feat: refactor four attention converters with TRT add_attention_v2()

py/torch_tensorrt/_utils.py

@@ -40,6 +40,22 @@ def check_cross_compile_trt_win_lib() -> bool:
     return False
 
 
+def is_tensorrt_rtx_version_supported(min_version: str) -> bool:
+    """
+    Check if the installed TensorRT-RTX version supports the specified minimum version.
+    Args:
+        min_version (str): Minimum required TensorRT-RTX version
+    Returns:
+        bool: True if TensorRT-RTX version is >= min_version, False otherwise
+    """
+    if trt._package_name != "tensorrt_rtx":
+        return True
+
+    from packaging.version import Version
+
+    return bool(Version(trt.__version__) >= Version(min_version))
+
+
 def is_tensorrt_version_supported(min_version: str) -> bool:
     """
     Check if the installed TensorRT version supports the specified minimum version.

py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py

@@ -4086,6 +4086,12 @@ def aten_ops_linear(
 def scaled_dot_product_attention_validator(
     node: Node, settings: Optional[CompilationSettings] = None
 ) -> bool:
+    attn_mask = args_bounds_check(node.args, 3, None)
+    is_causal = args_bounds_check(node.args, 5, False)
+    if is_causal and attn_mask is not None:
+        _LOGGER.debug("Explicit attn_mask should not be set when is_causal=True.")
+        return False
+
     enable_gqa = node.kwargs.get("enable_gqa", False)
 
     query_shape, key_shape, value_shape = None, None, None