How can I transfer t5 decoder_model_merged.onnx to tensorrt

[EASTERNTIGER]

Hi,when I try to use command:
trtexec --onnx=decoder_model_merged.onnx --saveEngine=decoder_model_merged.trt
in linux,it showed:
[07/30/2024-02:00:24] [E] Error[4]: [graphShapeAnalyzer.cpp::analyzeShapes::2084] Error Code 4: Miscellaneous (IConditionalOutputLayer optimum::if_OutputLayer_4945: optimum::if_OutputLayer_4945: dimensions not compatible for if-conditional outputs)
[07/30/2024-02:00:24] [E] Engine could not be created from network
[07/30/2024-02:00:24] [E] Building engine failed
[07/30/2024-02:00:24] [E] Failed to create engine from model or file.
[07/30/2024-02:00:24] [E] Engine set up failed

when I use the same command:
trtexec --onnx=decoder_model.onnx --saveEngine=decoder_model.trt
it works , so why in the same T5 onnx model,decoder_model.onnx、encoder_model.onnx、decoder_with_past_model.onnx could be transferred to trt successfully, decoder_model_merged.onnx could not,How can I fixed that?

[lix19937]

dimensions not compatible for if-conditional outputs)

Your model has if-then-else flow control. There can be no cross-edges connecting layers in the true-branch to layers in the false-branch, and vice versa. In other words, the outputs of one branch cannot depend on layers in the other branch.

You should export scripted PyTorch code to ONNX. Like follow:

import torch.onnx
import torch
import tensorrt as trt
import numpy as np

TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)

@torch.jit.script
def sum_even(items):
    s = torch.zeros(1, dtype=torch.float)
    for c in items:
        if c % 2 == 0:
            s += c
    return s

class ExampleModel(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, items):
        return sum_even(items)

def build_engine(model_file):
    builder = trt.Builder(TRT_LOGGER)
    network = builder.create_network(EXPLICIT_BATCH)
    config = builder.create_builder_config()
    parser = trt.OnnxParser(network, TRT_LOGGER)

    with open(model_file, 'rb') as model:
        assert parser.parse(model.read())
        return builder.build_engine(network, config)

def export_to_onnx():
    items = torch.zeros(4, dtype=torch.float)
    example = ExampleModel()
    torch.onnx.export(example, (items), "example.onnx", verbose=False, opset_version=13, enable_onnx_checker=False, do_constant_folding=True)

export_to_onnx()
build_engine("example.onnx")  # 
print("done") # lix19937 ---   

Or you replace the if-else control with fixed control branch.

[EASTERNTIGER]

dimensions not compatible for if-conditional outputs)

Your model has if-then-else flow control. There can be no cross-edges connecting layers in the true-branch to layers in the false-branch, and vice versa. In other words, the outputs of one branch cannot depend on layers in the other branch.

You should export scripted PyTorch code to ONNX. Like follow:

import torch.onnx
import torch
import tensorrt as trt
import numpy as np

TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)

@torch.jit.script
def sum_even(items):
    s = torch.zeros(1, dtype=torch.float)
    for c in items:
        if c % 2 == 0:
            s += c
    return s

class ExampleModel(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, items):
        return sum_even(items)

def build_engine(model_file):
    builder = trt.Builder(TRT_LOGGER)
    network = builder.create_network(EXPLICIT_BATCH)
    config = builder.create_builder_config()
    parser = trt.OnnxParser(network, TRT_LOGGER)

    with open(model_file, 'rb') as model:
        assert parser.parse(model.read())
        return builder.build_engine(network, config)

def export_to_onnx():
    items = torch.zeros(4, dtype=torch.float)
    example = ExampleModel()
    torch.onnx.export(example, (items), "example.onnx", verbose=False, opset_version=13, enable_onnx_checker=False, do_constant_folding=True)

export_to_onnx()
build_engine("example.onnx")  # 
print("done") # lix19937 ---   

Or you replace the if-else control with fixed control branch.

thank you so much for your reply!I will try that

[poweiw]

Closing this issue for now but feel free to reopen if you see further issues!