[BugFix][Relax][ONNX] Resolve param Vars in Concat to handle mixed Shape/Tensor inputs

[swjng]

Description

When from_onnx(model, keep_params_in_input=True) is used, every ONNX initializer becomes a relax.Var instead of a relax.Constant. The Concat handler's is_shape_like() check only recognizes relax.ShapeExpr and 1D-int64 relax.Constant, so a 1D-int64 shape value loaded as a Var is no longer recognized.

When such a Var is concatenated with a ShapeExpr — the standard pattern for dynamic-batch Reshape in PyTorch-exported ONNX models — the heterogeneous Tuple(ShapeExpr, Tensor) is rejected by relax.op.concat with:

InternalError: Op(relax.concat) expects the input to be a Tuple of Tensors.
However, the given input is R.Tuple(R.Shape([N]), R.Tensor((1,), dtype="int64"))

This effectively breaks keep_params_in_input=True for any model with dynamic-batch Reshape (extremely common in PyTorch ONNX exports).

Fix

Run each Concat input through the existing get_constant helper before the is_shape_like check. This resolves any Var that maps to a known param back to its baked Constant, restoring the all-shape-like fast path.

Minimal repro

An 8-node ONNX graph (Shape → Slice → Concat([dyn_n, [12]]) → Reshape) fails with keep_params_in_input=True before this PR and passes after. A regression test (test_concat_with_param_shape_value) covers this pattern.

Testing

pytest tests/python/relax/test_frontend_onnx.py -k concat

9 passed (1 new + 8 existing).

[gemini-code-assist]

Code Review

This pull request updates the ONNX frontend's Concat operator to resolve input parameters to their constant values, ensuring that 1D-int64 shape values can correctly follow the shape-like fast path when keep_params_in_input is set to true. This change specifically addresses issues with dynamic-batch Reshape patterns in PyTorch-exported models. A new test case has been added to verify this fix. I have no further feedback to provide.

[tlopex]

Hi @swjng
Thanks for the fix!

However, Concat._impl_v13 now calls get_constant on every input unconditionally. get_constant has a side effect: when the input is a param Var, it updates graph_nodes[name] to a baked relax.Constant.

This fixes the shape-construction case, but it can also change the semantics of ordinary tensor concat under keep_params_in_input=True. For example, an ONNX graph with Concat(x, weight) should keep weight as a runtime parameter. With this change, weight may be folded into the IR as a constant, so the compiled function no longer depends on the runtime value passed for that parameter.

Could we narrow the fix so param resolution is used only for the all-shape-like fast path, preferably without mutating _nodes? The tensor fallback should continue using the original inputs. It would also be good to add a regression test for normal tensor Concat(input, initializer) with keep_params_in_input=True, verifying that the initializer remains a real parameter input.

[swjng]

Thanks. Narrowed it as suggested in the follow-up commit: a local non-mutating resolve returns a Constant only for 1D-int64 param Vars, and only the all-shape-like fast path consumes the resolved list. The tensor fallback uses the original inputs, so Concat(input, weight) still references weight as a runtime Var.

Added test_concat_with_param_tensor_keeps_runtime_param covering exactly that case (asserts w stays in main's params and is recovered by detach_params).

[tlopex]

LGTM! Thank you