Use XNN fused RoPE for HF-style RoPE patterns

Author: GregoryComer

backends/xnnpack/_passes/__init__.py

@@ -19,6 +19,7 @@
     Conv1dUnsqueezePass,
 )
 from executorch.backends.xnnpack._passes.convert_to_linear import ConvertToLinearPass
+from executorch.backends.xnnpack._passes.convert_to_rope import ConvertToRopePass
 from executorch.backends.xnnpack._passes.convert_to_sdpa import ConvertToSDPAPass
 from executorch.backends.xnnpack._passes.convert_to_upsample_bilinear2d import (
     ConvertToUpsampleBilinear2d,
@@ -75,6 +76,7 @@ def __init__(
                 ConvertToLinearPass,
                 PropagateCustomMetaPass,
                 ConvertToSDPAPass,
+                ConvertToRopePass,
                 ConstPropPass,
                 FuseBatchNormPass,
                 DecomposeBatchNorm,

backends/xnnpack/_passes/convert_to_rope.py

@@ -0,0 +1,130 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from executorch.backends.xnnpack._passes.xnnpack_pass import XNNPACKPass
+from executorch.backends.xnnpack.partition.graphs import rope
+from executorch.exir.dialects._ops import ops as exir_ops
+
+from torch.fx.passes.infra.pass_base import PassResult
+from torch.fx.passes.utils.matcher_utils import InternalMatch, SubgraphMatcher
+
+logger = logging.getLogger(__name__)
+
+
+class ConvertToRopePass(XNNPACKPass):
+    def _build_weights(
+        self,
+        graph_module: torch.fx.GraphModule,
+        cos_node: torch.fx.Node,
+        sin_node: torch.fx.Node,
+        output_node: torch.fx.Node,
+    ) -> torch.fx.Node:
+        """
+        Construct the XNNPACK RoPE weights tensor from cos and sin inputs.
+
+        HF precompute_freqs_cis doubles the frequencies:
+            cos/sin shape: [batch, seq, head_dim] where head_dim = 2 * (dim // 2)
+            The first half and second half are identical.
+
+        XNNPACK expects weights: [tokens, channels] where:
+            weights[:, :C/2] = cos values (unique half)
+            weights[:, C/2:] = sin values (unique half)
+
+        We insert graph nodes to slice the unique halves and concatenate them.
+        """
+        cos_val = cos_node.meta.get("val")
+        head_dim = cos_val.shape[-1]
+        half_dim = head_dim // 2
+
+        with graph_module.graph.inserting_before(output_node):
+            cos_half = graph_module.graph.call_function(
+                exir_ops.edge.aten.slice_copy.Tensor,
+                args=(cos_node, -1, 0, half_dim),
+            )
+            cos_half.meta["val"] = cos_val.narrow(-1, 0, half_dim)
+
+            sin_val = sin_node.meta.get("val")
+            sin_half = graph_module.graph.call_function(
+                exir_ops.edge.aten.slice_copy.Tensor,
+                args=(sin_node, -1, 0, half_dim),
+            )
+            sin_half.meta["val"] = sin_val.narrow(-1, 0, half_dim)
+
+            weights = graph_module.graph.call_function(
+                exir_ops.edge.aten.cat.default,
+                args=([cos_half, sin_half], -1),
+            )
+            weights.meta["val"] = torch.cat(
+                [cos_half.meta["val"], sin_half.meta["val"]], -1
+            )
+
+        return weights
+
+    @staticmethod
+    def _trace_through_unsqueeze(node: torch.fx.Node) -> torch.fx.Node:
+        """If node is an unsqueeze_copy, return its input. Otherwise return node as-is."""
+        if (
+            node.op == "call_function"
+            and node.target == exir_ops.edge.aten.unsqueeze_copy.default
+        ):
+            return node.args[0]
+        return node
+
+    def create_rope(
+        self,
+        graph_module: torch.fx.GraphModule,
+        match: InternalMatch,
+    ):
+        logger.debug(f"Matched RoPE subgraph: {match}")
+
+        # placeholder_nodes are in the order of the pattern's placeholder ops:
+        # [x, cos_unsqueezed, sin_unsqueezed]
+        x_node = match.placeholder_nodes[0]
+        cos_unsqueezed = match.placeholder_nodes[1]
+        sin_unsqueezed = match.placeholder_nodes[2]
+        output_node = match.returning_nodes[0]
+
+        # Trace back through unsqueeze to get raw cos/sin for weight construction.
+        # The pattern excludes unsqueeze ops (they're shared between q/k RoPE),
+        # so the matched placeholders are the unsqueeze outputs.
+        cos_node = self._trace_through_unsqueeze(cos_unsqueezed)
+        sin_node = self._trace_through_unsqueeze(sin_unsqueezed)
+
+        weights = self._build_weights(graph_module, cos_node, sin_node, output_node)
+
+        with graph_module.graph.inserting_before(output_node):
+            rope_node = graph_module.graph.create_node(
+                "call_function",
+                torch.ops.xnnpack.rope.default,
+                args=(x_node, weights),
+            )
+
+        rope_node.meta["val"] = torch.empty_like(x_node.meta["val"])
+
+        output_node.replace_all_uses_with(rope_node)
+        graph_module.graph.eliminate_dead_code()
+
+    # override
+    def call(self, graph_module: torch.fx.GraphModule):
+        total_matches = 0
+        total_fused = 0
+        for pattern in rope.get_graphs():
+            sm = SubgraphMatcher(pattern.graph, ignore_literals=True)
+            matches = list(sm.match(graph_module.graph))
+            total_matches += len(matches)
+            for match in matches:
+                try:
+                    self.create_rope(graph_module, match)
+                    total_fused += 1
+                except Exception:
+                    logger.warning("Failed to fuse RoPE pattern", exc_info=True)
+        graph_module.recompile()
+        graph_module = super().call(graph_module).graph_module
+
+        return PassResult(graph_module, True)

backends/xnnpack/operators/__init__.py

@@ -43,6 +43,7 @@
     op_prelu,
     op_quant_dequant,
     op_relu,
+    op_rope,
     op_rsqrt,
     op_sigmoid,
     op_sin,

backends/xnnpack/operators/op_rope.py

@@ -0,0 +1,77 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Dict
+
+import torch
+from executorch.backends.xnnpack.operators.node_visitor import (
+    NodeVisitor,
+    register_node_visitor,
+)
+from executorch.backends.xnnpack.serialization.xnnpack_graph_schema import (
+    XNNGraph,
+    XNNRope,
+    XNode,
+)
+from executorch.backends.xnnpack.utils.utils import get_input_node
+
+# Register the custom op used by the fusion pass. The fused node targets
+# this op after ConvertToRopePass replaces the decomposed HF RoPE subgraph.
+lib = torch.library.Library("xnnpack", "FRAGMENT")
+lib.define("rope(Tensor input, Tensor weights) -> Tensor")
+
+
+@torch.library.impl(lib, "rope", "CompositeExplicitAutograd")
+def rope_impl(input: torch.Tensor, weights: torch.Tensor) -> torch.Tensor:
+    channels = input.shape[-1]
+    half_c = channels // 2
+    cos = weights[..., :half_c]
+    sin = weights[..., half_c:]
+
+    x_real = input[..., :half_c]
+    x_imag = input[..., half_c:]
+
+    out_real = x_real * cos - x_imag * sin
+    out_imag = x_real * sin + x_imag * cos
+    return torch.cat([out_real, out_imag], dim=-1)
+
+
+@torch.library.impl(lib, "rope", "Meta")
+def rope_meta(input: torch.Tensor, weights: torch.Tensor) -> torch.Tensor:
+    return torch.empty_like(input)
+
+
+@register_node_visitor
+class RopeVisitor(NodeVisitor):
+    target = "rope.default"
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        xnn_graph: XNNGraph,
+        vals_to_ids: Dict[torch.fx.Node, int],
+        debug_handle: int,
+    ) -> None:
+        self.define_nodes_tensor_inputs_outputs(node, xnn_graph, vals_to_ids)
+
+        input_id = vals_to_ids[get_input_node(node, 0)]
+        weights_id = vals_to_ids[get_input_node(node, 1)]
+        output_id = vals_to_ids[node]
+
+        ser_node = XNode(
+            xnode_union=XNNRope(
+                max_tokens=0,
+                input_id=input_id,
+                weights_id=weights_id,
+                output_id=output_id,
+                flags=0,
+            ),
+            debug_handle=debug_handle,
+        )
+        xnn_graph.xnodes.append(ser_node)

backends/xnnpack/operators/op_squeeze.py

@@ -13,6 +13,7 @@
 )
 from executorch.backends.xnnpack.serialization.xnnpack_graph_schema import (
     XNNGraph,
+    XNNStaticExpandDims,
     XNNStaticReshape,
     XNode,
 )
@@ -98,46 +99,21 @@ def define_node(
         vals_to_ids: Dict[torch.fx.Node, int],
         debug_handle: int,
     ) -> None:
-
-        dim = cast(int, node.args[1])
-        check_or_raise(
-            dim == -1 or dim == len(node.args[0].meta["val"].shape),
-            "XNNPACK currently only supports unsqueezing in last dimension",
-        )
-
         self.define_nodes_tensor_inputs_outputs(node, xnn_graph, vals_to_ids)
         input_node = get_input_node(node, 0)
 
-        # input
         input_id = vals_to_ids[input_node]
-
-        # output
         output_id = vals_to_ids[node]
 
-        check_or_raise(
-            "val" in input_node.meta,
-            "Missing val in tensor metadata for input when serializing XNNStaticReshape node",
-        )
-        dynamic_shape = node.meta["val"].shape
-        new_shape = []
-
-        num_dynamic_dims = 0
-        for dim in dynamic_shape:
-            if free_symbols(dim):
-                num_dynamic_dims += 1
-                new_shape.append(0)
-            else:
-                new_shape.append(dim)
-
-        check_or_raise(
-            num_dynamic_dims <= 1,
-            "XNNPACK reshape only supports 1 dynamic dimension. This may occur when ",
-        )
+        dim = cast(int, node.args[1])
+        input_ndim = len(input_node.meta["val"].shape)
+        if dim < 0:
+            dim = input_ndim + 1 + dim
 
         ser_node = XNode(
-            xnode_union=XNNStaticReshape(
-                num_dims=len(new_shape),
-                new_shape=new_shape,
+            xnode_union=XNNStaticExpandDims(
+                num_new_axes=1,
+                new_axes=[dim],
                 input_id=input_id,
                 output_id=output_id,
                 flags=0,

backends/xnnpack/partition/config/__init__.py

@@ -55,6 +55,7 @@
     SubConfig,
     TanhConfig,
     ToDimOrderCopyConfig,
+    UnsqueezeCopyConfig,
     UpsampleBilinear2dConfig,
 )
 from executorch.backends.xnnpack.partition.config.node_configs import (
@@ -116,6 +117,7 @@
     SoftmaxConfig,
     SquareRootConfig,
     SubConfig,
+    UnsqueezeCopyConfig,
     UpsampleBilinear2dConfig,
     # Quant/Dequant Op Configs
     QuantizedPerTensorConfig,