Add aten._trilinear support to torch_lib core

onnxscript/function_libs/torch_lib/ops/core.py

@@ -12,6 +12,7 @@
 from __future__ import annotations
 
 import math
+import string
 from typing import Any, Optional, Sequence, Tuple, Union
 
 import numpy as np
@@ -56,6 +57,7 @@
 _INT64_MAX = 9223372036854775807
 _INT64_MIN = -9223372036854775808
 _MATH_PI = math.pi
+_EINSUM_SYMBOLS = string.ascii_letters
 
 
 @torch_op("aten::_local_scalar_dense", trace_only=True)
@@ -9791,6 +9793,81 @@ def aten_tril_indices(row: int, col: int, offset: int = 0) -> TensorType:
     raise NotImplementedError()
 
 
+def _get_einsum_symbol(dim: int) -> str:
+    if dim >= len(_EINSUM_SYMBOLS):
+        raise ValueError("aten::_trilinear only supports up to 52 dimensions")
+    return _EINSUM_SYMBOLS[dim]
+
+
+def _validate_trilinear_dims(
+    total_dim: int, dims: Sequence[int], dims_name: str
+) -> None:
+    seen_dims = set()
+    for dim in dims:
+        if dim < 0 or dim >= total_dim:
+            raise ValueError(
+                f"aten::_trilinear {dims_name} values must be in [0, {total_dim})"
+            )
+        if dim in seen_dims:
+            raise ValueError(
+                f"aten::_trilinear {dims_name} values must be unique"
+            )
+        seen_dims.add(dim)
+
+
+def _build_trilinear_subscript(
+    total_dim: int, expanded_dims: Sequence[int], dims_name: str
+) -> str:
+    _validate_trilinear_dims(total_dim, expanded_dims, dims_name)
+    expanded_dims_set = set(expanded_dims)
+    return "".join(
+        _get_einsum_symbol(dim) for dim in range(total_dim) if dim not in expanded_dims_set
+    )
+
+
+def _build_trilinear_equation(
+    total_dim: int,
+    expand1: Sequence[int],
+    expand2: Sequence[int],
+    expand3: Sequence[int],
+    sumdim: Sequence[int],
+) -> str:
+    _validate_trilinear_dims(total_dim, sumdim, "sumdim")
+    sumdim_set = set(sumdim)
+    output_subscript = "".join(
+        _get_einsum_symbol(dim) for dim in range(total_dim) if dim not in sumdim_set
+    )
+    return (
+        f"{_build_trilinear_subscript(total_dim, expand1, 'expand1')},"
+        f"{_build_trilinear_subscript(total_dim, expand2, 'expand2')},"
+        f"{_build_trilinear_subscript(total_dim, expand3, 'expand3')}"
+        f"->{output_subscript}"
+    )
+
+
+@torch_op("aten::_trilinear", trace_only=True)
+def aten__trilinear(
+    i1: TReal,
+    i2: TReal,
+    i3: TReal,
+    expand1: Sequence[int],
+    expand2: Sequence[int],
+    expand3: Sequence[int],
+    sumdim: Sequence[int],
+    unroll_dim: int = 1,
+) -> TReal:
+    """_trilinear(Tensor i1, Tensor i2, Tensor i3, int[] expand1, int[] expand2, int[] expand3, int[] sumdim, int unroll_dim=1) -> Tensor"""
+
+    del unroll_dim
+
+    input_rank = getattr(i1, "rank", None)
+    if input_rank is None:
+        input_rank = len(i1.shape)
+    total_dim = input_rank + len(expand1)
+    equation = _build_trilinear_equation(total_dim, expand1, expand2, expand3, sumdim)
+    return op.Einsum(i1, i2, i3, equation=equation)
+
+
 def aten_triplet_margin_loss(
     anchor: TensorType,
     positive: TensorType,

tests/function_libs/torch_lib/extra_opinfo.py

@@ -68,6 +68,35 @@ def sample_inputs_bilinear(op_info, device, dtype, requires_grad, **kwargs):
             yield opinfo_core.SampleInput(input1, args=(input2, weight, None))
 
 
+def sample_inputs__trilinear(op_info, device, dtype, requires_grad, **kwargs):
+    """Sample inputs for aten._trilinear using bilinear's internal call pattern."""
+    del op_info
+    del kwargs
+
+    make_arg = functools.partial(
+        torch_testing.make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
+    )
+
+    cases = [
+        (2, 3, 4, 5),
+        (1, 2, 2, 1),
+        (3, 5, 2, 4),
+    ]
+    expand1 = (1, 3)
+    expand2 = (0,)
+    expand3 = (1, 2)
+    sumdim = (2, 3)
+
+    for batch_size, in1_features, in2_features, out_features in cases:
+        input1 = make_arg((batch_size, in1_features))
+        weight = make_arg((out_features, in1_features, in2_features))
+        input2 = make_arg((batch_size, in2_features))
+        yield opinfo_core.SampleInput(
+            input1,
+            args=(weight, input2, expand1, expand2, expand3, sumdim, 1),
+        )
+
+
 def sample_inputs_bernoulli_p(op_info, device, dtype, requires_grad, **kwargs):
     del op_info
 
@@ -2516,6 +2545,13 @@ def __init__(self):
         sample_inputs_func=sample_inputs_bilinear,
         supports_out=False,
     ),
+    opinfo_core.OpInfo(
+        "ops.aten._trilinear.default",
+        aten_name="_trilinear.default",
+        dtypes=common_dtype.floating_types(),
+        sample_inputs_func=sample_inputs__trilinear,
+        supports_out=False,
+    ),
     opinfo_core.OpInfo(
         "ops.aten.bernoulli.p",
         aten_name="bernoulli.p",

tests/function_libs/torch_lib/ops_test_data.py

@@ -1293,6 +1293,11 @@ def _where_input_wrangler(
         dtypes=(torch.int32,),
         reason="fixme: ORT does not have an implementation of Trilu for int32.",
     ),
+    TorchLibOpInfo(
+        "ops.aten._trilinear.default",
+        core_ops.aten__trilinear,
+        tolerance={torch.float32: (2e-5, 2e-5)},
+    ),
     TorchLibOpInfo("triu", core_ops.aten_triu).xfail(
         dtypes=(torch.int32,),
         reason="fixme: ORT does not have an implementation of Trilu for int32.",