[torchlib] Implement torch.ops.prims.broadcast_in_dim.default (#2382)

This PR implements the missing
`torch.ops.prims.broadcast_in_dim.default` operation that appears in
BERT_pytorch and other PyTorch models.

## Overview

The `broadcast_in_dim` operation is a primitive that broadcasts a tensor
to a target shape by specifying which dimensions of the output
correspond to the input tensor dimensions. This is different from
standard broadcasting operations.

## Implementation Details

**Function signature:**
```python
def prims_broadcast_in_dim(
    a: TensorType, shape: INT64, broadcast_dimensions: Sequence[int]
) -> TensorType:
```

**Parameters:**
- `a`: Input tensor to broadcast
- `shape`: Target output shape  
- `broadcast_dimensions`: Specifies which dimensions of the output shape
correspond to the input tensor dimensions

**Example:**
```python
# Input tensor: [3, 4]
# Target shape: [2, 3, 5, 4] 
# broadcast_dimensions: [1, 3]
# Result: Input dimension 0 (size 3) maps to output dimension 1
#         Input dimension 1 (size 4) maps to output dimension 3
#         Output dimensions 0 and 2 are broadcasted (filled from size 1)
```
Fixes #2218. Fix pytorch/pytorch#135343

---------

Signed-off-by: Justin Chu <[email protected]>
Co-authored-by: copilot-swe-agent[bot] <[email protected]>
Co-authored-by: justinchuby <[email protected]>
Co-authored-by: Justin Chu <[email protected]>

Author: Copilot

onnxscript/function_libs/torch_lib/ops/prims.py

@@ -176,12 +176,33 @@ def prims_bitwise_xor(self: TensorType, other: TensorType) -> TensorType:
     raise NotImplementedError()
 
 
+@torch_op("prims::broadcast_in_dim", trace_only=True)
 def prims_broadcast_in_dim(
-    a: TensorType, shape: INT64, broadcast_dimensions: Sequence[int]
+    a: TensorType, shape: Sequence[INT64], broadcast_dimensions: Sequence[int]
 ) -> TensorType:
     """broadcast_in_dim(Tensor(a) a, SymInt[] shape, int[] broadcast_dimensions) -> Tensor(a)"""
 
-    raise NotImplementedError()
+    target_rank = len(shape)
+
+    if not broadcast_dimensions:
+        # Special case: no broadcast dimensions - all target dims should be 1
+        return op.Expand(a, common_ops.merge_dims(shape))
+
+    # Create base shape of all 1s
+    ones = [1] * target_rank
+
+    # For each broadcast dimension, we'll replace the 1 with the actual input dimension
+    # Since broadcast_dimensions is compile-time known, we can do this with individual operations
+    intermediate_shape = ones
+
+    for i, broadcast_dim in enumerate(broadcast_dimensions):
+        # Get the input dimension value
+        input_dim_value = op.Shape(a, start=i, end=i + 1)
+        intermediate_shape[broadcast_dim] = input_dim_value
+
+    # Reshape input to intermediate shape and expand to target
+    reshaped = op.Reshape(a, common_ops.merge_dims(intermediate_shape))
+    return op.Expand(reshaped, shape)
 
 
 def prims_cat(tensors: Sequence[TensorType], dim: int) -> TensorType:

tests/function_libs/torch_lib/extra_opinfo.py

@@ -87,6 +87,35 @@ def sample_inputs_bernoulli_p_deterministic(op_info, device, dtype, requires_gra
             yield opinfo_core.SampleInput(t, kwargs={"p": p})
 
 
+def sample_inputs_broadcast_in_dim(op_info, device, dtype, requires_grad, **kwargs):
+    del op_info
+    del kwargs
+
+    # cases: (input_shape, target_shape, broadcast_dimensions)
+    # broadcast_dimensions maps each input dim to an axis in target_shape
+    cases = (
+        # scalar -> 1-D tensor
+        ((), (3,), ()),
+        # identity (no-op broadcast)
+        ((3,), (3,), (0,)),
+        # rank-preserving broadcast where singleton dims expand
+        ((1, 3, 1), (2, 3, 4), (0, 1, 2)),
+        # input rank 2 -> output rank 3, input dims map to trailing axes
+        ((3, 1), (2, 3, 4), (1, 2)),
+        # add leading broadcast axis
+        ((3, 4), (1, 3, 4), (1, 2)),
+        # insert broadcasting in middle axis
+        ((3,), (2, 3, 1), (1,)),
+    )
+    make_arg = functools.partial(
+        torch_testing.make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
+    )
+
+    for shape, target_shape, broadcast_dimensions in cases:
+        tensor = make_arg(shape)
+        yield opinfo_core.SampleInput(tensor, args=(target_shape, broadcast_dimensions))
+
+
 def sample_inputs_col2im(op_info, device, dtype, requires_grad, **kwargs):
     del op_info
     # input_shape, output_size, kernal, dilation, padding, stride
@@ -2687,6 +2716,13 @@ def __init__(self):
         sample_inputs_func=sample_inputs_upsample_trilinear3d_vec,
         supports_out=False,
     ),
+    opinfo_core.ReductionOpInfo(
+        "ops.prims.broadcast_in_dim.default",
+        op=torch.ops.prims.broadcast_in_dim.default,
+        dtypes=common_dtype.all_types(),
+        sample_inputs_func=sample_inputs_broadcast_in_dim,
+        supports_out=False,
+    ),
     opinfo_core.ReductionOpInfo(
         "ops.prims.var.default",
         nan_policy="propagate",

tests/function_libs/torch_lib/ops_test_data.py

@@ -2136,6 +2136,7 @@ def _where_input_wrangler(
             "Our implementation is based on that for CUDA"
         ),
     ),
+    TorchLibOpInfo("ops.prims.broadcast_in_dim.default", prims_ops.prims_broadcast_in_dim),
     TorchLibOpInfo(
         "ops.prims.var.default", prims_ops.prims_var, tolerance={torch.float16: (1e-3, 5e-2)}
     ),