You are viewing a plain text version of this content. The canonical link for it is here.
Posted to discuss-archive@tvm.apache.org by "siyiweigeHEW (via GitHub)" <gi...@apache.org> on 2026/08/27 21:08:57 UTC

[PR] [Fix][Relax][Frontend][ONNX] Squeeze: raise an error when axes select a non-unit dimension [tvm]

siyiweigeHEW opened a new pull request, #20211:
URL: https://github.com/apache/tvm/pull/20211

   
   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](https://onnx.ai/onnx/operators/onnx__Squeeze.html)
   (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`.
   
   ```python
   # 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:
   
   ```python
   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:
   
   ```bash
   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`.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Fix][Relax][Frontend][ONNX] Squeeze: raise an error when axes select a non-unit dimension [tvm]

Posted by "tlopex (via GitHub)" <gi...@apache.org>.
tlopex commented on PR #20211:
URL: https://github.com/apache/tvm/pull/20211#issuecomment-5460768694

   Closed as already fixed. Thanks @siyiweigeHEW @OmarAzizi 


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Fix][Relax][Frontend][ONNX] Squeeze: raise an error when axes select a non-unit dimension [tvm]

Posted by "OmarAzizi (via GitHub)" <gi...@apache.org>.
OmarAzizi commented on PR #20211:
URL: https://github.com/apache/tvm/pull/20211#issuecomment-5453370515

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


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org


Re: [PR] [Fix][Relax][Frontend][ONNX] Squeeze: raise an error when axes select a non-unit dimension [tvm]

Posted by "tlopex (via GitHub)" <gi...@apache.org>.
tlopex closed pull request #20211: [Fix][Relax][Frontend][ONNX] Squeeze: raise an error when axes select a non-unit dimension
URL: https://github.com/apache/tvm/pull/20211


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org