DSPy: The framework for programming—not prompting—language models
Python
Style fixing for leftovers (#8228)
Chen Qian816484737772
diff --git a/dspy/clients/lm.py b/dspy/clients/lm.pyindex ba48d9034..247db10b9 100644--- a/dspy/clients/lm.py+++ b/dspy/clients/lm.py@@ -11,7 +11,7 @@ from asyncer import syncifyimport dspyfrom dspy.clients.cache import request_cachefrom dspy.clients.openai import OpenAIProvider-from dspy.clients.provider import Provider, TrainingJob, ReinforceJob+from dspy.clients.provider import Provider, ReinforceJob, TrainingJobfrom dspy.clients.utils_finetune import TrainDataFormatfrom dspy.dsp.utils.settings import settingsfrom dspy.utils.callback import BaseCallback@@ -211,7 +211,7 @@ class LM(BaseLM):def reinforce(self, train_kwargs) -> ReinforceJob:# TODO(GRPO Team): Should we return an initialized job here?from dspy import settings as settings-+err = f"Provider {self.provider} does not implement the reinforcement learning interface."assert self.provider.reinforceable, errdiff --git a/dspy/clients/lm_local_arbor.py b/dspy/clients/lm_local_arbor.pyindex e7bf1debc..9427a19d6 100644--- a/dspy/clients/lm_local_arbor.py+++ b/dspy/clients/lm_local_arbor.py@@ -1,14 +1,13 @@import time-import requestsfrom datetime import datetime-from typing import Any, Dict, List, Optional, Union, TypedDict, TYPE_CHECKING+from typing import TYPE_CHECKING, Any, Dict, List, Optional, TypedDict, Unionimport openai+import requestsimport dspy-from dspy.clients.provider import TrainingJob, ReinforceJob, Provider-from dspy.clients.utils_finetune import TrainDataFormat, TrainingStatus, GRPOGroup, save_data-+from dspy.clients.provider import Provider, ReinforceJob, TrainingJob+from dspy.clients.utils_finetune import GRPOGroup, TrainDataFormat, TrainingStatus, save_dataif TYPE_CHECKING:from dspy.clients.lm import LM@@ -47,7 +46,7 @@ class ArborTrainingJob(TrainingJob):class ArborReinforceJob(ReinforceJob):- DEFAULT_TRAIN_KWARGS = {+ DEFAULT_TRAIN_KWARGS = { # noqa: RUF012"update_interval": 10,"temperature": 0.9,"beta": 0.04,@@ -56,7 +55,7 @@ class ArborReinforceJob(ReinforceJob):"learning_rate": 1e-6,"gradient_accumulation_steps": 1,# This is false by default in TRL, but I think it makes sense to be true for us- "gradient_checkpointing": True,+ "gradient_checkpointing": True,"lr_scheduler_type": "constant_with_warmup","max_prompt_length": None,"max_completion_length": None,@@ -69,7 +68,7 @@ class ArborReinforceJob(ReinforceJob):"logging_steps": 100,# By default, none is the model's max context length"max_context_length": None,- "lora": False+ "lora": False,}def __init__(self, lm: "LM", train_kwargs: GRPOTrainKwargs):@@ -90,75 +89,81 @@ class ArborReinforceJob(ReinforceJob):temperature = self.train_kwargs.get("temperature", self.DEFAULT_TRAIN_KWARGS["temperature"])beta = self.train_kwargs.get("beta", self.DEFAULT_TRAIN_KWARGS["beta"])num_iterations = self.train_kwargs.get("num_iterations", self.DEFAULT_TRAIN_KWARGS["num_iterations"])- per_device_train_batch_size = self.train_kwargs.get("per_device_train_batch_size", self.DEFAULT_TRAIN_KWARGS["per_device_train_batch_size"])+ per_device_train_batch_size = self.train_kwargs.get(+ "per_device_train_batch_size", self.DEFAULT_TRAIN_KWARGS["per_device_train_batch_size"]+ )learning_rate = self.train_kwargs.get("learning_rate", self.DEFAULT_TRAIN_KWARGS["learning_rate"])- gradient_accumulation_steps = self.train_kwargs.get("gradient_accumulation_steps", self.DEFAULT_TRAIN_KWARGS["gradient_accumulation_steps"])- gradient_checkpointing = self.train_kwargs.get("gradient_checkpointing", self.DEFAULT_TRAIN_KWARGS["gradient_checkpointing"])+ gradient_accumulation_steps = self.train_kwargs.get(+ "gradient_accumulation_steps", self.DEFAULT_TRAIN_KWARGS["gradient_accumulation_steps"]+ )+ gradient_checkpointing = self.train_kwargs.get(+ "gradient_checkpointing", self.DEFAULT_TRAIN_KWARGS["gradient_checkpointing"]+ )lr_scheduler_type = self.train_kwargs.get("lr_scheduler_type", self.DEFAULT_TRAIN_KWARGS["lr_scheduler_type"])max_prompt_length = self.train_kwargs.get("max_prompt_length", self.DEFAULT_TRAIN_KWARGS["max_prompt_length"])- max_completion_length = self.train_kwargs.get("max_completion_length", self.DEFAULT_TRAIN_KWARGS["max_completion_length"])+ max_completion_length = self.train_kwargs.get(+ "max_completion_length", self.DEFAULT_TRAIN_KWARGS["max_completion_length"]+ )bf16 = self.train_kwargs.get("bf16", self.DEFAULT_TRAIN_KWARGS["bf16"])scale_rewards = self.train_kwargs.get("scale_rewards", self.DEFAULT_TRAIN_KWARGS["scale_rewards"])- gradient_checkpointing_kwargs = self.train_kwargs.get("gradient_checkpointing_kwargs", self.DEFAULT_TRAIN_KWARGS["gradient_checkpointing_kwargs"])+ gradient_checkpointing_kwargs = self.train_kwargs.get(+ "gradient_checkpointing_kwargs", self.DEFAULT_TRAIN_KWARGS["gradient_checkpointing_kwargs"]+ )max_grad_norm = self.train_kwargs.get("max_grad_norm", self.DEFAULT_TRAIN_KWARGS["max_grad_norm"])report_to = self.train_kwargs.get("report_to", self.DEFAULT_TRAIN_KWARGS["report_to"])log_completions = self.train_kwargs.get("log_completions", self.DEFAULT_TRAIN_KWARGS["log_completions"])logging_steps = self.train_kwargs.get("logging_steps", self.DEFAULT_TRAIN_KWARGS["logging_steps"])- max_context_length = self.train_kwargs.get("max_context_length", self.DEFAULT_TRAIN_KWARGS["max_context_length"])+ max_context_length = self.train_kwargs.get(+ "max_context_length", self.DEFAULT_TRAIN_KWARGS["max_context_length"]+ )lora = self.train_kwargs.get("lora", self.DEFAULT_TRAIN_KWARGS["lora"])api_base = self.lm.kwargs["api_base"]suffix = "dspy"finetune_model = ArborProvider._remove_provider_prefix(self.lm.model)data = {- 'model': finetune_model,- 'suffix': suffix,- 'num_generations': num_generations,- 'update_interval': update_interval,- 'temperature': temperature,- 'beta': beta,- 'num_iterations': num_iterations,- 'per_device_train_batch_size': per_device_train_batch_size,- 'learning_rate': learning_rate,- 'gradient_accumulation_steps': gradient_accumulation_steps,- 'gradient_checkpointing': gradient_checkpointing,- 'lr_scheduler_type': lr_scheduler_type,- 'max_prompt_length': max_prompt_length,- 'max_completion_length': max_completion_length,- 'bf16': bf16,- 'scale_rewards': scale_rewards,- 'gradient_checkpointing_kwargs': gradient_checkpointing_kwargs,- 'max_grad_norm': max_grad_norm,- 'report_to': report_to,- 'log_completions': log_completions,- 'logging_steps': logging_steps,- 'max_context_length': max_context_length,- 'lora': lora,+ "model": finetune_model,+ "suffix": suffix,+ "num_generations": num_generations,+ "update_interval": update_interval,+ "temperature": temperature,+ "beta": beta,+ "num_iterations": num_iterations,+ "per_device_train_batch_size": per_device_train_batch_size,+ "learning_rate": learning_rate,+ "gradient_accumulation_steps": gradient_accumulation_steps,+ "gradient_checkpointing": gradient_checkpointing,+ "lr_scheduler_type": lr_scheduler_type,+ "max_prompt_length": max_prompt_length,+ "max_completion_length": max_completion_length,+ "bf16": bf16,+ "scale_rewards": scale_rewards,+ "gradient_checkpointing_kwargs": gradient_checkpointing_kwargs,+ "max_grad_norm": max_grad_norm,+ "report_to": report_to,+ "log_completions": log_completions,+ "logging_steps": logging_steps,+ "max_context_length": max_context_length,+ "lora": lora,}url = f"{api_base}fine_tuning/grpo/initialize"- headers = {'Content-Type': 'application/json'}- response = requests.post(- url=url,- headers=headers,- json=data- )+ headers = {"Content-Type": "application/json"}+ response = requests.post(url=url, headers=headers, json=data)assert response.status_code == 200, f"Failed to initialize GRPO: {response}"self.lm.model = ArborProvider._add_provider_prefix(finetune_model)# self.provider_job_id = response.json().get("job_id") # TODO: To be updated- def _run_grpo_step_one_group(self, train_group: GRPOGroup, train_data_format: Optional[Union[TrainDataFormat, str]] = None):+ def _run_grpo_step_one_group(+ self, train_group: GRPOGroup, train_data_format: Optional[Union[TrainDataFormat, str]] = None+ ):# TODO: Check that the data follows the intended formatapi_base = self.lm.kwargs["api_base"]# api_key = self.lm.kwargs["api_key"]finetune_model = ArborProvider._remove_provider_prefix(self.lm.model)- data = {- 'model': finetune_model,- 'update_inference_model': True,- 'batch': train_group- }+ data = {"model": finetune_model, "update_inference_model": True, "batch": train_group}url = f"{api_base}fine_tuning/grpo/step"- headers = {'Content-Type': 'application/json'}+ headers = {"Content-Type": "application/json"}response = requests.post(url, headers=headers, json=data)assert response.status_code == 200, f"Failed to run a GRPO step: {response.text}"response = response.json()@@ -175,38 +180,39 @@ class ArborReinforceJob(ReinforceJob):# different step methods or changing our smallets data format to be the# GRPO group.# TODO: Support step on the server side- assert train_data_format == TrainDataFormat.GRPO_CHAT, f"GRPO only supports the GRPO_CHAT data format. Got {train_data_format} instead."+ assert (+ train_data_format == TrainDataFormat.GRPO_CHAT+ ), f"GRPO only supports the GRPO_CHAT data format. Got {train_data_format} instead."for group in train_data:self._run_grpo_step_one_group(group, train_data_format)-+def update_model(self):api_base = self.lm.kwargs["api_base"]url = f"{api_base}fine_tuning/grpo/update_model"- headers = {'Content-Type': 'application/json'}+ headers = {"Content-Type": "application/json"}response = requests.post(url, headers=headers)- assert response.status_code == 200, f"Failed to update model: {response.text}"+ assert response.status_code == 200, f"Failed to update model: {response.text}"response = response.json()current_model = response["current_model"]self.lm.model = ArborProvider._add_provider_prefix(current_model)-- def save_checkpoint(self, checkpoint_name: str, score: float = None):++ def save_checkpoint(self, checkpoint_name: str, score: Optional[float] = None):api_base = self.lm.kwargs["api_base"]url = f"{api_base}fine_tuning/grpo/checkpoint"- headers = {'Content-Type': 'application/json'}+ headers = {"Content-Type": "application/json"}body = {"checkpoint_name": checkpoint_name}response = requests.post(url, headers=headers, json=body)assert response.status_code == 200, f"Failed to save checkpoint: {response.text}"response = response.json()- current_model = response["current_model"]last_checkpoint = response["last_checkpoint"]checkpoints = response["checkpoints"]checkpoint_model_path = checkpoints[last_checkpoint]self.checkpoints[last_checkpoint] = {- 'model_path':checkpoint_model_path,- 'score': score,+ "model_path": checkpoint_model_path,+ "score": score,}self.last_checkpoint = last_checkpoint@@ -214,7 +220,7 @@ class ArborReinforceJob(ReinforceJob):api_base = self.lm.kwargs["api_base"]url = f"{api_base}fine_tuning/grpo/terminate"- headers = {'Content-Type': 'application/json'}+ headers = {"Content-Type": "application/json"}response = requests.post(url, headers=headers)assert response.status_code == 200, f"Failed to terminate GRPO: {response.text}"@@ -238,7 +244,6 @@ class ArborReinforceJob(ReinforceJob):class ArborProvider(Provider):-def __init__(self):super().__init__()self.finetunable = True@@ -251,29 +256,23 @@ class ArborProvider(Provider):model = ArborProvider._remove_provider_prefix(lm.model)# TODO: Handle this on the server sideif model.startswith("huggingface/"):- model = model[len("huggingface/"):]+ model = model[len("huggingface/") :]api_base = lm.kwargs["api_base"]- api_key = lm.kwargs["api_key"]launch_kwargs = launch_kwargs or lm.launch_kwargs# Make request to launch endpoint- response = requests.post(- f"{api_base}chat/launch",- json={"model": model, "launch_kwargs": launch_kwargs}- )+ response = requests.post(f"{api_base}chat/launch", json={"model": model, "launch_kwargs": launch_kwargs})if response.status_code != 200:raise Exception(f"Failed to launch model. Status code: {response.status_code}, Response: {response.text}")print(f"Inference server for model {model} launched successfully")-@staticmethoddef kill(lm: "LM", launch_kwargs: Optional[Dict[str, Any]] = None):api_base = lm.kwargs["api_base"]- api_key = lm.kwargs["api_key"]response = requests.post(f"{api_base}chat/kill",@@ -291,7 +290,7 @@ class ArborProvider(Provider):if model.startswith("arbor:"):model = model[6:]return model-+@staticmethoddef _add_provider_prefix(model: str) -> str:if not model.startswith("openai/arbor:"):@@ -303,17 +302,20 @@ class ArborProvider(Provider):# TODO: We will delete this method once we start passing the LM object# to finetune.import dspy.settings as settings+if not hasattr(settings, "arbor_api_base"):- raise ValueError("Arbor API base not set. Please set the `dspy.settings.arbor_api_base` to the URL for the Arbor server (e.g. 'http://localhost:8000/v1/').")+ raise ValueError(+ "Arbor API base not set. Please set the `dspy.settings.arbor_api_base` to the URL for the Arbor server (e.g. 'http://localhost:8000/v1/')."+ )return dspy.settings.arbor_api_base@staticmethoddef finetune(- job: ArborTrainingJob,- model: str,- train_data: List[Dict[str, Any]],- train_data_format: Optional[TrainDataFormat],- train_kwargs: Optional[Dict[str, Any]] = None+ job: ArborTrainingJob,+ model: str,+ train_data: List[Dict[str, Any]],+ train_data_format: Optional[TrainDataFormat],+ train_kwargs: Optional[Dict[str, Any]] = None,) -> str:# TODO: We want to re-factor finetune so that it takes in an LM.# Until then, we use the following to get the api information. The@@ -440,11 +442,7 @@ class ArborProvider(Provider):return provider_file.id@staticmethod- def _start_remote_training(- train_file_id: str,- model: str,- train_kwargs: Dict[str, Any]- ) -> str:+ def _start_remote_training(train_file_id: str, model: str, train_kwargs: Dict[str, Any]) -> str:train_kwargs = train_kwargs or {}original_base_url = openai.base_urlopenai.base_url = ArborProvider._get_arbor_base_api()@@ -458,9 +456,9 @@ class ArborProvider(Provider):@staticmethoddef wait_for_job(- job: TrainingJob,- training_kwargs: Dict[str, Any],- poll_frequency: int = 20,+ job: TrainingJob,+ training_kwargs: Dict[str, Any],+ poll_frequency: int = 20,):done = Falsecur_event_id = None@@ -468,7 +466,6 @@ class ArborProvider(Provider):while not done:# Report estimated time if not already reportedif not reported_estimated_time:-original_base_url = openai.base_urlopenai.base_url = ArborProvider._get_arbor_base_api()remote_job = openai.fine_tuning.jobs.retrieve(job.provider_job_id)diff --git a/dspy/clients/provider.py b/dspy/clients/provider.pyindex 1722a586e..a56b35502 100644--- a/dspy/clients/provider.py+++ b/dspy/clients/provider.py@@ -35,7 +35,6 @@ class TrainingJob(Future):raise NotImplementedError-class ReinforceJob:def __init__(self, lm: "LM", train_kwargs: Optional[Dict[str, Any]] = None):self.lm = LM@@ -58,7 +57,7 @@ class ReinforceJob:@abstractmethoddef update_model(self):raise NotImplementedError-+@abstractmethoddef save_checkpoint(self, checkpoint_name: str):raise NotImplementedErrordiff --git a/dspy/primitives/tool.py b/dspy/primitives/tool.pyindex 0bd19d955..bf4a208f5 100644--- a/dspy/primitives/tool.py+++ b/dspy/primitives/tool.py@@ -172,13 +172,9 @@ class Tool:return convert_mcp_tool(session, tool)-def __repr__(self):- return (- f"Tool(name={self.name}, desc={self.desc}, args={self.args})"- )--+ return f"Tool(name={self.name}, desc={self.desc}, args={self.args})"+def __str__(self):desc = f", whose description is <desc>{self.desc}</desc>.".replace("\n", " ") if self.desc else "."arg_desc = f"It takes arguments {self.args} in JSON format."diff --git a/dspy/retrievers/embeddings.py b/dspy/retrievers/embeddings.pyindex ded5b1532..bf75aa3ce 100644--- a/dspy/retrievers/embeddings.py+++ b/dspy/retrievers/embeddings.py@@ -1,5 +1,7 @@-import numpy as npfrom typing import Any, List, Optional++import numpy as np+from dspy.utils.unbatchify import Unbatchify# TODO: Add .save and .load methods!@@ -14,7 +16,7 @@ class Embeddings:callbacks: Optional[List[Any]] = None,cache: bool = False,brute_force_threshold: int = 20_000,- normalize: bool = True+ normalize: bool = True,):assert cache is False, "Caching is not supported for embeddings-based retrievers"@@ -44,7 +46,7 @@ class Embeddings:pids = self._faiss_search(q_embeds, self.k * 10) if self.index else Nonepids = np.tile(np.arange(len(self.corpus)), (len(queries), 1)) if pids is None else pids-+return self._rerank_and_predict(q_embeds, pids)def _build_faiss(self):@@ -60,8 +62,10 @@ class Embeddings:quantizer = faiss.IndexFlatL2(dim)index = faiss.IndexIVFPQ(quantizer, dim, partitions, nbytes, 8)- print(f"Training a {nbytes}-byte FAISS index with {partitions} partitions, based on "- f"{len(self.corpus)} x {dim}-dim embeddings")+ print(+ f"Training a {nbytes}-byte FAISS index with {partitions} partitions, based on "+ f"{len(self.corpus)} x {dim}-dim embeddings"+ )index.train(self.corpus_embeddings)index.add(self.corpus_embeddings)index.nprobe = min(16, partitions)@@ -73,12 +77,12 @@ class Embeddings:def _rerank_and_predict(self, q_embeds: np.ndarray, candidate_indices: np.ndarray):candidate_embeddings = self.corpus_embeddings[candidate_indices]- scores = np.einsum('qd,qkd->qk', q_embeds, candidate_embeddings)+ scores = np.einsum("qd,qkd->qk", q_embeds, candidate_embeddings)- top_k_indices = np.argsort(-scores, axis=1)[:, :self.k]+ top_k_indices = np.argsort(-scores, axis=1)[:, : self.k]top_indices = candidate_indices[np.arange(len(q_embeds))[:, None], top_k_indices]- return [([self.corpus[idx] for idx in indices], [idx for idx in indices]) for indices in top_indices]+ return [([self.corpus[idx] for idx in indices], [idx for idx in indices]) for indices in top_indices] # noqa: C416def _normalize(self, embeddings: np.ndarray):norms = np.linalg.norm(embeddings, axis=1, keepdims=True)diff --git a/dspy/teleprompt/simba.py b/dspy/teleprompt/simba.pyindex 21461b84e..45fce7c69 100644--- a/dspy/teleprompt/simba.py+++ b/dspy/teleprompt/simba.py@@ -30,21 +30,21 @@ class SIMBA(Teleprompter):Initializes SIMBA.Args:- metric (Callable): A function that takes an Example and a prediction_dict+ metric (Callable): A function that takes an Example and a prediction_dictas input and returns a float.bsize (int, optional): Mini-batch size. Defaults to 32.- num_candidates (int, optional): Number of new candidate programs to produce+ num_candidates (int, optional): Number of new candidate programs to produceper iteration. Defaults to 6.max_steps (int, optional): Number of optimization steps to run. Defaults to 8.- max_demos (int, optional): Maximum number of demos a predictor can hold+ max_demos (int, optional): Maximum number of demos a predictor can holdbefore dropping some. Defaults to 4.- demo_input_field_maxlen (int, optional): Maximum number of characters to keep+ demo_input_field_maxlen (int, optional): Maximum number of characters to keepin an input field when building a new demo. Defaults to 100,000.- num_threads (int, optional): Number of threads for parallel execution.+ num_threads (int, optional): Number of threads for parallel execution.Defaults to None.- temperature_for_sampling (float, optional): Temperature used for picking+ temperature_for_sampling (float, optional): Temperature used for pickingprograms during the trajectory-sampling step. Defaults to 0.2.- temperature_for_candidates (float, optional): Temperature used for picking+ temperature_for_candidates (float, optional): Temperature used for pickingthe source program for building new candidates. Defaults to 0.2."""self.metric = metric@@ -296,8 +296,8 @@ class SIMBA(Teleprompter):sys_scores = [outputs[i]["score"] for i in range(start, end)]register_new_program(cand_sys, sys_scores)- M = len(winning_programs) - 1 # noqa: N806- N = self.num_candidates + 1 # noqa: N806+ M = len(winning_programs) - 1 # noqa: N806+ N = self.num_candidates + 1 # noqa: N806if M < 1:program_idxs = [0] * Nelse: