Fix attention mask to use float_lowest instead of -inf and add unit test for softmax NaN case

Author: Aniketsy

onnxscript/function_libs/torch_lib/ops/nn.py

@@ -14,6 +14,7 @@
 
 from __future__ import annotations
 
+import numpy as np
 import math
 from typing import Optional, Sequence, Tuple, TypeVar, Union
 
@@ -2048,6 +2049,9 @@ def _aten_scaled_dot_product_attention_no_mask_onnx(
         attn_weight, _ = op.Dropout(attn_weight, dropout_p)
     return op.MatMul(attn_weight, value)
 
+def float_lowest(dtype):
+    """Returns the lowest representable value for the given numpy dtype."""
+    return np.finfo(np.dtype(dtype)).min
 
 def _aten_scaled_dot_product_attention_bool_mask_onnx(
     query: TFloat,
@@ -2078,7 +2082,7 @@ def _aten_scaled_dot_product_attention_bool_mask_onnx(
     key_transposed_scaled = op.Mul(key_transposed, op.Sqrt(scale))
     # Turn the Boolean mask to float: attn_mask.masked_fill(not attn_mask, -float('inf'))
     zero = op.Constant(value=ir.tensor(0.0, dtype=query.dtype))
-    neg_inf = op.Constant(value=ir.tensor(-float("inf"), dtype=query.dtype))
+    neg_inf = op.Constant(value=ir.tensor(float_lowest(query.dtype)), dtype=query.dtype)
     attn_mask = op.Where(attn_mask, zero, neg_inf)
     attn_weight = op.Softmax(
         op.Add(op.MatMul(query_scaled, key_transposed_scaled), attn_mask),

tests/common/testutils.py

@@ -14,6 +14,7 @@
 import torch
 
 from onnxscript import optimizer
+from onnxscript.onnx_opset import opset18 as op
 from onnxscript.rewriter import onnxruntime as ort_rewriter
 from onnxscript.utils import evaluation_utils
 
@@ -101,3 +102,9 @@ def test_onnxruntime_rewrite(
                     f"Failed for model {model_name} and output {i} with rtol={rtol} and atol={atol}\n{e}"
                 )
                 raise
+
+def test_softmax_with_all_inf_mask():
+    # GH #2561
+    input = np.array([[-float("inf"), -float("inf")]], dtype=np.float32)
+    output = op.Softmax(input, axis=-1)
+    assert np.isnan(output).all(), "Softmax should return NaN when all inputs are -inf"