ai-edge-torch-nightly 0.3.0.dev20250114__py3-none-any.whl

ai_edge_torch/odml_torch/export.py

@@ -0,0 +1,403 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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.
+# ==============================================================================
+"""APIs to convert and lower a PyTorch ExportedProgram to MLIR."""
+
+import dataclasses
+import enum
+import io
+import operator
+from typing import Any, Callable, Optional
+
+from jax.lib import xla_extension
+from jax._src.lib.mlir import ir
+from jax._src.lib.mlir.dialects import func
+from jax._src.lib.mlir.dialects import hlo as stablehlo
+import torch
+import torch.utils._pytree as pytree
+
+from . import _torch_future
+from . import debuginfo
+from . import export_utils
+from . import lowerings
+
+LoweringContext = lowerings.context.LoweringContext
+
+
+def _build_flat_inputs(exported_program: torch.export.ExportedProgram):
+  """Build flattened inputs and metadata from exported program's signature."""
+  placeholder_nodes = [
+      n for n in exported_program.graph.nodes if n.op == "placeholder"
+  ]
+  export_flat_args = _torch_future.graph_module_flat_inputs(
+      exported_program, *exported_program.example_inputs
+  )
+
+  ir_inputs = []
+  tensor_metas = []
+  for node, arg in zip(placeholder_nodes, export_flat_args):
+    tensor_meta = node.meta.get("tensor_meta") or node.meta.get("val")
+    if tensor_meta is None:
+      raise RuntimeError(
+          f"{type(arg)} (for {node.name}) does not have tensor meta"
+      )
+
+    tensor_metas.append(tensor_meta)
+    # Assume all dynamic dimensions are unbounded.
+    # TODO: Add checks for ep.range_constraints in MLIR.
+    shape = tuple(
+        export_utils.IR_DYNAMIC if export_utils.is_torch_dynamic(s) else s
+        for s in tensor_meta.shape
+    )
+    ir_inputs.append(
+        ir.RankedTensorType.get(
+            shape,
+            export_utils.torch_dtype_to_ir_element_type(tensor_meta.dtype),
+        )
+    )
+  return tuple(ir_inputs), tuple(export_flat_args), tuple(tensor_metas)
+
+
+def _get_output_metas(exported_program: torch.export.ExportedProgram):
+  """Get the output node's tensor_meta from the exported program."""
+  outputs = [n for n in exported_program.graph.nodes if n.op == "output"]
+  assert len(outputs) == 1
+  outputs, _ = pytree.tree_flatten(outputs[0].args[0])
+  assert all(isinstance(output, torch.fx.Node) for output in outputs)
+  return tuple(output.meta["tensor_meta"] for output in outputs)
+
+
+class LoweringInterpreter(torch.fx.Interpreter):
+  """The FX interpreter to iterate and invoke corresponding lowering for each PyTorch op in the graph."""
+
+  def __init__(self, module: torch.fx.GraphModule, lctx: LoweringContext):
+    super().__init__(module)
+    self.lctx = lctx
+    self.outputs = None
+
+  def _build_loc(self, node: torch.fx.Node):
+
+    info = debuginfo.build_mlir_debuginfo(node)
+    if info is None:
+      return ir.Location.unknown()
+
+    return ir.Location.name(name=info)
+
+  def run_node(self, node: torch.fx.Node):
+    loc = self._build_loc(node)
+    with loc:
+      self.lctx = self.lctx.replace(ir_location=loc, node=node)
+      res = super().run_node(node)
+      self.lctx = self.lctx.replace(ir_location=None, node=None)
+    return res
+
+  def call_function(self, target, args, kwargs):
+    if target is operator.getitem:
+      return super().call_function(target, args, kwargs)
+
+    if hasattr(target, "_schema"):
+      new_args = []
+      for arg, spec in zip(args, target._schema.arguments):
+        if isinstance(spec.type, torch.TensorType):
+          if isinstance(arg, int):
+            arg = lowerings.utils.splat(arg, ir.IntegerType.get_signless(32))
+          elif isinstance(arg, float):
+            arg = lowerings.utils.splat(arg, ir.F32Type.get())
+
+        new_args.append(arg)
+      args = tuple(new_args)
+
+    lowering = lowerings.lookup(target)
+    if lowering is None:
+      raise RuntimeError(f"Lowering not found: {target}")
+    return lowering(self.lctx, *args, **kwargs)
+
+  def output(self, target, args, kwargs):
+    flat_outputs = pytree.tree_flatten(args[0])[0]
+    self.outputs = flat_outputs
+
+
+@dataclasses.dataclass
+class InputSpec:
+
+  class VariableType(enum.Enum):
+    USER_INPUT = "user_input"
+    PARAMETER = "parameter"
+
+  type_: VariableType
+  i: int = -1
+  name: str = ""
+
+  @classmethod
+  def parameter(cls, name: str):
+    return cls(type_=cls.VariableType.PARAMETER, name=name)
+
+  @classmethod
+  def user_input(cls, i: int):
+    return cls(type_=cls.VariableType.USER_INPUT, i=i)
+
+  @property
+  def is_parameter(self):
+    return self.type_ == self.VariableType.PARAMETER
+
+  @property
+  def is_user_input(self):
+    return self.type_ == self.VariableType.USER_INPUT
+
+
+@dataclasses.dataclass
+class VariableSignature:  # either argument or parameters
+  shape: list[int]
+  dtype: str
+  input_spec: InputSpec = None
+
+
+@dataclasses.dataclass
+class MlirLowered:
+  """The lowered MLIR module, metadata, and weight tensors bundle from exported program."""
+
+  ctx: ir.Context
+  module: ir.Module
+  state_dict: dict[str, torch.Tensor]
+  input_signature: list[VariableSignature]
+  output_signature: list[VariableSignature]
+
+  _tf_function: Optional[Callable[Any, Any]] = None
+
+  def __str__(self):
+    return str(self.get_text(enable_debug_info=False))
+
+  def __repr__(self):
+    return str(self.get_text(enable_debug_info=False))
+
+  def get_text(self, enable_debug_info=False):
+    return str(
+        self.module.operation.get_asm(enable_debug_info=enable_debug_info)
+    )
+
+  @property
+  def module_bytecode(self) -> bytes:
+    output = io.BytesIO()
+    self.module.operation.write_bytecode(file=output)
+    return output.getvalue()
+
+  @property
+  def module_bytecode_vhlo(self) -> bytes:
+    # HACK: In OSS, we use MLIR pybinding and StableHLO dialect from JAX's
+    # build, which may not have the same StableHLO version as what used in
+    # TFLite converter. Therefore we always serialize MLIR module in VHLO.
+    # TODO(b/362798610) Build MLIR pybinding in ai-edge-torch release.
+    if stablehlo.get_api_version() < 9:
+      target_version = stablehlo.get_minimum_version()
+    else:
+      target_version = stablehlo.get_version_from_compatibility_requirement(
+          stablehlo.StablehloCompatibilityRequirement.WEEK_12
+      )
+    module_bytecode = xla_extension.mlir.serialize_portable_artifact(
+        self.module_bytecode, target_version
+    )
+    return module_bytecode
+
+  @property
+  def tf_function(self):
+    # Lazy import
+    from . import tf_integration
+
+    if self._tf_function is None:
+      self._tf_function = tf_integration.mlir_to_tf_function(self)
+    return self._tf_function
+
+  def __call__(self, *args):
+    # Lazy importing TF when execution is needed.
+    return self.tf_function(*args)
+
+
+# TODO(b/331481564) Make this a ai_edge_torch FX pass.
+def _convert_i64_to_i32(exported_program: torch.export.ExportedProgram):
+  """Convert internal constant aten ops' output from int64 to int32.
+
+  Int32 generally has better performance and compatibility than int64 in
+  runtime. This pass converts aten op where the output(s) are int64 constant
+  tensors to return int32 constant tensors.
+
+  Args:
+    exported_program: The exported program to apply the pass.
+  """
+
+  def in_i32(x: int):
+    return -2147483648 <= x <= 2147483647
+
+  def to_int32(x: torch.Tensor):
+    return torch.ops.aten._to_copy.default(x, dtype=torch.int32)
+
+  def rewrite_arange(node: torch.fx.Node):
+    tensor_meta = node.meta.get("tensor_meta", None)
+    if not tensor_meta:
+      return
+
+    start, end = node.args[:2]
+    if tensor_meta.dtype != torch.int64:
+      return
+    if not (in_i32(start) and in_i32(end)):
+      return
+    op = node.target
+    node.target = lambda *args, **kwargs: to_int32(op(*args, **kwargs))
+
+  graph_module = exported_program.graph_module
+  for node in graph_module.graph.nodes:
+
+    if node.target == torch.ops.aten.arange.start_step:
+      rewrite_arange(node)
+
+
+# TODO(b/331481564) Make this a ai_edge_torch FX pass.
+def _convert_q_dq_per_channel_args_to_list(
+    exported_program: torch.export.ExportedProgram,
+):
+  """Resolve tensor inputs to Q/DQ ops as static number list for lowering.
+
+  This pass makes the ExportedProgram in a non-executable state. This pass must
+  be run after all run_decompositions calls.
+  """
+  placeholder_nodes = [
+      n for n in exported_program.graph.nodes if n.op == "placeholder"
+  ]
+  export_flat_args = _torch_future.graph_module_flat_inputs(
+      exported_program, *exported_program.example_inputs
+  )
+
+  placeholder_tensor = {
+      n: tensor for n, tensor in zip(placeholder_nodes, export_flat_args)
+  }
+
+  graph_module = exported_program.graph_module
+  for node in graph_module.graph.nodes:
+    if node.target in (
+        torch.ops.quantized_decomposed.quantize_per_channel.default,
+        torch.ops.quantized_decomposed.quantize_per_tensor.tensor,
+        torch.ops.quantized_decomposed.dequantize_per_channel.default,
+        torch.ops.quantized_decomposed.dequantize_per_tensor.tensor,
+    ):
+      input, scale_node, zero_point_node = node.args[:3]
+      scale = placeholder_tensor[scale_node]
+      zero_point = placeholder_tensor[zero_point_node]
+
+      scale = scale.detach().numpy().tolist()
+      zero_point = zero_point.detach().numpy().tolist()
+      node.args = (input, scale, zero_point, *node.args[3:])
+
+
+def exported_program_to_mlir(
+    exported_program: torch.export.ExportedProgram,
+) -> MlirLowered:
+  """Lower the exported program to MLIR."""
+  exported_program = _torch_future.safe_run_decompositions(
+      exported_program, lowerings.decompositions()
+  )
+
+  _convert_i64_to_i32(exported_program)
+
+  # No decompositions but just retracing/cananicalization.
+  exported_program = _torch_future.safe_run_decompositions(
+      exported_program, _torch_future.dummy_decomp_table()
+  )
+
+  # Passes below mutate the exported program to a state not executable by torch.
+  # Do not call run_decompositions after applying the passes.
+  _convert_q_dq_per_channel_args_to_list(exported_program)
+
+  with export_utils.create_ir_context() as context, ir.Location.unknown():
+
+    module = ir.Module.create()
+    lctx = LoweringContext(context, module)
+    interpreter = LoweringInterpreter(exported_program.graph_module, lctx)
+    ir_flat_inputs, export_flat_args, tensor_metas = _build_flat_inputs(
+        exported_program
+    )
+
+    # HACK: OSS MLIR pybinding could mysteriously transform func.func under
+    # construction into a func.return op after calling ir.Module.parse(..)
+    # in the context, which happens in JAX bridge. This is a bug in MLIR
+    # pybinding.
+    # Workaround steps:
+    # 1. Create a temp func.func.
+    # 2. Create and insert ops to temp's entry block. During the process
+    #   the temp func.func would be broken, but the ops in the block are fine.
+    # 3. Create the main func.func and copy all the ops in temp's entry block
+    #   to main.
+    # 4. Erase the temp func.func.
+    temp_func = func.FuncOp(
+        "temp",
+        ir.FunctionType.get(ir_flat_inputs, []),
+        ip=ir.InsertionPoint.at_block_begin(module.body),
+    )
+    with ir.InsertionPoint(temp_func.add_entry_block()):
+      interpreter.run(*temp_func.arguments, enable_io_processing=False)
+      num_mutations = len(exported_program.graph_signature.buffers_to_mutate)
+      outputs = interpreter.outputs[num_mutations:]
+      func.ReturnOp(interpreter.outputs[num_mutations:])
+
+    main_func = func.FuncOp(
+        "main",
+        ir.FunctionType.get(ir_flat_inputs, [o.type for o in outputs]),
+        ip=ir.InsertionPoint.at_block_begin(module.body),
+    )
+    with ir.InsertionPoint(main_func.add_entry_block()):
+      outputs = export_utils.clone_func_body_ops(temp_func, main_func.arguments)
+      func.ReturnOp(outputs)
+
+    main_func.attributes["sym_visibility"] = ir.StringAttr.get("public")
+    temp_func.erase()
+
+    module.operation.verify()
+
+  input_signature = []
+  state_dict = {}
+
+  user_inputs_cnt = 0
+  for arg, tensor_meta, input_spec in zip(
+      export_flat_args,
+      tensor_metas,
+      exported_program.graph_signature.input_specs,
+  ):
+    # Assumption:
+    # All states comes first in the list of args, and user provided inputs
+    # comes later. Also there is no kwargs.
+    if input_spec.kind == torch.export.graph_signature.InputKind.USER_INPUT:
+      input_signature.append(
+          VariableSignature(
+              tensor_meta.shape,
+              tensor_meta.dtype,
+              input_spec=InputSpec.user_input(user_inputs_cnt),
+          )
+      )
+      user_inputs_cnt += 1
+    else:
+      # Parameter or constant
+      state_dict[input_spec.target] = arg
+      input_signature.append(
+          VariableSignature(
+              tensor_meta.shape,
+              tensor_meta.dtype,
+              input_spec=InputSpec.parameter(input_spec.target),
+          )
+      )
+
+  output_signature = [
+      VariableSignature(tensor_meta.shape, tensor_meta.dtype)
+      for tensor_meta in _get_output_metas(exported_program)
+  ]
+  return MlirLowered(
+      context, module, state_dict, input_signature, output_signature
+  )

ai_edge_torch/odml_torch/export_utils.py

@@ -0,0 +1,157 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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.
+# ==============================================================================
+"""Utilities for ODML Torch export."""
+
+import re
+from typing import Sequence, cast
+from ai_edge_torch.odml_torch.lowerings import utils as lowering_utils
+import jax._src.interpreters.mlir
+from jax._src.lib.mlir import ir
+from jax._src.lib.mlir.dialects import func
+import torch
+
+# std::numeric_limits<int64_t>::min()
+IR_DYNAMIC = -9223372036854775808
+
+
+def is_ir_dynamic(v):
+  return v == IR_DYNAMIC
+
+
+def is_torch_dynamic(v):
+  return isinstance(v, torch.SymInt)
+
+
+def is_iterable(v):
+  try:
+    iter(v)
+  except TypeError:
+    return False
+  return True
+
+
+def create_ir_context():
+  # HACK: Use ir context from JAX as base for better stability in OSS.
+  # TODO(b/362798610) Build MLIR pybinding in ai-edge-torch release.
+  context = jax._src.interpreters.mlir.make_ir_context()
+  context.allow_unregistered_dialects = True
+  return context
+
+
+def inline(
+    symbol_table: ir.SymbolTable,
+    block: ir.Block,
+):
+  """Recursively inlines all func.call ops in the block.
+
+  The symbol_table must include all func.func called by func.call ops.
+  This inliner in Python is implemented because MLIR inline pass from JAX's
+  MLIR pybinding build in OSS cannot properly inline func.call ops.
+  """
+  while True:
+    is_changed = False
+    for op in block.operations:
+      if op.OPERATION_NAME != func.CallOp.OPERATION_NAME:
+        continue
+
+      call_op = cast(func.CallOp, op)
+      func_op = cast(func.FuncOp, symbol_table[call_op.callee.value])
+      with ir.InsertionPoint(op):
+        new_results = clone_func_body_ops(func_op, call_op.operands)
+
+      for old_result, new_result in zip(call_op.results, new_results):
+        old_result = cast(ir.Value, old_result)
+        old_result.replace_all_uses_with(new_result)
+      call_op.erase()
+      is_changed = True
+
+    if not is_changed:
+      break
+
+  for op in block.operations:
+    for region in op.regions:
+      for block in region.blocks:
+        inline(symbol_table, block)
+
+
+def clone_func_body_ops(func_op: func.FuncOp, ir_inputs: Sequence[ir.Value]):
+  """Clone operations in the func_op's body by one into the current context."""
+  func_args = list(func_op.arguments)
+  ir_inputs = list(ir_inputs)
+  assert len(func_args) == len(ir_inputs)
+
+  value_mapping = {arg: ir_input for arg, ir_input in zip(func_args, ir_inputs)}
+
+  for op in list(func_op.entry_block.operations):
+    cloned_operands = [value_mapping[val] for val in op.operands]
+    if op.OPERATION_NAME == func.ReturnOp.OPERATION_NAME:
+      return cloned_operands
+
+    cloned = cast(ir.Operation, op.operation.clone())
+
+    for i in range(len(op.operands)):
+      cloned.operands[i] = cloned_operands[i]
+
+    for i in range(len(op.results)):
+      value_mapping[op.results[i]] = cloned.results[i]
+
+  return []
+
+
+def sanitize_aten_op_name(op, chars=":."):
+  return re.sub("[{}]".format(chars), "_", str(op))
+
+
+def build_ir_attr(val):
+  if val is None:
+    return ir.StringAttr.get("py_None")
+  if isinstance(val, bool):
+    return ir.BoolAttr.get(val)
+  if isinstance(val, int):
+    return ir.IntegerAttr.get(ir.IntegerType.get_signless(64), val)
+  if isinstance(val, float):
+    return ir.BoolAttr.get(val)
+  if isinstance(val, str):
+    return ir.StringAttr.get(val)
+  if isinstance(val, dict):
+    return ir.DictAttr.get({k: build_ir_attr(v) for k, v in val.items()})
+  if isinstance(val, (list, tuple)):
+    return ir.ArrayAttr.get([build_ir_attr(v) for v in val])
+
+  # Stringify the value to a StringAttr by default
+  return ir.StringAttr.get(str(val))
+
+
+torch_dtype_to_ir_element_type = lowering_utils.torch_dtype_to_ir_element_type
+
+
+def ir_element_type_to_torch_dtype(ty):
+  if isinstance(ty, ir.F32Type):
+    return torch.float32
+  if isinstance(ty, ir.F64Type):
+    return torch.float64
+  if isinstance(ty, ir.F16Type):
+    return torch.half
+  if isinstance(ty, ir.IntegerType):
+    if ty.is_signless:
+      if ty.width == 64:
+        return torch.long
+      if ty.width == 32:
+        return torch.int32
+      if ty.width == 16:
+        return torch.int16
+      if ty.width == 1:
+        return torch.bool
+  raise RuntimeError(f"Unsupported ir element type: {ty}")

ai_edge_torch/odml_torch/jax_bridge/__init__.py

@@ -0,0 +1,18 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# 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.
+# ==============================================================================
+from ai_edge_torch.odml_torch.jax_bridge import _wrap
+from ai_edge_torch.odml_torch.jax_bridge import utils
+
+wrap = _wrap.wrap