Add corrected kl with importance sampling

[garyzhang99]

Description

Implement the corrected kl with importance sampling, as mentioned in DeepseekV3.2 paper.

Checklist

Please check the following items before code is ready to be reviewed.

• Code has passed all tests
• Docstrings have been added/updated in Google Style
• Documentation has been updated
• Code is ready for review

[gemini-code-assist]

Summary of Changes

Hello @garyzhang99, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request enhances the KL divergence calculation framework by integrating a "corrected K3" function that leverages importance sampling. This modification, inspired by the DeepseekV3.2 paper, allows for more robust and accurate KL divergence estimation, particularly when comparing current policy log probabilities against a reference and an old policy. The changes involve extending the KL function API, implementing the new corrected K3 logic, and updating the training actors to pass the necessary old policy log probabilities.

Highlights

• New KL Function: Introduced CorrectedK3Fn for KL divergence calculation with importance sampling, based on the DeepseekV3.2 paper. This new function provides a more robust KL estimation by incorporating an 'old policy' log probability.
• API Extension for Importance Sampling: The calculate_kl and calculate_kl_loss methods in the base KLFn class and all its existing implementations (DummyKLFn, K1Fn, K2Fn, K3Fn, LowVarKLFn, AbsFn) now accept an optional old_logprob parameter to facilitate importance sampling.
• Integration into Training Actors: Updated dp_actor.py and megatron_actor.py to pass the old_logprob parameter to the kl_fn.calculate_kl_loss method, enabling the use of importance sampling in the training loop.
• Comprehensive Unit Testing: Added a new test file tests/algorithm/kl_fn_test.py with extensive unit tests for the new corrected_k3 function, covering its fallback mechanism to standard K3 when old_logprob is absent, its behavior with old_logprob, and various loss aggregation modes.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a corrected KL divergence calculation using importance sampling, CorrectedK3Fn. The changes are well-structured: the new logic is encapsulated in its own class, the base KLFn and its subclasses are updated for a consistent interface, and the trainers are modified to pass the necessary old_logprob. The addition of a comprehensive test suite is excellent, covering new functionality, fallback behavior, and edge cases, which increases confidence in the implementation's correctness. My main feedback is a minor suggestion to improve code clarity by replacing hardcoded values with named constants. Overall, this is a high-quality contribution.

[garyzhang99]

/unittest-module-algorithm

[github-actions]

Summary

Tests 📝	Passed ✅	Failed ❌	Skipped ⏭️	Other ❓	Flaky 🍂	Duration ⏱️
25	22	3	0	0	0	18.3s

Failed Tests

Failed Tests ❌	Fail Message
❌ tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_fallback	The test failed in the call phase due to an assertion error
❌ tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_same_policy	The test failed in the call phase due to an assertion error
❌ tests/algorithm/kl_fn_test.py::KLFnTest::test_kl_loss_aggregation_modes	The test failed in the call phase

Tests

Test Name	Status	Flaky	Duration
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_batch_level_std_grpo	✅		41ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_batch_level_step_wise_grpo_advantage	✅		2ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_duplicate_grpo	✅		5ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_grpo_advantage	✅		3ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_grpo_correct_bias	✅		2ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_grpo_reward_std	✅		1ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_step_wise_grpo_advantage	✅		2ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_step_wise_grpo_with_std_threshold	✅		2ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_abs_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_fallback	❌		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_loss	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_same_policy	❌		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_with_old_logprob	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_dummy_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_k1_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_k2_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_k3_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_kl_loss_aggregation_modes	❌		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_low_var_kl_fn	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_dpo_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_gspo_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_mix_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_opmd_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_ppo_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_sft_policy_loss	✅		1ms

Github Test Reporter by CTRF 💚

[garyzhang99]

/unittest-module-algorithm

[github-actions]

Summary

Tests 📝	Passed ✅	Failed ❌	Skipped ⏭️	Other ❓	Flaky 🍂	Duration ⏱️
25	25	0	0	0	0	18.1s

Tests

Test Name	Status	Flaky	Duration
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_batch_level_std_grpo	✅		42ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_batch_level_step_wise_grpo_advantage	✅		2ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_duplicate_grpo	✅		5ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_grpo_advantage	✅		3ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_grpo_correct_bias	✅		2ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_grpo_reward_std	✅		1ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_step_wise_grpo_advantage	✅		2ms
tests/algorithm/advantage_fn_test.py::TestGroupedAdvantageFn::test_step_wise_grpo_with_std_threshold	✅		2ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_abs_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_fallback	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_loss	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_same_policy	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_corrected_k3_with_old_logprob	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_dummy_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_k1_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_k2_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_k3_kl_fn	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_kl_loss_aggregation_modes	✅		1ms
tests/algorithm/kl_fn_test.py::KLFnTest::test_low_var_kl_fn	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_dpo_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_gspo_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_mix_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_opmd_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_ppo_policy_loss	✅		1ms
tests/algorithm/policy_loss_test.py::VerlPolicyLossTest::test_sft_policy_loss	✅		1ms

Github Test Reporter by CTRF 💚

[garyzhang99]

/unittest-module-trainer

[github-actions]

Summary

Tests 📝	Passed ✅	Failed ❌	Skipped ⏭️	Other ❓	Flaky 🍂	Duration ⏱️
22	20	0	2	0	0	42m 38s

Skipped

Tests	Status
tests/trainer/trainer_test.py::TestMultiModalGRPO::test_trainer	skipped ⏭️
tests/trainer/trainer_test.py::TestMultiModalSFT::test_trainer	skipped ⏭️

Tests

Test Name	Status	Flaky	Duration
tests/trainer/trainer_test.py::TestTrainerCountdown_0_fsdp::test_trainer	✅		3m 16s
tests/trainer/trainer_test.py::TestTrainerCountdown_1_megatron::test_trainer	✅		4m 45s
tests/trainer/trainer_test.py::TestStepAheadAsyncRL::test_trainer	✅		1m 29s
tests/trainer/trainer_test.py::TestTrainerGSM8K_0_fsdp::test_trainer	✅		1m 19s
tests/trainer/trainer_test.py::TestTrainerGSM8K_1_fsdp2::test_trainer	✅		1m 18s
tests/trainer/trainer_test.py::TestTrainerGSM8K_2_fsdp::test_trainer	✅		1m 24s
tests/trainer/trainer_test.py::TestTrainerGSM8K_3_fsdp2::test_trainer	✅		1m 33s
tests/trainer/trainer_test.py::TestTrainerSFTWarmupGSM8K::test_trainer	✅		2m 31s
tests/trainer/trainer_test.py::TestTrainerDPO::test_trainer	✅		1m
tests/trainer/trainer_test.py::TestTrainerSFT::test_trainer	✅		57.7s
tests/trainer/trainer_test.py::TestTrainerToolsSFT::test_trainer_tools	✅		57.8s
tests/trainer/trainer_test.py::TestFullyAsyncMode_0_fsdp::test_fully_async_mode	✅		1m 55s
tests/trainer/trainer_test.py::TestFullyAsyncMode_1_fsdp::test_fully_async_mode	✅		1m 53s
tests/trainer/trainer_test.py::TestFullyAsyncMode_2_megatron::test_fully_async_mode	✅		2m 37s
tests/trainer/trainer_test.py::TestTrainerCheckpointSave_0_fsdp::test_trainer	✅		2m 19s
tests/trainer/trainer_test.py::TestTrainerCheckpointSave_1_megatron::test_trainer	✅		4m 25s
tests/trainer/trainer_test.py::TestTrainerMIX::test_trainer	✅		2m 28s
tests/trainer/trainer_test.py::TestMultiModalGRPO::test_trainer	⏭️		810ms
tests/trainer/trainer_test.py::TestMultiModalSFT::test_trainer	⏭️		807ms
tests/trainer/trainer_test.py::TestTrainerLoRA::test_trainer	✅		3m 31s
tests/trainer/trainer_test.py::TestOverRollout::test_trainer	✅		1m 22s
tests/trainer/trainer_test.py::TestTrainerPromptTruncation::test_trainer	✅		1m 11s

Github Test Reporter by CTRF 💚