huenwei-arch opened a new issue, #18750:
URL: https://github.com/apache/tvm/issues/18750
### Expected behavior
The TVM ONNX frontend should correctly handle the `approximate` attribute
for the `Gelu` operator (introduced in Opset 20).
- If `approximate="none"` (default), it should map to `relax.op.nn.gelu`,
using the exact CDF formula:
$$y = 0.5x(1 + erf(\frac{x}{\sqrt{2}}))$$
- If `approximate="tanh"`, it should map to `relax.op.nn.gelu_tanh`, using
the Tanh approximation:
$$y = 0.5x(1 + \tanh(\sqrt{\frac{2}{\pi}}(x + 0.044715x^3)))$$
### Actual behavior
The TVM ONNX frontend currently ignores the `approximate` attribute and
hardcodes the mapping to `R.nn.gelu` (the exact version).
This leads to a systematic numerical mismatch when the source model expects
the Tanh-based approximation. For example, at $x = -1.0$:
- **Exact (TVM current):** $\approx -0.158655$
- **Tanh Approx (Expected):** $\approx -0.158808$
- **Delta:** $\approx 1.5 \times 10^{-4}$ (Exceeds `float32` tolerance for
identical algorithms).
**Observed Relax IR (incorrectly mapped to `gelu`):**
```python
@R.function
def main(X: R.Tensor((4,), dtype="float32")) -> R.Tensor((4,),
dtype="float32"):
R.func_attr({"num_input": 1})
with R.dataflow():
# Bug: Should be R.nn.gelu_tanh when approximate="tanh"
gv: R.Tensor((4,), dtype="float32") = R.nn.gelu(X)
R.output(gv)
return gv
```
### Environment
OS: Ubuntu 20.04.6 LTS (Focal Fossa)
TVM Version: 0.19.0 (Relax)
ONNX Version: 1.18.0
ONNX Runtime Version: 1.24.1
NumPy Version: 2.4.2
### Steps to reproduce
```python
import onnx
from onnx import helper, TensorProto
import numpy as np
import tvm
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx
import onnxruntime as ort
def reproduce_gelu_tanh_mismatch():
"""
Reproduce the issue where TVM ONNX frontend ignores the 'approximate'
attribute
of the Gelu operator (Opset 20).
"""
# 1. Create an ONNX model with Gelu (approximate='tanh')
# According to ONNX Opset 20, this should trigger the Tanh-based
approximation.
dtype_proto = TensorProto.FLOAT
dtype_np = 'float32'
node_def = helper.make_node(
'Gelu',
inputs=['X'],
outputs=['Y'],
approximate='tanh'
)
graph_def = helper.make_graph(
[node_def],
'gelu-tanh-repro',
[helper.make_tensor_value_info('X', dtype_proto, [4])],
[helper.make_tensor_value_info('Y', dtype_proto, [4])],
)
opset_info = helper.make_opsetid("", 20)
model_def = helper.make_model(
graph_def,
ir_version=9,
opset_imports=[opset_info]
)
# 2. Prepare input data
x_np = np.array([-1.0, 0.0, 1.0, 2.0], dtype=dtype_np)
# 3. Get reference output from ONNX Runtime
sess = ort.InferenceSession(model_def.SerializeToString())
onnx_outputs = sess.run(None, {'X': x_np})[0]
# 4. Get TVM Relax output
tvm_mod = from_onnx(model_def)
# Check IR: If bug exists, it will show R.nn.gelu instead of
R.nn.gelu_tanh
print("--- Generated Relax IR ---")
print(tvm_mod["main"].script())
target = tvm.target.Target("llvm")
dev = tvm.cpu(0)
ex = relax.build(tvm_mod, target)
vm = relax.VirtualMachine(ex, dev)
tvm_outputs = vm["main"](tvm.nd.array(x_np, dev)).asnumpy()
# 5. Compare Results
print(f"Reference (ORT): {onnx_outputs}")
print(f"TVM Relax: {tvm_outputs}")
try:
np.testing.assert_allclose(
onnx_outputs,
tvm_outputs,
rtol=1e-5, atol=1e-5,
err_msg="Numerical mismatch! TVM likely ignored
approximate='tanh'."
)
print("\nResult: Success (No mismatch detected)")
except AssertionError as e:
print("\nResult: Bug Confirmed (Numerical mismatch detected)")
print(e)
if __name__ == "__main__":
reproduce_gelu_tanh_mismatch()
```
### Triage
* relax:frontend:onnx
* needs-triage
--
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]
