onnxruntime

c156e933 - [webgpu] Correct definition of large numbers, fixes softmax(max_negative_number) in float32 (#26670)

[webgpu] Correct definition of large numbers, fixes softmax(max_negative_number) in float32 (#26670) ### Description The correct definition of the most negative number is `-3.40282346638528e+38`, according to IEEE 754, but it is being incorrectly registered inline as a truncated version `-3.402823e+38f`. ```py >>> import numpy as np >>> np.finfo(np.float32).min np.float32(-3.4028235e+38) >>> np.finfo(np.float32).min.item() -3.4028234663852886e+38 ``` For this reason, values less than this threshold were handled incorrectly. While this may seem like a small/irrelevant detail, it's essential in attention masking, where we do in fact use this value, leading to large numerical errors down the line. Reproduction: ```py from onnx import helper, TensorProto import onnxruntime as ort import numpy as np # 1. Create the ONNX model # Define input and output input_shape = [1, 2] input_info = helper.make_tensor_value_info('X', TensorProto.FLOAT, input_shape) output_info = helper.make_tensor_value_info('Y', TensorProto.FLOAT, input_shape) # Create the Softmax node # Softmax takes one input: X softmax_node = helper.make_node( 'Softmax', inputs=['X'], outputs=['Y'], name='SoftmaxNode', axis=-1 # Default axis is -1, usually applied to the last dimension ) # Create the graph graph_def = helper.make_graph( [softmax_node], 'test-model', [input_info], [output_info] ) # Create the model model_def = helper.make_model(graph_def, producer_name='onnx-example') opset = model_def.opset_import[0] opset.version = 13 # Ensure opset version supports the operations # 2. Convert model to string (bytes) model_str = model_def.SerializeToString() # 3. Prepare input data np.random.seed(0) input_data = np.array( [[-3.40282346638528e+38, -3.40282346638528e+38]] # [[-3.4028234663852886e+38, -3.4028234663852886e+38]] ).astype(np.float32) print(input_data.tolist()) # 4. Run on CPUExecutionProvider sess_cpu = ort.InferenceSession(model_str, providers=['CPUExecutionProvider']) res_cpu = sess_cpu.run(['Y'], {'X': input_data})[0] print("CPU Result:", res_cpu) # 5. Run on WebGpuExecutionProvider sess_webgpu = ort.InferenceSession(model_str, providers=['WebGpuExecutionProvider']) res_webgpu = sess_webgpu.run(['Y'], {'X': input_data})[0] print("WebGPU Result:", res_webgpu) # Compare results diff = np.abs(res_cpu - res_webgpu) max_diff = diff.max().item() print(diff) print(f"Max diff: {max_diff}") assert max_diff < 1e-5, f"Results do not match within tolerance! Max diff: {max_diff}" print("Results match!") ``` Before: ``` [[-3.4028234663852886e+38, -3.4028234663852886e+38]] CPU Result: [[0.5 0.5]] WebGPU Result: [[0. 0.]] [[0.5 0.5]] Max diff: 0.5 AssertionError: Results do not match within tolerance! Max diff: 0.5 ``` After: ``` [[-3.4028234663852886e+38, -3.4028234663852886e+38]] CPU Result: [[0.5 0.5]] WebGPU Result: [[0.5 0.5]] [[0. 0.]] Max diff: 0.0 Results match! ``` cc @guschmue