Skip to content

armgit / dspy

Public Imported from GitHub
0

DSPy: The framework for programming—not prompting—language models

Python

Standardize coding style in dspy/clients (#8163)

Chen Qian348c45c63738
diff --git a/dspy/clients/__init__.py b/dspy/clients/__init__.py
index 3eeba1e2c..52be745ab 100644
--- a/dspy/clients/__init__.py
+++ b/dspy/clients/__init__.py
@@ -1,15 +1,17 @@
-from dspy.clients.lm import LM
-from dspy.clients.provider import Provider, TrainingJob
-from dspy.clients.base_lm import BaseLM, inspect_history
-from dspy.clients.embedding import Embedder
-import litellm
+import logging
import os
from pathlib import Path
-from dspy.clients.cache import Cache
-import logging
from typing import Optional
+
+import litellm
from litellm.caching.caching import Cache as LitellmCache
+from dspy.clients.base_lm import BaseLM, inspect_history
+from dspy.clients.cache import Cache
+from dspy.clients.embedding import Embedder
+from dspy.clients.lm import LM
+from dspy.clients.provider import Provider, TrainingJob
+
logger = logging.getLogger(__name__)
DISK_CACHE_DIR = os.environ.get("DSPY_CACHEDIR") or os.path.join(Path.home(), ".dspy_cache")
diff --git a/dspy/clients/base_lm.py b/dspy/clients/base_lm.py
index 4c7d15cc3..9080d4cbe 100644
--- a/dspy/clients/base_lm.py
+++ b/dspy/clients/base_lm.py
@@ -1,6 +1,5 @@
import datetime
import uuid
-from abc import ABC
from dspy.dsp.utils import settings
from dspy.utils.callback import with_callbacks
@@ -9,7 +8,7 @@ MAX_HISTORY_SIZE = 10_000
GLOBAL_HISTORY = []
-class BaseLM(ABC):
+class BaseLM:
"""Base class for handling LLM calls.
Most users can directly use the `dspy.LM` class, which is a subclass of `BaseLM`. Users can also implement their
@@ -178,7 +177,10 @@ def _inspect_history(history, n: int = 1):
image_str = ""
if "base64" in c["image_url"].get("url", ""):
len_base64 = len(c["image_url"]["url"].split("base64,")[1])
- image_str = f"<{c['image_url']['url'].split('base64,')[0]}base64,<IMAGE BASE 64 ENCODED({str(len_base64)})>"
+ image_str = (
+ f"<{c['image_url']['url'].split('base64,')[0]}base64,"
+ f"<IMAGE BASE 64 ENCODED({len_base64!s})>"
+ )
else:
image_str = f"<image_url: {c['image_url']['url']}>"
print(_blue(image_str.strip()))
diff --git a/dspy/clients/cache.py b/dspy/clients/cache.py
index 9dff88136..d3bad5084 100644
--- a/dspy/clients/cache.py
+++ b/dspy/clients/cache.py
@@ -83,7 +83,7 @@ class Cache:
try:
# For regular functions, we can get the source code
return f"<callable_source:{inspect.getsource(value)}>"
- except (TypeError, OSError, IOError):
+ except (TypeError, OSError):
# For lambda functions or other callables where source isn't available,
# use a string representation
return f"<callable:{value.__name__ if hasattr(value, '__name__') else 'lambda'}>"
@@ -170,7 +170,7 @@ class Cache:
def request_cache(
cache_arg_name: Optional[str] = None,
- ignored_args_for_cache_key: Optional[list[str]] = ["api_key", "api_base", "base_url"],
+ ignored_args_for_cache_key: Optional[list[str]] = None,
enable_memory_cache: bool = True,
*, # everything after this is keyword-only
maxsize: Optional[int] = None, # legacy / no-op
@@ -185,7 +185,7 @@ def request_cache(
enable_memory_cache: Whether to enable in-memory cache at call time. If False, the memory cache will not be
written to on new data.
"""
-
+ ignored_args_for_cache_key = ignored_args_for_cache_key or ["api_key", "api_base", "base_url"]
# Deprecation notice
if maxsize is not None:
logger.warning(
diff --git a/dspy/clients/lm_local.py b/dspy/clients/lm_local.py
index 8329db978..79a0d8e02 100644
--- a/dspy/clients/lm_local.py
+++ b/dspy/clients/lm_local.py
@@ -49,12 +49,10 @@ class LocalProvider(Provider):
if model.startswith("local:"):
model = model[6:]
if model.startswith("huggingface/"):
- model = model[len("huggingface/"):]
+ model = model[len("huggingface/") :]
logger.info(f"Grabbing a free port to launch an SGLang server for model {model}")
- logger.info(
- f"We see that CUDA_VISIBLE_DEVICES is {os.environ.get('CUDA_VISIBLE_DEVICES', 'unset')}"
- )
+ logger.info(f"We see that CUDA_VISIBLE_DEVICES is {os.environ.get('CUDA_VISIBLE_DEVICES', 'unset')}")
port = get_free_port()
timeout = launch_kwargs.get("timeout", 1800)
command = f"python -m sglang.launch_server --model-path {model} --port {port} --host 0.0.0.0"
@@ -111,9 +109,7 @@ class LocalProvider(Provider):
return "".join(logs_buffer)
# Let the user know server is up
- logger.info(
- f"Server ready on random port {port}! Logs are available via lm.get_logs() method on returned lm."
- )
+ logger.info(f"Server ready on random port {port}! Logs are available via lm.get_logs() method on returned lm.")
lm.kwargs["api_base"] = f"http://localhost:{port}/v1"
lm.kwargs["api_key"] = "local"
@@ -121,10 +117,10 @@ class LocalProvider(Provider):
lm.process = process
lm.thread = thread
-
@staticmethod
def kill(lm: "LM", launch_kwargs: Optional[Dict[str, Any]] = None):
from sglang.utils import terminate_process
+
if not hasattr(lm, "process"):
logger.info("No running server to kill.")
return
@@ -168,7 +164,7 @@ class LocalProvider(Provider):
"bf16": True,
"output_dir": output_dir,
}
- train_kwargs={**default_train_kwargs, **(train_kwargs or {})}
+ train_kwargs = {**default_train_kwargs, **(train_kwargs or {})}
output_dir = train_kwargs["output_dir"] # user might have changed the output_dir
logger.info(f"Starting local training, will save to {output_dir}")
@@ -185,7 +181,7 @@ class LocalProvider(Provider):
def create_output_dir(model_name, data_path):
model_str = model_name.replace("/", "-")
time_str = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
- rnd_str = ''.join(random.choices(string.ascii_lowercase + string.digits, k=6))
+ rnd_str = "".join(random.choices(string.ascii_lowercase + string.digits, k=6))
model_identifier = f"{rnd_str}_{model_str}_{time_str}"
output_dir = data_path.replace(".jsonl", "_" + model_identifier)
return output_dir
@@ -204,16 +200,10 @@ def train_sft_locally(model_name, train_data, train_kwargs):
device = train_kwargs.get("device", None)
if device is None:
- device = (
- "cuda"
- if torch.cuda.is_available()
- else "mps" if torch.backends.mps.is_available() else "cpu"
- )
+ device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
logger.info(f"Using device: {device}")
- model = AutoModelForCausalLM.from_pretrained(
- pretrained_model_name_or_path=model_name
- ).to(device)
+ model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path=model_name).to(device)
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name)
# Set up the chat format; generally only for non-chat model variants, hence the try-except.
@@ -229,21 +219,25 @@ def train_sft_locally(model_name, train_data, train_kwargs):
logger.info("Creating dataset")
if "max_seq_length" not in train_kwargs:
train_kwargs["max_seq_length"] = 4096
- logger.info(f"The 'train_kwargs' parameter didn't include a 'max_seq_length', defaulting to {train_kwargs['max_seq_length']}")
+ logger.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"])
+ return encode_sft_example(example, tokenizer, train_kwargs["max_seq_length"]) # noqa: F821
+
tokenized_dataset = hf_dataset.map(tokenize_function, batched=False)
tokenized_dataset.set_format(type="torch")
tokenized_dataset = tokenized_dataset.filter(lambda example: (example["labels"] != -100).any())
- USE_PEFT = train_kwargs.get("use_peft", False)
+ use_peft = train_kwargs.get("use_peft", False)
peft_config = None
- if USE_PEFT:
+ if use_peft:
from peft import LoraConfig
rank_dimension = 32
@@ -293,8 +287,8 @@ def train_sft_locally(model_name, train_data, train_kwargs):
# Save the model!
trainer.save_model()
- MERGE = True
- if USE_PEFT and MERGE:
+ merge = True
+ if use_peft and merge:
from peft import AutoPeftModelForCausalLM
# Load PEFT model on CPU
@@ -305,9 +299,7 @@ def train_sft_locally(model_name, train_data, train_kwargs):
)
merged_model = model_.merge_and_unload()
- merged_model.save_pretrained(
- sft_config.output_dir, safe_serialization=True, max_shard_size="5GB"
- )
+ merged_model.save_pretrained(sft_config.output_dir, safe_serialization=True, max_shard_size="5GB")
# Clean up!
import gc
@@ -330,7 +322,7 @@ def get_free_port() -> int:
return s.getsockname()[1]
-def wait_for_server(base_url: str, timeout: int = None) -> None:
+def wait_for_server(base_url: str, timeout: Optional[int] = None) -> None:
"""
Wait for the server to be ready by polling the /v1/models endpoint.
diff --git a/dspy/clients/openai.py b/dspy/clients/openai.py
index 6cfaf1a1c..1f5952ceb 100644
--- a/dspy/clients/openai.py
+++ b/dspy/clients/openai.py
@@ -5,50 +5,49 @@ from typing import Any, Dict, List, Optional
import openai
-from dspy.clients.provider import TrainingJob, Provider
+from dspy.clients.provider import Provider, TrainingJob
from dspy.clients.utils_finetune import TrainDataFormat, TrainingStatus, save_data
-
_OPENAI_MODELS = [
- 'gpt-4-turbo',
- 'gpt-4-turbo-2024-04-09',
- 'tts-1',
- 'tts-1-1106',
- 'chatgpt-4o-latest',
- 'dall-e-2',
- 'whisper-1',
- 'gpt-3.5-turbo-instruct',
- 'gpt-3.5-turbo',
- 'gpt-3.5-turbo-0125',
- 'babbage-002',
- 'davinci-002',
- 'gpt-4o-mini-2024-07-18',
- 'gpt-4o',
- 'dall-e-3',
- 'gpt-4o-mini',
- 'gpt-4o-2024-08-06',
- 'gpt-4o-2024-05-13',
- 'o1-preview',
- 'gpt-4o-audio-preview-2024-10-01',
- 'o1-mini-2024-09-12',
- 'gpt-4o-audio-preview',
- 'tts-1-hd',
- 'tts-1-hd-1106',
- 'o1-preview-2024-09-12',
- 'o1-mini',
- 'gpt-4-1106-preview',
- 'text-embedding-ada-002',
- 'gpt-3.5-turbo-16k',
- 'text-embedding-3-small',
- 'text-embedding-3-large',
- 'gpt-4o-realtime-preview-2024-10-01',
- 'gpt-4o-realtime-preview',
- 'gpt-3.5-turbo-1106',
- 'gpt-4-0613',
- 'gpt-4-turbo-preview',
- 'gpt-4-0125-preview',
- 'gpt-4',
- 'gpt-3.5-turbo-instruct-0914'
+ "gpt-4-turbo",
+ "gpt-4-turbo-2024-04-09",
+ "tts-1",
+ "tts-1-1106",
+ "chatgpt-4o-latest",
+ "dall-e-2",
+ "whisper-1",
+ "gpt-3.5-turbo-instruct",
+ "gpt-3.5-turbo",
+ "gpt-3.5-turbo-0125",
+ "babbage-002",
+ "davinci-002",
+ "gpt-4o-mini-2024-07-18",
+ "gpt-4o",
+ "dall-e-3",
+ "gpt-4o-mini",
+ "gpt-4o-2024-08-06",
+ "gpt-4o-2024-05-13",
+ "o1-preview",
+ "gpt-4o-audio-preview-2024-10-01",
+ "o1-mini-2024-09-12",
+ "gpt-4o-audio-preview",
+ "tts-1-hd",
+ "tts-1-hd-1106",
+ "o1-preview-2024-09-12",
+ "o1-mini",
+ "gpt-4-1106-preview",
+ "text-embedding-ada-002",
+ "gpt-3.5-turbo-16k",
+ "text-embedding-3-small",
+ "text-embedding-3-large",
+ "gpt-4o-realtime-preview-2024-10-01",
+ "gpt-4o-realtime-preview",
+ "gpt-3.5-turbo-1106",
+ "gpt-4-0613",
+ "gpt-4-turbo-preview",
+ "gpt-4-0125-preview",
+ "gpt-4",
+ "gpt-3.5-turbo-instruct-0914",
]
@@ -84,7 +83,6 @@ class TrainingJobOpenAI(TrainingJob):
class OpenAIProvider(Provider):
-
def __init__(self):
super().__init__()
self.finetunable = True
@@ -113,7 +111,7 @@ class OpenAIProvider(Provider):
return True
return False
-
+
@staticmethod
def _remove_provider_prefix(model: str) -> str:
provider_prefix = "openai/"
@@ -179,7 +177,6 @@ class OpenAIProvider(Provider):
except Exception:
return False
-
@staticmethod
def is_terminal_training_status(status: TrainingStatus) -> bool:
return status in [
@@ -187,7 +184,7 @@ class OpenAIProvider(Provider):
TrainingStatus.failed,
TrainingStatus.cancelled,
]
-
+
@staticmethod
def get_training_status(job_id: str) -> TrainingStatus:
provider_status_to_training_status = {
@@ -234,11 +231,7 @@ class OpenAIProvider(Provider):
return provider_file.id
@staticmethod
- def _start_remote_training(
- train_file_id: str,
- model: str,
- train_kwargs: Optional[Dict[str, Any]] = None
- ) -> str:
+ def _start_remote_training(train_file_id: str, model: str, train_kwargs: Optional[Dict[str, Any]] = None) -> str:
train_kwargs = train_kwargs or {}
provider_job = openai.fine_tuning.jobs.create(
model=model,
diff --git a/dspy/clients/provider.py b/dspy/clients/provider.py
index a77ba8202..b73f577c7 100644
--- a/dspy/clients/provider.py
+++ b/dspy/clients/provider.py
@@ -1,14 +1,14 @@
from abc import abstractmethod
from concurrent.futures import Future
from threading import Thread
-from typing import Any, Dict, List, Optional, Union
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
from dspy.clients.utils_finetune import TrainDataFormat
-from typing import TYPE_CHECKING
if TYPE_CHECKING:
from dspy.clients.lm import LM
+
class TrainingJob(Future):
def __init__(
self,
diff --git a/dspy/clients/utils_finetune.py b/dspy/clients/utils_finetune.py
index 6842b7ccf..7d83c741f 100644
--- a/dspy/clients/utils_finetune.py
+++ b/dspy/clients/utils_finetune.py
@@ -62,12 +62,12 @@ def save_data(
def validate_data_format(
- data: List[Dict[str, Any]],
- data_format: TrainDataFormat,
- ):
+ data: List[Dict[str, Any]],
+ data_format: TrainDataFormat,
+):
find_err_funcs = {
TrainDataFormat.CHAT: find_data_error_chat,
- TrainDataFormat.COMPLETION: find_data_errors_completion
+ TrainDataFormat.COMPLETION: find_data_errors_completion,
}
err = f"Data format {data_format} is not supported."
assert data_format in find_err_funcs, err
@@ -83,7 +83,7 @@ def validate_data_format(
if isinstance(data_dict, dict):
err = find_err_func(data_dict)
if err:
- err_dict = dict(index=ind, error=err)
+ err_dict = {"index": ind, "error": err}
data_dict_errors.append(err_dict)
if data_dict_errors:
@@ -96,9 +96,7 @@ def validate_data_format(
raise ValueError(err)
-def find_data_errors_completion(
- data_dict: Dict[str, str]
- ) -> Optional[str]:
+def find_data_errors_completion(data_dict: Dict[str, str]) -> Optional[str]:
keys = ["prompt", "completion"]
assert isinstance(data_dict, dict)
@@ -106,7 +104,7 @@ def find_data_errors_completion(
found_keys = sorted(data_dict.keys())
if set(expected_keys) != set(found_keys):
return f"Expected Keys: {expected_keys}; Found Keys: {found_keys}"
-
+
for key in keys:
if not isinstance(data_dict[key], str):
return f"Expected `{key}` to be of type `str`. Found: {type(data_dict[key])}"
@@ -114,9 +112,7 @@ def find_data_errors_completion(
# Following functions are modified from the OpenAI cookbook:
# https://cookbook.openai.com/examples/chat_finetuning_data_prep
-def find_data_error_chat(
- messages: Dict[str, Any]
- ) -> Optional[str]:
+def find_data_error_chat(messages: Dict[str, Any]) -> Optional[str]:
assert isinstance(messages, dict)
expected_keys = ["messages"]
@@ -133,9 +129,7 @@ def find_data_error_chat(
return f"Error in message at index {ind}: {err}"
-def find_data_error_chat_message(
- message: Dict[str, Any]
- ) -> Optional[str]:
+def find_data_error_chat_message(message: Dict[str, Any]) -> Optional[str]:
assert isinstance(message, dict)
message_keys = sorted(["role", "content"])