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

ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/pass_body.py

@@ -0,0 +1,303 @@
+# 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.
+# ==============================================================================
+"""Optimize layout transposes pass."""
+
+import operator
+import os
+from typing import Union
+
+from ai_edge_torch import fx_pass_base
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_check  # NOQA
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_mark  # NOQA
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_partitioners  # NOQA
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import layout_rewrite  # NOQA
+from ai_edge_torch._convert.fx_passes.optimize_layout_transposes_pass import utils  # NOQA
+import torch
+import torch.ao.quantization.quantize_pt2e
+
+TransposeFunc = Union[utils.tensor_to_nchw, utils.tensor_to_nhwc]
+
+
+class OptimizeLayoutTransposesPass(fx_pass_base.ExportedProgramPassBase):
+
+  def get_source_meta(self, node: torch.fx.Node):
+    keys = ["stack_trace", "nn_module_stack", "source_fn_stack", "from_node"]
+    meta = {}
+    for key in keys:
+      if key in node.meta:
+        meta[key] = node.meta[key]
+    return meta
+
+  def insert_t_q_dq(
+      self,
+      graph: torch.fx.Graph,
+      input_dq: torch.fx.Node,
+      target: torch.fx.Node,
+      transpose_func: TransposeFunc,
+      transpose_node_meta: dict,
+  ) -> list[torch.fx.Node]:
+    """original:
+
+        input_dq -> target
+    insert the node as:
+        input_dq -> (T q dq) -> target
+    """
+    assert utils.is_dq_node(input_dq)
+
+    q_args = input_dq.args[1:]
+    q_kwargs = input_dq.kwargs
+    q_op, dq_op = utils.get_paired_q_dq_ops(input_dq.target)
+    with graph.inserting_before(target):
+      t = graph.call_function(transpose_func, (input_dq,))
+      # Q and DQ inserted here may required updating the `axis` arg when they
+      # are per_channel ops. However, instead of updating here, the nodes would
+      # be marked as NHWC/NCHW and applied rewriters after partitioning.
+      q = graph.call_function(q_op, (t,) + q_args, q_kwargs)
+      dq = graph.call_function(dq_op, (q,) + q_args, q_kwargs)
+
+    input_dq.meta = transpose_node_meta
+    t.meta = transpose_node_meta
+    q.meta = transpose_node_meta
+    dq.meta = self.get_source_meta(target)
+
+    target.replace_input_with(input_dq, dq)
+    return [t, q, dq]
+
+  def insert_dq_t_q(
+      self,
+      graph: torch.fx.Graph,
+      input_q: torch.fx.Node,
+      target: torch.fx.Node,
+      transpose_func: TransposeFunc,
+      transpose_node_meta: dict,
+  ) -> list[torch.fx.Node]:
+    """original:
+
+        input_q -> target
+    insert the node as:
+        input_q -> (dq T q) -> target
+    """
+    assert utils.is_q_node(input_q)
+
+    q_args = input_q.args[1:]
+    q_kwargs = input_q.kwargs
+    q_op, dq_op = utils.get_paired_q_dq_ops(input_q.target)
+    with graph.inserting_before(target):
+      # Q and DQ inserted here may required updating the `axis` arg when they
+      # are per_channel ops. However, instead of updating here, the nodes would
+      # be marked as NHWC/NCHW and applied rewriters after partitioning.
+      dq = graph.call_function(dq_op, (input_q,) + q_args, q_kwargs)
+      t = graph.call_function(transpose_func, (dq,))
+      q = graph.call_function(q_op, (t,) + q_args, q_kwargs)
+
+    dq.meta = transpose_node_meta
+    t.meta = transpose_node_meta
+    q.meta = transpose_node_meta
+
+    target.replace_input_with(input_q, q)
+    return [dq, t, q]
+
+  def insert_layout_transpose(
+      self,
+      graph: torch.fx.Graph,
+      input_node: torch.fx.Node,
+      target_node: torch.fx.Node,
+      transpose_func: TransposeFunc,
+      transpose_node_meta: dict,
+  ) -> None:
+    assert transpose_func in (utils.tensor_to_nchw, utils.tensor_to_nhwc)
+
+    # new_nodes only contains Q/DQ/Transpose nodes, which are all SISO.
+    # Insertion order input nodes -> output nodes
+    new_nodes = []
+
+    # Constraint Q2: the NHWC partition's entry and exit must not be output
+    # edges of Q/DQ ops that are connected to a constant/weight tensor.
+    while layout_mark.is_const_node(input_node) and (
+        utils.is_dq_node(input_node) or utils.is_q_node(input_node)
+    ):
+      with graph.inserting_before(target_node):
+        new_input_node = graph.node_copy(input_node)
+
+      target_node.replace_input_with(input_node, new_input_node)
+
+      new_nodes = [new_input_node] + new_nodes
+      input_node, target_node = new_input_node.args[0], new_input_node
+
+    if utils.is_q_node(input_node):
+      # Constraint Q3: when the entry and exit is right after a q op (occur after a (dq-op-q)
+      # triplet), the transpose must be added as a quantized transpose in (dq-T-q)
+      # input_q -> (dq T q) -> target
+      new_nodes = (
+          self.insert_dq_t_q(
+              graph,
+              input_node,
+              target_node,
+              transpose_func,
+              transpose_node_meta,
+          )
+          + new_nodes
+      )
+    elif utils.is_dq_node(input_node):
+      # Constraint Q1: the NHWC partition's entry and exit cannot be edges
+      # within (dq-op-q) triplet.
+      # input_dq -> (T q dq) -> target
+      new_nodes = (
+          self.insert_t_q_dq(
+              graph,
+              input_node,
+              target_node,
+              transpose_func,
+              transpose_node_meta,
+          )
+          + new_nodes
+      )
+    else:
+      # input -> target
+      with graph.inserting_before(target_node):
+        t = graph.call_function(transpose_func, (input_node,))
+      t.meta = transpose_node_meta
+      target_node.replace_input_with(input_node, t)
+      new_nodes = [t] + new_nodes
+
+    # Mark new nodes as NCHW or NHWC
+    # For all nodes before the transpose, mark it as input_marker
+    # For all nodes after the transpose (incl. transpose), mark it as output_marker
+    if transpose_func == utils.tensor_to_nchw:
+      input_marker, target_marker = (
+          layout_mark.mark_as_nhwc_node,
+          layout_mark.mark_as_nchw_node,
+      )
+    else:
+      input_marker, target_marker = (
+          layout_mark.mark_as_nchw_node,
+          layout_mark.mark_as_nhwc_node,
+      )
+
+    marker = input_marker
+    for node in new_nodes:
+      if node.target == transpose_func:
+        marker = target_marker
+      marker(node)
+    assert marker == target_marker
+
+  def input_to_nhwc(
+      self,
+      graph: torch.fx.Graph,
+      input_node: torch.fx.Node,
+      target_node: torch.fx.Node,
+  ) -> None:
+    if layout_mark.is_nhwc_node(input_node):
+      return
+
+    if not layout_check.is_4d(input_node):
+      raise AssertionError(
+          "Attempting to convert non-NHWC compatible node to NHWC:"
+          f" {input_node}"
+      )
+
+    # Assign target node's source meta to the to_NHWC node, because the transpose
+    # is added for the existence of target node.
+    self.insert_layout_transpose(
+        graph,
+        input_node,
+        target_node,
+        utils.tensor_to_nhwc,
+        self.get_source_meta(target_node),
+    )
+
+  def input_to_nchw(
+      self,
+      graph: torch.fx.Graph,
+      input_node: torch.fx.Node,
+      target_node: torch.fx.Node,
+  ) -> None:
+    if layout_mark.is_nchw_node(input_node):
+      return
+
+    self.insert_layout_transpose(
+        graph,
+        input_node,
+        target_node,
+        utils.tensor_to_nchw,
+        self.get_source_meta(input_node),
+    )
+
+  def mark_const_nodes(self, exported_program: torch.export.ExportedProgram):
+    graph_module = exported_program.graph_module
+    graph = graph_module.graph
+
+    input_specs = exported_program.graph_signature.input_specs
+    non_user_input_names = set()
+    for spec in input_specs:
+      if spec.kind != torch.export.graph_signature.InputKind.USER_INPUT:
+        non_user_input_names.add(spec.arg.name)
+
+    for node in graph.nodes:
+      has_input_nodes = len(node.all_input_nodes) > 0
+      all_inputs_are_const = all(
+          map(layout_mark.is_const_node, node.all_input_nodes)
+      )
+      if (
+          node.name in non_user_input_names
+          or (has_input_nodes and all_inputs_are_const)
+          or (node.op != "placeholder" and not has_input_nodes)
+      ):
+        layout_mark.mark_as_const_node(node)
+
+  def call(self, exported_program: torch.export.ExportedProgram):
+    self.mark_const_nodes(exported_program)
+
+    graph_module = exported_program.graph_module
+    partitioner = os.environ.get(
+        "AIEDGETORCH_LAYOUT_OPTIMIZE_PARTITIONER", None
+    )
+    if partitioner == "MINCUT":
+      graph_module = layout_partitioners.min_cut.partition(graph_module)
+    elif partitioner == "GREEDY":
+      graph_module = layout_partitioners.greedy.partition(graph_module)
+    else:
+      # By default use min cut partitioner if possible
+      if layout_partitioners.min_cut.can_partition(graph_module):
+        graph_module = layout_partitioners.min_cut.partition(graph_module)
+      else:
+        graph_module = layout_partitioners.greedy.partition(graph_module)
+
+    graph = graph_module.graph
+    for node in list(graph.nodes):
+      if node.target == operator.getitem:
+        # force the layout mark of a getitem node to follow its producer.
+        if layout_mark.is_nchw_node(node.args[0]):
+          layout_mark.mark_as_nchw_node(node)
+        else:
+          layout_mark.mark_as_nhwc_node(node)
+
+    for node in list(graph.nodes):
+      if layout_mark.is_nhwc_node(node):
+        for input_node in layout_check.get_layout_sensitive_inputs(node):
+          self.input_to_nhwc(graph, input_node, node)
+        layout_rewrite.rewrite_nhwc_node(node)
+      else:
+        for input_node in layout_check.get_layout_sensitive_inputs(node):
+          # Note: for non-4D tensors input_to_nchw is always noop.
+          self.input_to_nchw(graph, input_node, node)
+
+    graph_module.graph.eliminate_dead_code()
+    graph_module.recompile()
+    graph_module.graph.lint()
+    # Mark const node again for debugging
+    self.mark_const_nodes(exported_program)
+
+    return fx_pass_base.ExportedProgramPassResult(exported_program, True)

ai_edge_torch/_convert/fx_passes/optimize_layout_transposes_pass/utils.py

@@ -0,0 +1,64 @@
+# 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.
+# ==============================================================================
+"""Utils for the optimized layout transposes pass."""
+
+from typing import Callable
+
+import torch
+import torch.ao.quantization.quantize_pt2e
+
+
+def tensor_to_nhwc(t: torch.Tensor):
+  return torch.ops.aten.permute(t.contiguous(), [0, 2, 3, 1]).contiguous()
+
+
+def tensor_to_nchw(t: torch.Tensor):
+  return torch.ops.aten.permute(t.contiguous(), [0, 3, 1, 2]).contiguous()
+
+
+def flatten_torch_op_overloads(op):
+  if isinstance(op, torch._ops.OpOverloadPacket):
+    return [getattr(op, overload) for overload in op.overloads()]
+  return [op]
+
+
+_TORCH_Q_OPS = [
+    torch.ops.quantized_decomposed.quantize_per_tensor.default,
+    torch.ops.quantized_decomposed.quantize_per_tensor.tensor,
+    torch.ops.quantized_decomposed.quantize_per_tensor.tensor2,
+    torch.ops.quantized_decomposed.quantize_per_channel.default,
+]
+
+_TORCH_DQ_OPS = [
+    torch.ops.quantized_decomposed.dequantize_per_tensor.default,
+    torch.ops.quantized_decomposed.dequantize_per_tensor.tensor,
+    torch.ops.quantized_decomposed.dequantize_per_tensor.tensor2,
+    torch.ops.quantized_decomposed.dequantize_per_channel.default,
+]
+
+
+def is_q_node(node: torch.fx.Node):
+  return node.target in _TORCH_Q_OPS
+
+
+def is_dq_node(node: torch.fx.Node):
+  return node.target in _TORCH_DQ_OPS
+
+
+def get_paired_q_dq_ops(op: Callable) -> tuple[Callable, Callable]:
+  for q, dq in zip(_TORCH_Q_OPS, _TORCH_DQ_OPS):
+    if op in (q, dq):
+      return q, dq
+  raise AssertionError(f"{op} is not a Q/DQ op.")

ai_edge_torch/_convert/fx_passes/remove_non_user_outputs_pass.py

@@ -0,0 +1,52 @@
+# 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.
+# ==============================================================================
+"""Pass to remove all non user outputs from exported program."""
+
+
+from ai_edge_torch import fx_pass_base
+import torch
+
+
+class RemoveNonUserOutputsPass(fx_pass_base.ExportedProgramPassBase):
+  """This pass removes all non user outputs from the exported program's output.
+
+  The FX graph may output more tensors/data than what user's original model
+  returns. Those additional outputs include user input mutations, gradient to
+  parameter, etc. Those outputs are not supported by our inference only
+  conversion or runtime. This pass remove all those outputs to ensure the
+  converted models' outputs match what returned from user's model in eval mode.
+  """
+
+  def call(self, exported_program: torch.export.ExportedProgram):
+    for node in exported_program.graph.nodes:
+      if node.op != "output":
+        continue
+
+      outputs = node.args[0]
+      output_specs = exported_program.graph_signature.output_specs
+
+      new_outputs = []
+      new_output_specs = []
+      for output, spec in zip(outputs, output_specs):
+        if spec.kind == torch.export.graph_signature.OutputKind.USER_OUTPUT:
+          new_outputs.append(output)
+          new_output_specs.append(spec)
+
+      node.args = (tuple(new_outputs),)
+      exported_program.graph_signature.output_specs = new_output_specs
+
+    exported_program.graph_module.graph.lint()
+    exported_program.graph_module.recompile()
+    return fx_pass_base.ExportedProgramPassResult(exported_program, True)

ai_edge_torch/_convert/signature.py

@@ -0,0 +1,66 @@
+# 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.
+# ==============================================================================
+
+import dataclasses
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from ai_edge_torch import lowertools
+import torch
+import torch.utils._pytree as pytree
+
+
+@dataclasses.dataclass
+class Signature:
+  name: str
+  module: torch.nn.Module
+  sample_args: tuple[torch.Tensor]
+  sample_kwargs: dict[str, torch.Tensor]
+  dynamic_shapes: Optional[Union[Dict[str, Any], Tuple[Any]]] = None
+
+  @property
+  def _normalized_sample_args_kwargs(self):
+    args, kwargs = self.sample_args, self.sample_kwargs
+    if args is not None:
+      if not isinstance(args, tuple):
+        # TODO(b/352584188): Check value types
+        raise ValueError("sample_args must be a tuple of torch tensors.")
+    if kwargs is not None:
+      if not isinstance(kwargs, dict) or not all(
+          isinstance(key, str) for key in kwargs.keys()
+      ):
+        # TODO(b/352584188): Check value types
+        raise ValueError("sample_kwargs must be a dict of string to tensor.")
+    args = args if args is not None else tuple()
+    kwargs = kwargs if kwargs is not None else {}
+    return args, kwargs
+
+  @property
+  def flat_arg_names(self) -> list[str]:
+    spec = pytree.tree_flatten(self._normalized_sample_args_kwargs)[1]
+    args_spec, kwargs_spec = spec.children_specs
+    names = []
+    for i in range(args_spec.num_leaves):
+      names.append(f"args_{i}")
+
+    kwargs_names = lowertools.flat_dict_names(
+        kwargs_spec.children_specs, kwargs_spec.context
+    )
+    names.extend(kwargs_names)
+    return names
+
+  @property
+  def flat_args(self) -> tuple[Any]:
+    args, kwargs = self._normalized_sample_args_kwargs
+    return tuple([*args, *kwargs.values()])

ai_edge_torch/_convert/test/__init__.py

@@ -0,0 +1,14 @@
+# 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.
+# ==============================================================================