[Pipeline Refactor] Update routes, text generation initial functionality

[dsikka]

Summary

• The operators, pipeline, and routes added for text_generation so far include the operators and the flow shown above

Testing

Testing script to evaluate various situations:

1. Just running the single-token engine as prompt_sequence_length > number of prompt tokens
2. Just running the multi-token engine as the number of tokens % prompt_sequence_length == 0
3. Running a combination of multi-token engine and single-token engine, as the number of prompt_sequence_length <= number of prompt tokens but prompt tokens % prompt_sequence_length != 0

All 3 cases are tested using the test script below, by changing the prompt_sequence_length. Prompt logits are evaluated against the ground truth logits, produced using the transformers model.

import numpy as np
from transformers import AutoModelForCausalLM, AutoTokenizer

import torch
from deepsparse.transformers.pipelines.text_generation import TextGenerationInput
from deepsparse.v2.text_generation.pipeline import TextGenerationPipeline
from huggingface_hub import snapshot_download

def create_tokenizer(model_name):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    tokenizer.padding_side = "left"
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    return tokenizer

def get_ground_truth(prompt):
    model = AutoModelForCausalLM.from_pretrained("roneneldan/TinyStories-1M")
    tokenizer = create_tokenizer("roneneldan/TinyStories-1M")

    input_ids = tokenizer.encode(prompt, return_tensors="pt")
    out = model(input_ids=input_ids)
    prompt_logits = out.logits.detach().numpy()
    return prompt_logits

cur_len = 16
prompt = "Hello there, how are you?"
model_path = "hf:mgoin/TinyStories-1M-deepsparse"
pipeline = TextGenerationPipeline(model_path, engine_kwargs={"engine_type": "onnxruntime"}, prompt_sequence_length=cur_len)
input_values = TextGenerationInput(prompt=prompt)
logits = pipeline(input_values)
ground_truth = get_ground_truth(prompt)

print("All Close?", np.allclose(logits, ground_truth, atol=0.0001))