[BUG] `pipeline.log` cache location not consistent within the same `Pipeline`

[alvarobartt]

Describe the bug

Apparently, the cache location is different in the Pipeline.run method before and after calling the super().run, since the signature is updated, and it modifies the path, so that the pipeline.log is dumped in one directory (signature calculated before super().run) and then the rest of the files into another one (signature calculated after super().run).

To Reproduce

import time

from distilabel.llms import LlamaCppLLM, OpenAILLM
from distilabel.pipeline import Pipeline
from distilabel.steps import KeepColumns, LoadHubDataset, PushToHub
from distilabel.steps.combine import CombineColumns
from distilabel.steps.generators.data import LoadDataFromDicts
from distilabel.steps.tasks import TextGeneration, UltraFeedback

if __name__ == "__main__":
    start_time = time.time()

    with Pipeline(name="ultrafeedback-dpo") as pipeline:
        load_dataset = LoadHubDataset(
            name="load_dataset",
            output_mappings={"prompt": "instruction"},
        )

        text_generation_zephyr = TextGeneration(
            name="text_generation_zephyr",
            llm=LlamaCppLLM(
                model_path="./models/zephyr-7b-beta.Q4_K_M.gguf",  # type: ignore
                n_gpu_layers=-1,
                n_ctx=1024,
            ),
            input_batch_size=10,
            output_mappings={"model_name": "generation_model"},
        )
        load_dataset.connect(text_generation_zephyr)

        text_generation_gemma = TextGeneration(
            name="text_generation_gemma",
            llm=LlamaCppLLM(
                model_path="./models/gemma-7b-it-Q4_K_M.gguf",  # type: ignore
                n_gpu_layers=-1,
                n_ctx=1024,
            ),
            input_batch_size=10,
            output_mappings={"model_name": "generation_model"},
        )
        load_dataset.connect(text_generation_gemma)

        combine_columns = CombineColumns(
            name="combine_columns",
            columns=["generation", "generation_model"],
            output_columns=["generations", "generation_models"],
        )
        text_generation_zephyr.connect(combine_columns)
        text_generation_gemma.connect(combine_columns)

        ultrafeedback = UltraFeedback(
            name="ultrafeedback",
            llm=OpenAILLM(
                model="gpt-4",
                api_key="sk-...",  # type: ignore
            ),
            aspect="overall-rating",
            input_batch_size=10,
            output_mappings={"model_name": "ultrafeedback_model"},
        )
        combine_columns.connect(ultrafeedback)

    distiset = pipeline.run(
        parameters={
            "load_dataset": {
                "repo_id": "distilabel-internal-testing/instruction-dataset-mini",
                "split": "test",
            },
            "text_generation_zephyr": {
                "llm": {
                    "generation_kwargs": {
                        "max_new_tokens": 1024,
                        "temperature": 0.7,
                    },
                },
            },
            "text_generation_gemma": {
                "llm": {
                    "generation_kwargs": {
                        "max_new_tokens": 1024,
                        "temperature": 0.7,
                    },
                },
            },
            "ultrafeedback": {
                "llm": {
                    "generation_kwargs": {
                        "max_new_tokens": 1024,
                        "temperature": 0.7,
                    },
                },
            },
        }
    )
    print("--- %s seconds ---" % (time.time() - start_time))

    distiset.push_to_hub(
        "alvarobartt/example-distilabel", token="..."
    )

Expected behaviour

The cache location to be the same within the Pipeline.run method.

Desktop:

• Package version: develop
• Python version: 3.11