Skip to content

Reimplement batched_rewards to fix compatibility, broken in #2698 #2876

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Reimplement batched_rewards to fix compatibility, broken in #2698 #2876

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
47 changes: 46 additions & 1 deletion torchtune/dev/rl/rewards.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,10 +7,15 @@
import re
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from typing import Optional
from typing import Optional, Union

import torch

from torchtune.modules.transforms.tokenizers import (
HuggingFaceModelTokenizer,
ModelTokenizer,
)


@dataclass
class RewardOutput:
Expand Down Expand Up @@ -216,3 +221,43 @@ def __call__(
},
successes=successes,
)


def batched_rewards(
tokenizer: Union[ModelTokenizer, HuggingFaceModelTokenizer],
completions: torch.Tensor,
answers: list[list[str]],
device: torch.device,
Copy link

Choose a reason for hiding this comment

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

You will need to pass the device parameter of batched_rewards in https://github.com/pytorch/torchtune/blob/main/recipes/dev/grpo_full_finetune_distributed.py#L649

But even then an error occurs regarding how the answers per batch per roll out are handled.

[rank0]: Traceback (most recent call last):
[rank0]:   File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 948, in <module>
[rank0]:     sys.exit(recipe_main())
[rank0]:              ^^^^^^^^^^^^^
[rank0]:   File "/tmp/torchtune/torchtune/config/_parse.py", line 99, in wrapper
[rank0]:     sys.exit(recipe_main(conf))
[rank0]:              ^^^^^^^^^^^^^^^^^
[rank0]:   File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 943, in recipe_main
[rank0]:     recipe.train()
[rank0]:   File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 812, in train
[rank0]:     trajectory = self.generate_trajectory_batched(tokens, answers)
[rank0]:                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 703, in generate_trajectory_batched
[rank0]:     self.generate_trajectory(batch_input_ids, batch_answers)
[rank0]:   File "/tmp/torchtune/recipes/dev/grpo_full_finetune_distributed.py", line 649, in generate_trajectory
[rank0]:     rewards, successes = batched_rewards(self._tokenizer, responses, answers, self._device)
[rank0]:                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/tmp/torchtune/torchtune/dev/rl/rewards.py", line 251, in batched_rewards
[rank0]:     answers_list.append(answers[b][g])
[rank0]:                         ~~~~~~~~~~^^^
[rank0]: IndexError: string index out of range

) -> tuple[torch.Tensor, torch.Tensor]:
reward_funcs = [
ThinkingAnswerFormattingReward("think", "answer", 1.0),
FormattedMathCorrectnessReward("think", 50.0),
]

num_reward_funcs = len(reward_funcs)
batch_size, group_size, _ = completions.shape

rewards_tensor = torch.zeros(
batch_size, group_size, num_reward_funcs, dtype=torch.bfloat16, device=device
)
successes_tensor = torch.zeros_like(rewards_tensor)

completions_list = []
answers_list = []
for b in range(batch_size):
for g in range(group_size):
completion_text = tokenizer.decode(completions[b, g].tolist())
completions_list.append(completion_text)
answers_list.append(answers[b][g])

completions_ids = torch.zeros(len(completions_list)) # dummy

for rw_idx, reward_func in enumerate(reward_funcs):
reward_obj = reward_func(completions_ids, completions_list, answers_list)
batch_reward = reward_obj.total_reward.view(batch_size, group_size)
batch_success = reward_obj.successes.view(batch_size, group_size)

rewards_tensor[:, :, rw_idx] = batch_reward
successes_tensor[:, :, rw_idx] = batch_success

return rewards_tensor, successes_tensor