[ONNX] Create fake implementations for onnx ops; fix boolean mask in attention (pytorch#165780)

Previously we rely on the concreate implementation to generate fake implementation. This makes the fake implementation overly complicated and breaks in some cases when there are dynamic shapes.

This PR updates onnx op registration to instead take a dedicated fake implementation.

**Also fixed: When boolean mask is supplied to torch sdpa, it was previously taken the negation, which is incorrect.**

Fix pytorch#164909 Also taken changes from pytorch#156635

Pull Request resolved: pytorch#165780
Approved by: https://github.com/titaiwangms

Author: justinchuby

test/onnx/exporter/test_small_models_e2e.py

@@ -5,22 +5,16 @@
 
 import logging
 
-import onnxruntime
 import pytest
 import transformers
 from onnxscript import ir
-from packaging import version
 
 import torch
 from torch.onnx._internal.exporter import _testing as onnx_testing
 from torch.testing._internal import common_utils
 from torch.utils import _pytree as torch_pytree
 
 
-def has_onnxruntime_opset_23() -> bool:
-    return version.parse(onnxruntime.__version__) >= version.parse("1.23")
-
-
 class _WithExport:
     def export(self, model, args=(), kwargs=None, **options) -> torch.onnx.ONNXProgram:
         onnx_program = torch.onnx.export(
@@ -746,13 +740,7 @@ def forward(self, query, key, value):
         onnx_program = self.export(Model(), (query, key, value), opset_version=23)
         self.assertEqual(["Attention"], [n.op_type for n in onnx_program.model.graph])
 
-        if has_onnxruntime_opset_23():
-            onnx_testing.assert_onnx_program(onnx_program, atol=1e-2, rtol=1)
-        else:
-            # Test with reference evaluator because ORT does not support the op as of version 1.22
-            onnx_testing.assert_onnx_program(
-                onnx_program, atol=1e-2, rtol=1, backend="reference"
-            )
+        onnx_testing.assert_onnx_program(onnx_program, atol=1e-2, rtol=1)
 
     def test_rms_norm(self):
         """Test RMS normalization with various configurations."""
@@ -789,8 +777,7 @@ def forward(self, x):
 
         onnx_program = self.export(RMSNormWithWeight(), (x,), opset_version=23)
 
-        # Test with reference evaluator because ORT does not support the op as of version 1.22
-        onnx_testing.assert_onnx_program(onnx_program, backend="reference")
+        onnx_testing.assert_onnx_program(onnx_program)
 
     def test_rms_norm_with_eps(self):
         """Test RMS normalization with custom epsilon."""
@@ -803,7 +790,6 @@ def forward(self, x):
 
         onnx_program = self.export(RMSNormWithEps(), (x,), opset_version=23)
 
-        # Test with reference evaluator because ORT does not support the op as of version 1.22
         onnx_testing.assert_onnx_program(onnx_program, backend="reference")
 
     def test_enable_gqa_in_attention_23_with_dropout(self):