Skip to content

Add GRPOConfig for Arbor #8882

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 7 commits into from
Closed

Add GRPOConfig for Arbor #8882

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
733d891
made new GRPO data class for train_kwargs, integrated to lm_local_arb…
zhassan223 Oct 1, 2025
fbfd080
refactored grpo into own file and added tests to lm local arbor, upda…
zhassan223 Oct 2, 2025
ba0332f
fixed rl_papillon implementation to use grpoConfig and fixed outdate…
zhassan223 Oct 2, 2025
50fedce
reconfigured arguments for GRPO to use config argument and phased out…
zhassan223 Oct 2, 2025
78093d2
Merge branch 'main' into main
zhassan223 Oct 2, 2025
8755d88
removed comments, restructured grpo dictionary function, fixed ruff w…
zhassan223 Oct 7, 2025
059c785
renamed grpoConfig -> arborGRPOConfig and adjusted files and tests ac…
zhassan223 Oct 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
39 changes: 20 additions & 19 deletions docs/docs/tutorials/rl_multihop/index.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -226,29 +226,30 @@
"metadata": {},
"outputs": [],
"source": [
"from dspy.teleprompt.grpo import GRPO\n",
"from dspy.teleprompt.grpo import GRPO, GRPOConfig\n",
"from dspy.clients.utils_finetune import MultiGPUConfig\n",
"\n",
"program = ResearchHop(num_docs=4, num_hops=2)\n",
"program.set_lm(local_lm)\n",
"\n",
"# NOTE: Training on 6 GPUs.\n",
"train_kwargs = {\n",
" \"per_device_train_batch_size\": 2,\n",
" \"gradient_accumulation_steps\": 8,\n",
" \"temperature\": 1.0,\n",
" \"beta\": 0.04,\n",
" \"learning_rate\": 1e-5,\n",
" \"gradient_checkpointing\": True,\n",
" \"gradient_checkpointing_kwargs\": {\"use_reentrant\": False},\n",
" \"bf16\": True,\n",
" \"lr_scheduler_type\": \"constant_with_warmup\",\n",
" \"max_prompt_length\": None,\n",
" \"max_completion_length\": None,\n",
" \"scale_rewards\": True,\n",
" \"max_grad_norm\": 0.5,\n",
" \"lora\": True,\n",
"}\n",
"config = GRPOConfig(\n",
" num_generations=4,\n",
" temperature=1.0,\n",
" beta=0.04,\n",
" per_device_train_batch_size=2,\n",
" learning_rate=1e-5,\n",
" gradient_accumulation_steps=8,\n",
" gradient_checkpointing=True,\n",
" gradient_checkpointing_kwargs={\"use_reentrant\": False},\n",
" bf16=True,\n",
" lr_scheduler_type=\"constant_with_warmup\",\n",
" max_prompt_length=None,\n",
" max_completion_length=None,\n",
" scale_rewards=True,\n",
" max_grad_norm=0.5,\n",
" lora=True,\n",
")\n",
"\n",
"compiler = GRPO(\n",
" metric=recall,\n",
Expand All @@ -259,7 +260,7 @@
" num_threads=16,\n",
" use_train_as_val=False,\n",
" num_steps_for_val=10,\n",
" train_kwargs=train_kwargs,\n",
" config=config,#now using GRPOConfig directly\n",
" report_train_scores=False,\n",
" gpu_config=MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1),\n",
")\n",
Expand All @@ -268,7 +269,7 @@
" student=program,\n",
" trainset=trainset,\n",
" valset=devset,\n",
")\n"
")"
]
},
{
Expand Down
41 changes: 21 additions & 20 deletions docs/docs/tutorials/rl_papillon/index.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -255,41 +255,42 @@
"metadata": {},
"outputs": [],
"source": [
"from dspy.teleprompt.grpo import GRPO\n",
"from dspy.teleprompt.grpo import GRPO, GRPOConfig\n",
"from dspy.clients.utils_finetune import MultiGPUConfig\n",
"\n",
"papillon = PAPILLON(untrusted_model=openai_lm)\n",
"papillon.set_lm(local_lm)\n",
"\n",
"# NOTE: Training on 3 GPUs.\n",
"train_kwargs = {\n",
" \"per_device_train_batch_size\": 8,\n",
" \"gradient_accumulation_steps\": 4,\n",
" \"temperature\": 1.0,\n",
" \"beta\": 0.04,\n",
" \"learning_rate\": 2e-6,\n",
" \"gradient_checkpointing\": True,\n",
" \"gradient_checkpointing_kwargs\": {\"use_reentrant\": False},\n",
" \"bf16\": True,\n",
" \"lr_scheduler_type\": \"constant_with_warmup\",\n",
" \"max_prompt_length\": None,\n",
" \"max_completion_length\": None,\n",
" \"scale_rewards\": True,\n",
" \"max_grad_norm\": 0.5,\n",
" \"lora\": True,\n",
"}\n",
"config = GRPOConfig(\n",
" num_generations=8,\n",
" temperature=1.0,\n",
" beta=0.04,\n",
" per_device_train_batch_size=8,\n",
" learning_rate=2e-6,\n",
" gradient_accumulation_steps=4,\n",
" gradient_checkpointing=True,\n",
" gradient_checkpointing_kwargs={\"use_reentrant\": False},\n",
" bf16=True,\n",
" lr_scheduler_type=\"constant_with_warmup\",\n",
" max_prompt_length=None,\n",
" max_completion_length=None,\n",
" scale_rewards=True,\n",
" max_grad_norm=0.5,\n",
" lora=True,\n",
")\n",
"\n",
"compiler = GRPO(\n",
" metric=compute_overall_score,\n",
" multitask=True,\n",
" num_dspy_examples_per_grpo_step=4,\n",
" num_samples_per_input=8,\n",
" num_rollouts_per_grpo_step=8,#changed from num_generations since that parameter doesn't exist anymore\n",
Copy link
Collaborator

@Ziems Ziems Oct 2, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good catch!

" exclude_demos=True,\n",
" num_train_steps=500,\n",
" num_threads=24,\n",
" use_train_as_val=False,\n",
" num_steps_for_val=10,\n",
" train_kwargs=train_kwargs,\n",
" config=config,#now using GRPOConfig directly\n",
" report_train_scores=False,\n",
" gpu_config=MultiGPUConfig(num_inference_gpus=2, num_training_gpus=2),\n",
")\n",
Expand All @@ -298,7 +299,7 @@
" student=papillon,\n",
" trainset=trainset,\n",
" valset=devset,\n",
")\n"
")"
]
},
{
Expand Down
6 changes: 2 additions & 4 deletions dspy/clients/lm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,16 +227,14 @@ def thread_function_wrapper():

return job

def reinforce(
self, train_kwargs, gpu_config: MultiGPUConfig = MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1)
) -> ReinforceJob:
def reinforce(self, config, gpu_config: MultiGPUConfig = MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1)) -> 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, err

job = self.provider.ReinforceJob(lm=self, train_kwargs=train_kwargs, gpu_config=gpu_config)
job = self.provider.ReinforceJob(lm=self, config=config, gpu_config=gpu_config)
job.initialize()
return job

Expand Down
104 changes: 13 additions & 91 deletions dspy/clients/lm_local_arbor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
import time
from datetime import datetime
from typing import TYPE_CHECKING, Any, TypedDict
from typing import TYPE_CHECKING, Any
from urllib.parse import urljoin

import openai
Expand All @@ -9,15 +9,12 @@
import dspy
from dspy.clients.provider import Provider, ReinforceJob, TrainingJob
from dspy.clients.utils_finetune import GRPOGroup, MultiGPUConfig, TrainDataFormat, TrainingStatus, save_data
from dspy.teleprompt.grpo.grpo_config import GRPOConfig

if TYPE_CHECKING:
from dspy.clients.lm import LM


class GRPOTrainKwargs(TypedDict):
num_generations: int


class ArborTrainingJob(TrainingJob):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
Expand Down Expand Up @@ -47,107 +44,32 @@ def status(self) -> TrainingStatus:


class ArborReinforceJob(ReinforceJob):
DEFAULT_TRAIN_KWARGS = { # noqa: RUF012
"temperature": 0.9,
"beta": 0.04,
"num_iterations": 1,
"per_device_train_batch_size": 8,
"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,
"lr_scheduler_type": "constant_with_warmup",
"max_prompt_length": None,
"max_completion_length": None,
"gradient_checkpointing_kwargs": None,
"bf16": False,
"scale_rewards": True,
"max_grad_norm": 1.0,
"report_to": "none",
"log_completions": True,
"logging_steps": 100,
# By default, none is the model's max context length
"max_context_length": None,
"lora": False,
}

def __init__(self, lm: "LM", train_kwargs: GRPOTrainKwargs, gpu_config: MultiGPUConfig = MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1)):
def __init__(self, lm: "LM", config: GRPOConfig, gpu_config: MultiGPUConfig = MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1)):
# The teleprompter must ensure that this is set
if "num_generations" not in train_kwargs:
raise ValueError("num_generations must be set in the training kwargs")
if not isinstance(config, GRPOConfig):
raise TypeError(f"Expected config to be of type GRPOConfig, but got {type(config)}")

self.lm = lm
self.train_kwargs = train_kwargs
self.config: GRPOConfig = config
self.provider_job_id = None
self.checkpoints = {}
self.last_checkpoint = None
self.gpu_config = gpu_config

def initialize(self):
# TODO(GRPO Team): Set provider job ID
num_generations = self.train_kwargs.get("num_generations")
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"]
)
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"]
)
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"]
)
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"]
)
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"]
)
lora = self.train_kwargs.get("lora", self.DEFAULT_TRAIN_KWARGS["lora"])
api_base = self.lm.kwargs["api_base"]

finetune_model = ArborProvider._remove_provider_prefix(self.lm.model)
# Only multi-GPU is supported for now
gpu_config_type = "multi"
data = {
"model": finetune_model,
"num_generations": num_generations,
"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,
"gpu_config": {
"type": gpu_config_type,
gpu_config_type: self.gpu_config,
},

# Create data payload from GRPOConfig
data = self.config.to_dict()
data["model"] = finetune_model
data["gpu_config"] = {
"type": gpu_config_type,
gpu_config_type: self.gpu_config,
}
url = urljoin(api_base, "fine_tuning/grpo/initialize")
headers = {"Content-Type": "application/json"}
Expand Down
4 changes: 2 additions & 2 deletions dspy/clients/provider.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,9 +36,9 @@ def status(self):


class ReinforceJob:
def __init__(self, lm: "LM", train_kwargs: dict[str, Any] | None = None, gpu_config: MultiGPUConfig = MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1)):
def __init__(self, lm: "LM", config: dict[str, Any] | None = None, gpu_config: MultiGPUConfig = MultiGPUConfig(num_inference_gpus=1, num_training_gpus=1)):
self.lm = lm
self.train_kwargs = train_kwargs or {}
self.config = config or {}
self.gpu_config = gpu_config
self.checkpoints = {}
self.last_checkpoint = None
Expand Down
4 changes: 4 additions & 0 deletions dspy/teleprompt/grpo/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,4 @@
from dspy.teleprompt.grpo.grpo import GRPO
from dspy.teleprompt.grpo.grpo_config import GRPOConfig

__all__ = ["GRPO", "GRPOConfig"]
Loading
Loading