Mr-Neutr0n opened a new pull request, #18773:
URL: https://github.com/apache/tvm/pull/18773

   ## Description
   
   Fixes #18750
   
   The ONNX frontend `Gelu` converter was hardcoded to always emit `R.nn.gelu`, 
ignoring the `approximate` attribute introduced in ONNX Opset 20. When a model 
uses `approximate="tanh"`, TVM would silently produce incorrect results by 
applying the exact erf-based Gelu instead of the tanh approximation.
   
   ## Changes
   
   Added `_impl_v20` to the `Gelu` converter class that checks the 
`approximate` attribute:
   - `approximate="tanh"` → routes to `relax.op.nn.gelu_tanh`
   - `approximate="none"` (default) → routes to `relax.op.nn.gelu`
   
   The existing `_impl_v1` is left untouched for backward compatibility with 
the Microsoft onnxruntime contrib opset. The version dispatch mechanism will 
automatically select `_impl_v20` for Opset 20+ models.
   
   ## Testing
   
   - The change is a straightforward attribute check that routes to an 
already-existing and tested operator (`R.nn.gelu_tanh`).
   - No new dependencies introduced.


-- 
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: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to