DSPy: The framework for programming—not prompting—language models
Python
Decrease the memory usage at `import dspy` (#7915)
Chen Qianc0b79b07dc32
diff --git a/dspy/clients/lm_local.py b/dspy/clients/lm_local.pyindex 3610271eb..6094e4c91 100644--- a/dspy/clients/lm_local.py+++ b/dspy/clients/lm_local.py@@ -1,18 +1,17 @@import datetimeimport loggingimport random-import requestsimport socketimport stringimport subprocess-import timeimport threading+import time+from typing import TYPE_CHECKING, Any, Dict, List, Optional-from datasets import Dataset-from typing import Any, Dict, List, Optional-from dspy.clients.provider import TrainingJob, Provider+import requests++from dspy.clients.provider import Provider, TrainingJobfrom dspy.clients.utils_finetune import TrainDataFormat, save_data-from typing import TYPE_CHECKINGif TYPE_CHECKING:from dspy.clients.lm import LM@@ -29,7 +28,7 @@ class LocalProvider(Provider):@staticmethoddef launch(lm: "LM", launch_kwargs: Optional[Dict[str, Any]] = None):try:- import sglang # noqa: F401+ import sglang # noqa: F401except ImportError:raise ImportError("For local model launching, please install sglang by running "@@ -229,6 +228,8 @@ def train_sft_locally(model_name, train_data, train_kwargs):train_kwargs["max_seq_length"] = 4096logger.info(f"The 'train_kwargs' parameter didn't include a 'max_seq_length', defaulting to {train_kwargs['max_seq_length']}")+ from datasets import Dataset+hf_dataset = Dataset.from_list(train_data)def tokenize_function(example):return encode_sft_example(example, tokenizer, train_kwargs["max_seq_length"])diff --git a/dspy/clients/utils_finetune.py b/dspy/clients/utils_finetune.pyindex b77320011..6842b7ccf 100644--- a/dspy/clients/utils_finetune.py+++ b/dspy/clients/utils_finetune.py@@ -3,7 +3,6 @@ from enum import Enumfrom typing import Any, Dict, List, Optionalimport ujson-from datasets.fingerprint import Hasherimport dspyfrom dspy.adapters.base import Adapter@@ -47,6 +46,8 @@ def write_lines(file_path, data):def save_data(data: List[Dict[str, Any]],) -> str:+ from datasets.fingerprint import Hasher+# Assign a unique name to the file based on the data hashhash = Hasher.hash(data)file_name = f"{hash}.jsonl"diff --git a/dspy/datasets/dataloader.py b/dspy/datasets/dataloader.pyindex 351b84418..a45d2c057 100644--- a/dspy/datasets/dataloader.py+++ b/dspy/datasets/dataloader.py@@ -1,13 +1,13 @@import randomfrom collections.abc import Mapping-from typing import List, Tuple, Union--import pandas as pd-from datasets import load_dataset+from typing import TYPE_CHECKING, List, Tuple, Unionimport dspyfrom dspy.datasets.dataset import Dataset+if TYPE_CHECKING:+ import pandas as pd+class DataLoader(Dataset):def __init__(self):@@ -27,6 +27,8 @@ class DataLoader(Dataset):if not isinstance(input_keys, tuple):raise ValueError("Invalid input keys provided. Please provide a tuple of input keys.")+ from datasets import load_dataset+dataset = load_dataset(dataset_name, *args, **kwargs)if isinstance(dataset, list) and isinstance(kwargs["split"], list):@@ -59,6 +61,8 @@ class DataLoader(Dataset):]def from_csv(self, file_path: str, fields: List[str] = None, input_keys: Tuple[str] = ()) -> List[dspy.Example]:+ from datasets import load_dataset+dataset = load_dataset("csv", data_files=file_path)["train"]if not fields:@@ -68,7 +72,7 @@ class DataLoader(Dataset):def from_pandas(self,- df: pd.DataFrame,+ df: "pd.DataFrame",fields: list[str] = None,input_keys: tuple[str] = (),) -> list[dspy.Example]:@@ -80,6 +84,8 @@ class DataLoader(Dataset):]def from_json(self, file_path: str, fields: List[str] = None, input_keys: Tuple[str] = ()) -> List[dspy.Example]:+ from datasets import load_dataset+dataset = load_dataset("json", data_files=file_path)["train"]if not fields:@@ -88,6 +94,8 @@ class DataLoader(Dataset):return [dspy.Example({field: row[field] for field in fields}).with_inputs(*input_keys) for row in dataset]def from_parquet(self, file_path: str, fields: List[str] = None, input_keys: Tuple[str] = ()) -> List[dspy.Example]:+ from datasets import load_dataset+dataset = load_dataset("parquet", data_files=file_path)["train"]if not fields:diff --git a/dspy/datasets/gsm8k.py b/dspy/datasets/gsm8k.pyindex ec1acf6a4..d2bb7ac08 100644--- a/dspy/datasets/gsm8k.py+++ b/dspy/datasets/gsm8k.py@@ -1,13 +1,14 @@import randomimport tqdm-from datasets import load_datasetclass GSM8K:def __init__(self):self.do_shuffle = False+ from datasets import load_dataset+dataset = load_dataset("gsm8k", "main")hf_official_train = dataset["train"]diff --git a/dspy/datasets/hotpotqa.py b/dspy/datasets/hotpotqa.pyindex e5cb9576a..995418388 100644--- a/dspy/datasets/hotpotqa.py+++ b/dspy/datasets/hotpotqa.py@@ -1,7 +1,5 @@import random-from datasets import load_dataset-from dspy.datasets.dataset import Dataset@@ -20,6 +18,8 @@ class HotPotQA(Dataset):"Dev must be all hard to match official dev, but training can be flexible.")+ from datasets import load_dataset+hf_official_train = load_dataset("hotpot_qa", "fullwiki", split="train", trust_remote_code=True)hf_official_dev = load_dataset("hotpot_qa", "fullwiki", split="validation", trust_remote_code=True)diff --git a/dspy/evaluate/evaluate.py b/dspy/evaluate/evaluate.pyindex 407789cc0..b9c709897 100644--- a/dspy/evaluate/evaluate.py+++ b/dspy/evaluate/evaluate.py@@ -1,8 +1,10 @@import loggingimport types-from typing import Any, Callable, List, Optional+from typing import TYPE_CHECKING, Any, Callable, List, Optional++if TYPE_CHECKING:+ import pandas as pd-import pandas as pdimport tqdmimport dspy@@ -183,6 +185,7 @@ class Evaluate:]+ import pandas as pd# Truncate every cell in the DataFrame (DataFrame.applymap was renamed to DataFrame.map in Pandas 2.1.0)result_df = pd.DataFrame(data)result_df = result_df.map(truncate_cell) if hasattr(result_df, "map") else result_df.applymap(truncate_cell)@@ -252,7 +255,7 @@ def truncate_cell(content) -> str:return content-def stylize_metric_name(df: pd.DataFrame, metric_name: str) -> pd.DataFrame:+def stylize_metric_name(df: "pd.DataFrame", metric_name: str) -> "pd.DataFrame":"""Stylize the cell contents of a pandas DataFrame corresponding to the specified metric name.@@ -265,12 +268,14 @@ def stylize_metric_name(df: pd.DataFrame, metric_name: str) -> pd.DataFrame:return df-def display_dataframe(df: pd.DataFrame):+def display_dataframe(df: "pd.DataFrame"):"""Display the specified Pandas DataFrame in the console.:param df: The Pandas DataFrame to display."""+ import pandas as pd+if is_in_ipython_notebook_environment():display(configure_dataframe_for_ipython_notebook_display(df))else:@@ -281,8 +286,10 @@ def display_dataframe(df: pd.DataFrame):print(df)-def configure_dataframe_for_ipython_notebook_display(df: pd.DataFrame) -> pd.DataFrame:+def configure_dataframe_for_ipython_notebook_display(df: "pd.DataFrame") -> "pd.DataFrame":"""Set various pandas display options for DataFrame in an IPython notebook environment."""+ import pandas as pd+pd.options.display.max_colwidth = 70return dfdiff --git a/dspy/experimental/synthesizer/synthesizer.py b/dspy/experimental/synthesizer/synthesizer.pyindex 9a2bfcd74..e1184c5e5 100644--- a/dspy/experimental/synthesizer/synthesizer.py+++ b/dspy/experimental/synthesizer/synthesizer.py@@ -2,7 +2,6 @@ import randomfrom collections.abc import Mappingfrom typing import List, Optional, Union-from datasets import Datasetfrom rich import print as rprintfrom tqdm import tqdm, trange@@ -241,6 +240,8 @@ class Synthesizer:def export(self, data: List[dspy.Example], path: str, mode: str = None, **kwargs):extension = mode or path.split(".")[-1]+ from datasets import Dataset+dataset = Dataset.from_list([example.toDict() for example in data],)