[siyiweigeHEW]

Fixes: #20185

Summary

The Relax ONNX frontend silently accepted a Squeeze whose axes select a
dimension with size not equal to 1, returning the input unchanged with no
error. The ONNX Squeeze spec
(opset 11+ and earlier) requires the opposite: "If an axis is selected with shape
entry not equal to one, an error is raised." Both onnxruntime and
onnx.reference reject such models; TVM diverged. The converter now validates
statically-known axes against the input shape and raises a ValueError before
delegating to relax.op.squeeze.

Root cause

Squeeze._squeeze (python/tvm/relax/frontend/onnx/onnx_frontend.py) delegates
the non-constant path directly to relax.op.squeeze(data, list(axis)). The
underlying op's shape inference
(src/relax/op/tensor/manipulate.cc:1269-1284) deliberately implements PyTorch
semantics — "If a dimension is not 1, silently skip it (no-op), matching PyTorch
behavior" (and carries a Todo(relax-team): revisit here for better check on if the axis being squeezed has length 1). So a non-unit selected axis was silently
ignored and the input returned unchanged. The ONNX frontend performed no
dim == 1 check — unlike its constant path, where np.squeeze already raises
the same ValueError as onnx.reference.

# Before: silently accepted e.g. data(2,3) with axes=[0]
if isinstance(axis, tuple):
    return relax.op.squeeze(data, list(axis))

Fix

Validate before delegating: when axis is statically known (a tuple) and the
input shape is statically known, every selected (normalized, deduplicated,
bounds-checked) axis must have size 1:

if isinstance(axis, tuple):
    cls._validate_axes_unit_dims(data, axis)
    return relax.op.squeeze(data, list(axis))

_validate_axes_unit_dims reuses _normalize_constant_axes (negative-axis
normalization, bounds and uniqueness checks) and raises the same ValueError
message as np.squeeze/onnx.reference when a known dim is != 1. Dimensions
that are only known at runtime cannot be validated at import time and keep the
previous behavior — the shared relax.op.squeeze is intentionally left unchanged
because other frontends (e.g. PyTorch) rely on its no-op semantics.

Validation

Differential test: Relax (build + VirtualMachine) vs onnxruntime on the same
model, cross-checked with onnx.reference (which is the reference for the
rejection semantics and for fp64), across valid, invalid-axis, constant, and
out-of-bounds cases.

Case	onnxruntime	onnx.reference	TVM (after fix)	Result
(2,3) axes=[0] (non-unit)	ERR	ERR	ValueError	OK
(1,2,3) axes=[1] (non-unit)	ERR	ERR	ValueError	OK
(1,2,1,3) axes=[1] (only non-unit)	ERR	ERR	ValueError	OK
(2,3,1) axes=[0,2] (mixed)	ERR	ERR	ValueError	OK
(2,3) axes=[-1] (neg, non-unit)	ERR	ERR	ValueError	OK
(1,2,1,3) axes=[0,2] (valid, regression)	(2,3)	(2,3)	(2,3) max|diff|=0	OK
(1,2,1) axes=[-1] (valid neg, regression)	(1,2)	(1,2)	(1,2) max|diff|=0	OK
(1,) axes=[0] (scalar output, regression)	()	()	() max|diff|=0	OK

Totals: 25/25 OK (pre-fix: 6/6 legal models diverged — onnxruntime/ref
rejected, TVM silently returned the input unchanged). Valid axes (incl. negative
axes, scalar output, opset 13/21), the constant-data path, and out-of-bounds
rejection are unchanged. Run:

TVM_LIBRARY_PATH=<tvm23>/build PYTHONPATH=<patched 262c6d2e04 source>/python \
  python results/TVM/deepseek-v4-flash/prove_hum/onnx_Squeeze/5修复_差分验证.py

New in-tree tests cover the raise behavior for both opset 11 (axes attribute) and
opset 13 (axes input), the constant path, and numeric correctness of valid axes
vs onnxruntime.

Files changed

• python/tvm/relax/frontend/onnx/onnx_frontend.py — Squeeze._squeeze: validate
  that every selected axis has a size-1 dimension (ONNX spec) before delegating to
  relax.op.squeeze.
• tests/python/relax/test_frontend_onnx.py — add test_squeeze_non_unit_axis_raises,
  test_squeeze_constant_non_unit_axis_raises, test_squeeze_valid_axes_correctness.

[OmarAzizi]

Fixed in #20188, sorry I didn't know you were working on it

[tlopex]

Closed as already fixed. Thanks @siyiweigeHEW @OmarAzizi