Add text generation

[jimypbr]

Implement text generation for decoder and encoder/decoder models: GPT2, BART, T5

Adds:

• Greedy search
• Sampling
• Beam search
• Beam sample
• Pipelines for text2text, translation, and summarization

Example of using generate with GPT2:

from optimum.graphcore.modeling_utils import to_pipelined
from optimum.graphcore import IPUConfig
from transformers import GPT2TokenizerFast, GPT2LMHeadModel

tokenizer = GPT2TokenizerFast.from_pretrained("gpt2-medium")
model = GPT2LMHeadModel.from_pretrained("gpt2-medium", pad_token_id=tokenizer.eos_token_id)
input_ids = tokenizer.encode('I enjoy walking with my cute dog', return_tensors="pt")

ipu_config = IPUConfig(layers_per_ipu=[8, 16])
pipelined_model = to_pipelined(model, ipu_config)
pipelined_model = pipelined_model.half().parallelize()

sample_output = pipelined_model.generate(
    input_ids, 
    do_sample=True, 
    max_length=200, 
    top_k=0,
    top_p=0.92,
    use_cache=False,
)

print("Output:\n" + 100 * '-')
print(tokenizer.decode(sample_output[0], skip_special_tokens=True))

Pipelines are also supported:

from optimum.graphcore import pipeline

ipu_config = IPUConfig(layers_per_ipu=[16], ipus_per_replica=1)
text2text_generator = opipeline("text2text-generation", model="google/flan-t5-small", input_max_length=45, ipu_config=ipu_config.to_dict())

text2text_generator("question: What is 42 ? context: 42 is the answer to life, the universe and everything")
text2text_generator("Translate to German:  My name is Arthur.")

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[katalinic-gc]

LGTM