min_fp = 0
             clamp_max_fp = float("inf")

backends/arm/operators/op_quant.py

@@ -12,12 +12,7 @@
 import torch
 import torch.fx
 from executorch.backends.arm.operators.node_visitor import NodeVisitor
-from executorch.backends.arm.tosa_mapping import map_dtype, TosaArg
-from executorch.backends.arm.tosa_quant_utils import (
-    dq_op,
-    get_quantized_node_output_dtype,
-    is_node_quantized,
-)
+from executorch.backends.arm.tosa_mapping import TosaArg
 from executorch.backends.arm.tosa_specification import TosaSpecification
 from executorch.backends.arm.tosa_utils import getNodeArgs, tosa_shape
 from torch.export.exported_program import ExportedProgram
@@ -35,15 +30,8 @@ def process_call_function(
     # Convert output (this node itself)
     output = TosaArg(node)
 
-    is_dq_node = node.target == dq_op
-    if is_dq_node:
-        output_dtype = ts.DType.INT8
-    else:
-        output_dtype = output.dtype
     tosa_graph.currRegion.currBasicBlock.addTensor(
-        output.name,
-        tosa_shape(output.shape, output.dim_order),
-        output_dtype,
+        output.name, tosa_shape(output.shape, output.dim_order), output.dtype
     )
 
     # Visiting each Node
@@ -79,11 +67,7 @@ def process_inputs(
     tensor = ts.TosaSerializerTensor(
         inputs[0].name,
         tosa_shape(input_shape, input_dim_order),
-        (
-            map_dtype(get_quantized_node_output_dtype(node))
-            if is_node_quantized(node)
-            else inputs[0].dtype
-        ),
+        inputs[0].dtype,
         data=None,
         placeholderFilename=inputs[0].name + ".npy",
     )

backends/arm/operators/op_relu.py

@@ -145,14 +145,3 @@ def get_u85_compile_spec_unbuilt(
         .dump_intermediate_artifacts_to(artifact_path)
     )
     return compile_spec
-
-
-def get_target_board(compile_spec: list[CompileSpec]) -> str | None:
-    for spec in compile_spec:
-        if spec.key == "compile_flags":
-            flags = spec.value.decode()
-            if "u55" in flags:
-                return "corstone-300"
-            elif "u85" in flags:
-                return "corstone-320"
-    return None

backends/arm/process_node.py

@@ -98,7 +98,7 @@ def test_mv2_u55_BI(self):
         )
         if conftest.is_option_enabled("corstone_fvp"):
             tester.run_method_and_compare_outputs(
-                atol=1.0, qtol=1, inputs=self.model_inputs, target_board="corstone-300"
+                atol=1.0, qtol=1, inputs=self.model_inputs
             )
 
     @pytest.mark.slow
@@ -118,5 +118,5 @@ def test_mv2_u85_BI(self):
         )
         if conftest.is_option_enabled("corstone_fvp"):
             tester.run_method_and_compare_outputs(
-                atol=1.0, qtol=1, inputs=self.model_inputs, target_board="corstone-320"
+                atol=1.0, qtol=1, inputs=self.model_inputs
             )

backends/arm/test/common.py

@@ -173,9 +173,7 @@ def test_maxpool2d_tosa_u55_BI(
             (test_data,),
         )
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1, inputs=(test_data,), target_board="corstone-300"
-            )
+            tester.run_method_and_compare_outputs(qtol=1, inputs=(test_data,))
 
     @parameterized.expand(test_data_suite)
     @pytest.mark.corstone_fvp
@@ -191,9 +189,7 @@ def test_maxpool2d_tosa_u85_BI(
             (test_data,),
         )
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1, inputs=(test_data,), target_board="corstone-320"
-            )
+            tester.run_method_and_compare_outputs(qtol=1, inputs=(test_data,))
 
     @parameterized.expand(test_data_suite_mult_batches)
     def test_maxpool2d_tosa_MI_mult_batches(
@@ -232,9 +228,7 @@ def test_maxpool2d_tosa_u55_BI_mult_batches(
             (test_data,),
         )
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1, inputs=(test_data,), target_board="corstone-300"
-            )
+            tester.run_method_and_compare_outputs(qtol=1, inputs=(test_data,))
 
     @parameterized.expand(test_data_suite_mult_batches)
     @pytest.mark.corstone_fvp
@@ -251,6 +245,4 @@ def test_maxpool2d_tosa_u85_BI_mult_batches(
             (test_data,),
         )
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1, inputs=(test_data,), target_board="corstone-320"
-            )
+            tester.run_method_and_compare_outputs(qtol=1, inputs=(test_data,))

backends/arm/test/models/test_mobilenet_v2_arm.py

@@ -121,9 +121,7 @@ def test_maximum_u55_BI(self, operand1: torch.Tensor, operand2: torch.Tensor):
             self.Maximum(), common.get_u55_compile_spec(), test_data
         )
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1, inputs=test_data, target_board="corstone-300"
-            )
+            tester.run_method_and_compare_outputs(qtol=1, inputs=test_data)
 
     @parameterized.expand(Maximum.test_parameters)
     def test_maximum_u85_BI(self, operand1: torch.Tensor, operand2: torch.Tensor):
@@ -132,6 +130,4 @@ def test_maximum_u85_BI(self, operand1: torch.Tensor, operand2: torch.Tensor):
             self.Maximum(), common.get_u85_compile_spec(), test_data
         )
         if conftest.is_option_enabled("corstone_fvp"):
-            tester.run_method_and_compare_outputs(
-                qtol=1, inputs=test_data, target_board="corstone-320"
-            )
+            tester.run_method_and_compare_outputs(qtol=1, inputs=test_data)