[BugFix] Delay torch import until needed for deepsparse.transformers.eval_downstream

[rahul-tuli]

Bug Description

A dependency issue where torch isn't installed with deepsparse[transformers]

#!/usr/bin/env bash

# Exit on error, undefined variables, and errors in piped commands
set -euf -o pipefail

# This script is used to reproduce the bug in the following issue and test the fix:
#  https://app.asana.com/0/1205229323407165/1205234254788114/f

# The bug is that torch is not installed as a part of deepsparse[transformers]
#  but is needed for `deepsparse.transformers.eval_downstream`
# for metrics calculation.


# Delete existing virtual environment
rm -rf venv

# Create a new virtual environment
python3 -m venv venv

# Activate the virtual environment
source venv/bin/activate

# Upgrade pip
pip install --upgrade pip

# Install deepsparse[transformers]
pip install -e "./[transformers]"

# Invoke problem command
deepsparse.transformers.eval_downstream --help

Before This PR, above script fails with the following error:

None of PyTorch, TensorFlow >= 2.0, or Flax have been found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
Traceback (most recent call last):
  File "/home/rahul/projects/deepsparse/venv/bin/deepsparse.transformers.eval_downstream", line 5, in <module>
    from deepsparse.transformers.eval_downstream import main
  File "/home/rahul/projects/deepsparse/src/deepsparse/transformers/eval_downstream.py", line 72, in <module>
    from deepsparse.transformers.metrics import Perplexity, PrecisionRecallF1
  File "/home/rahul/projects/deepsparse/src/deepsparse/transformers/metrics.py", line 26, in <module>
    import torch
ModuleNotFoundError: No module named 'torch'

After this PR, above script succeeds with the following output:

None of PyTorch, TensorFlow >= 2.0, or Flax have been found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
usage: deepsparse.transformers.eval_downstream [-h] -d {squad,squad_v2,mnli,qqp,sst2,sst2_zero_shot,imdb,conll2003,go_emotions,openai_humaneval} [-v VAL_RATIO] [-s VAL_SPLIT_SEED] [-c NUM_CORES] [-e {deepsparse,onnxruntime}]
                                               [--max-sequence-length MAX_SEQUENCE_LENGTH] [--max-samples MAX_SAMPLES] [-o OUTPUT_DIR] [--profile] [--max-answer-length MAX_ANSWER_LENGTH]
                                               [--version-2-with-negative VERSION_2_WITH_NEGATIVE] [--pad-to-max-length PAD_TO_MAX_LENGTH] [--n-best-size N_BEST_SIZE] [--zero-shot ZERO_SHOT]
                                               model_path

Evaluate a Hugging Face Transformers ONNX model on a downstream dataset

positional arguments:
  model_path            The path to a directory containing model.onnx, config.json, and tokenizer.json files or SparseZoo stub to the model

optional arguments:
  -h, --help            show this help message and exit
  -d {squad,squad_v2,mnli,qqp,sst2,sst2_zero_shot,imdb,conll2003,go_emotions,openai_humaneval}, --dataset {squad,squad_v2,mnli,qqp,sst2,sst2_zero_shot,imdb,conll2003,go_emotions,openai_humaneval}
  -v VAL_RATIO, --val_ratio VAL_RATIO
                        Ratio between 0.0 and 1.0 representing the proportion of the dataset include in the validation set
  -s VAL_SPLIT_SEED, --val_split_seed VAL_SPLIT_SEED
                        Random seed used to split the validation set, used with the --val_ratio flag. Default to 42.
  -c NUM_CORES, --num-cores NUM_CORES
                        The number of physical cores to run the eval on, defaults to all physical cores available on the system
  -e {deepsparse,onnxruntime}, --engine {deepsparse,onnxruntime}
                        Inference engine backend to run eval on. Choices are 'deepsparse', 'onnxruntime'. Default is 'deepsparse'
  --max-sequence-length MAX_SEQUENCE_LENGTH
                        the max sequence length for model inputs. Default is 384
  --max-samples MAX_SAMPLES
                        the max number of samples to evaluate. Default is None or all samples
  -o OUTPUT_DIR, --output-dir OUTPUT_DIR
                        Folder to save output predictions, used for debugging
  --profile             Run with profiling, used for debugging
  --max-answer-length MAX_ANSWER_LENGTH
                        The maximum length of an answer that can be generated. This is needed because the start and end predictions are not conditioned on one another.
  --version-2-with-negative VERSION_2_WITH_NEGATIVE
                        Whether or not the underlying dataset contains examples with no answers
  --pad-to-max-length PAD_TO_MAX_LENGTH
                        Whether to pad all samples to `max_seq_length`. If False, will pad the samples dynamically when batching to the maximum length in the batch (which can be faster on GPU but will be slower on TPU).
  --n-best-size N_BEST_SIZE
                        The total number of n-best predictions to generate when looking for an answer.
  --zero-shot ZERO_SHOT
                        Whether to run the dataset with a zero shot pipeline. Currently supports sst2. Default is False

[rahul-tuli]

Failing test should be green after: neuralmagic/sparsezoo#357