[Draft] feat: flag-gated upstream complex decomposition adapter

[apbose]

Adds an alternative complex-numerics lowering path that delegates the interior complex->real expansion to PyTorch's upstream decomposition (decompose_complex_in_graph, pytorch/pytorch#169832) instead of the hand-maintained per-op rewriter in complex_graph_rewrite.py.

The new complex_decomposition_adapter pass:

• captures the complex I/O signature (reusing the existing helpers) so the post-partition _insert_complex_io_adapters boundary code works unchanged;
• calls decompose_complex_in_graph on the exported graph, feeding it the placeholders' fake tensors to preserve the export ShapeEnv/SymInts;
• normalizes the SoA aten.complex/real/imag seams it leaves behind into the interleaved [...,2] layout the rest of the TRT flow expects.

Wiring:

• use_complex_decomposition setting (default False) selects new vs legacy pass via complex_lowering_pass in _aten_lowering_pass;
• torch-version capability gate has_complex_decomposition() in _features (torch>=2.14.dev), with graceful fallback to complex_graph_detection.

Tests:

• test_complex_ops.py parametrized [legacy]/[decomp] via an autouse fixture so the existing suite runs on both paths (decomp auto-skips on torch<2.14);
• new GPU-free test_complex_decomposition_adapter.py covering the SoA->[...,2] normalization and the fallback gate.

Default off; legacy path unchanged. Boundary normalization assumptions (exact upstream op set, dynamic-shape survival through make_fx) still to be confirmed by a spike; see TODO(spike) in the adapter.

#4390

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/_aten_lowering_pass.py	2026-07-23 21:36:54.866947+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/_aten_lowering_pass.py	2026-07-23 21:37:15.722131+00:00
@@ -42,10 +42,11 @@
    otherwise the legacy hand-rolled rewriter runs.
    """
    if getattr(settings, "use_complex_decomposition", False):
        return complex_decomposition_adapter(gm, settings)
    return complex_graph_detection(gm, settings)
+

post_lowering_pass_list = [
    replace_fused_rms_norm,
    remove_input_alias_fixing_clones,
    constant_fold,
--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/lowering/test_complex_decomposition_adapter.py	2026-07-23 21:36:54.908527+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/lowering/test_complex_decomposition_adapter.py	2026-07-23 21:37:19.335969+00:00
@@ -69,13 +69,11 @@
    assert aten.complex.default not in targets
    assert aten.cat.default in targets
    assert aten.unsqueeze.default in targets

    # the packed cat node must carry the complex-layout tag
-    cat_nodes = [
-        n for n in gm.graph.nodes if n.target == aten.cat.default
-    ]
+    cat_nodes = [n for n in gm.graph.nodes if n.target == aten.cat.default]
    assert cat_nodes and cat_nodes[0].meta.get("is_complex_layout") is True

    # numerically: output is the [...,2] interleaved layout of re/im
    out = gm(re, im)
    assert out.shape == (3, 2)