[Feature] Add callback for Text Generation Pipelines

[rahul-tuli]

This PR adds a callback attribute to TextGenerationInput

callback is a callable that will be invoked on each generated token, and returns a boolean value.

If callback returns false, further generation of tokens is stopped

Test Code:

from deepsparse import Pipeline

start = 0
def dummy_callback(token):
    global start
    start += 1
    return start < 3


if __name__ == "__main__":
    codegen_pipeline = Pipeline.create(
        task="text_generation",
        model_path = "zoo:nlg/text_generation/codegen_mono-350m/pytorch/huggingface/bigpython_bigquery_thepile/base-none"
    )
    
    inputs = {
        "sequences": ["def fib(a, b, accumulator=0)"]*3,
        "callback": dummy_callback,
        "return_logits": True,
    }
    
    outputs = codegen_pipeline(**inputs)
    assert ( shape:=outputs.logits.shape[1]) == 3, f"Expected 3 tokens, got {shape}"

Also added above logic in an automated test (This is not yet active on GHA, but tested locally)

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1205276886236977

[dbogunowicz]

LGTM, but still not sure how useful this function is. Was this a request from product/QA?