Expand doesn't handle dynamic shaped tensors.

[jxchenus]

System Info

AWS p3.8xlarge (https://aws.amazon.com/ec2/instance-types/p3/)

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

test_expand_dyn.py.txt

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest

import numpy as np

# isort: off
import torch
import tensorrt as trt
# isort: on
from polygraphy.backend.trt import (CreateConfig, EngineFromNetwork, Profile,
                                    TrtRunner)

import tensorrt_llm
from tensorrt_llm import Tensor


class TestFunctional(unittest.TestCase):

    def setUp(self):
        tensorrt_llm.logger.set_level('error')

    def test_expand_dyn_input(self):
        # test data
        dtype = 'float32'
        input_shape = (-1, 10)

        # construct trt network
        builder = tensorrt_llm.Builder()
        net = builder.create_network()
        with tensorrt_llm.net_guard(net):
            network = tensorrt_llm.default_trtnet()
            input = Tensor(name='input',
                           shape=input_shape,
                           dtype=tensorrt_llm.str_dtype_to_trt(dtype))
            shape = tensorrt_llm.functional.constant(np.array([1, 10]))
            output = tensorrt_llm.functional.expand(input, shape)
            output_shape = tensorrt_llm.functional.shape(output, 0)

        # trt run
        profiles = [Profile().add('shape', (1, 1), input_shape, (10, 10))]
        build_engine = EngineFromNetwork((builder.trt_builder, net.trt_network),
                                         config=CreateConfig(profiles=profiles))

    def test_expand_dyn_inout(self):
        # test data
        dtype = 'float32'
        input_shape = (-1, 10)

        # construct trt network
        builder = tensorrt_llm.Builder()
        net = builder.create_network()
        with tensorrt_llm.net_guard(net):
            network = tensorrt_llm.default_trtnet()
            input = Tensor(name='input',
                           shape=input_shape,
                           dtype=tensorrt_llm.str_dtype_to_trt(dtype))
            shape = tensorrt_llm.functional.constant(np.array([-1, 10]))
            output = tensorrt_llm.functional.expand(input, shape)
            output_shape = tensorrt_llm.functional.shape(output, 0)

        # trt run
        profiles = [Profile().add('shape', (1, 1), input_shape, (10, 10))]
        build_engine = EngineFromNetwork((builder.trt_builder, net.trt_network),
                                         config=CreateConfig(profiles=profiles))

Rename the above file to test_expand_dyn.py and run:

python3 -m unittest test_expand_dyn.py

Expected behavior

The test should succeed.

actual behavior

The test fails with the following:

[TensorRT-LLM] TensorRT-LLM version: 0.10.0
[07/03/2024-22:05:51] [TRT] [E] ITensor::getDimensions: Error Code 4: Internal Error (Tensor (Unnamed Layer* 1) [Slice]_output has axis 0 with inherently negative length. Proven upper bound is -1. Network must have an instance where axis has non-negative length.)
[07/03/2024-22:05:51] [TRT] [E] ITensor::getDimensions: Error Code 4: Internal Error (Output shape can not be computed for node SHAPE_4.)
E.
======================================================================
ERROR: test_expand_dyn_inout (test_expand_dyn.TestFunctional)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/TensorRT-LLM/tests/functional/test_expand_dyn.py", line 72, in test_expand_dyn_inout
    output_shape = tensorrt_llm.functional.shape(output, 0)
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/functional.py", line 1779, in shape
    return gather(res, dim=0, indices=dim).view([])
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/functional.py", line 1831, in gather
    assert input.rank() == indices.rank()
  File "/usr/local/lib/python3.10/dist-packages/tensorrt_llm/functional.py", line 476, in rank
    return len(self.trt_tensor.shape)
ValueError: __len__() should return >= 0

----------------------------------------------------------------------
Ran 2 tests in 1.425s

FAILED (errors=1)

additional notes

When compiling a llama-like model using trtllm-build, tensors with a batch size dimension have that dimension set to -1. But operators like expand and inherently slice seem to have issues supporting an input or output tensor with a dimension with size -1, so subsequent operations using the output fail to build.

[jxchenus]

I found a proper way to do it now: Basically the dynamic shape itself should be a model input.

[yjjinjie]

@jxchenus hello,I meet the same problem. how do you use the dynamic shape as a model_input?

pytorch/TensorRT#3140

when the batch_size used as. input, I must use batch_size= grouped_features["batch_size"].item() to get, it will encounted anther bug
queries = query.unsqueeze(1).expand(batch_size, max_seq_length, -1)

[jxchenus]

I found a workaround by using slice fill mode instead. The shape of the output can be dynamically passed in:

NVIDIA/TensorRT#3979 (comment)