drgrpo_learner_test.py

# Copyright 2025 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from unittest import mock
from absl.testing import absltest
from absl.testing import parameterized
from flax import nnx
import jax
import jax.numpy as jnp
import numpy as np
from tunix.rl import function_registry as fr
from tunix.rl.grpo import drgrpo_learner as drgrpo_lib
from tunix.rl.grpo import grpo_learner as grpo_lib
from tunix.tests import test_common as tc

jax.config.update("jax_threefry_partitionable", False)


class DrGRPOlearnerTest(parameterized.TestCase):

  def setUp(self):
    super().setUp()
    self.mock_model = mock.MagicMock()
    self.pad_id = 0
    self.eos_id = 1

    # Common data shapes
    self.batch_size = 2
    self.seq_len = 4
    self.prompt_ids = jnp.zeros(
        (self.batch_size, self.seq_len), dtype=jnp.int32
    )
    self.completion_ids = jnp.ones(
        (self.batch_size, self.seq_len), dtype=jnp.int32
    )
    self.completion_mask = jnp.array(
        [[1, 1, 1, 0], [1, 1, 0, 0]], dtype=jnp.float32
    )
    self.advantages = jnp.array([0.5, -0.2], dtype=jnp.float32)
    self.ref_per_token_logps = (
        jnp.ones_like(self.completion_ids, dtype=jnp.float32) * -0.2
    )
    self.old_per_token_logps = (
        jnp.ones_like(self.completion_ids, dtype=jnp.float32) * -0.15
    )

  def create_train_example(self):
    example = mock.MagicMock()
    example.prompt_ids = self.prompt_ids
    example.completion_ids = self.completion_ids
    example.completion_mask = self.completion_mask
    example.advantages = self.advantages
    example.ref_per_token_logps = self.ref_per_token_logps
    example.old_per_token_logps = self.old_per_token_logps
    example.segment_ids = None
    example.segment_positions = None
    return example

  def test_create_config(self):
    drgrpo_config = drgrpo_lib.DrGRPOConfig(
        epsilon=0.1, num_generations=5, num_iterations=3, beta=0.123
    )
    self.assertEqual(drgrpo_config.algo_variant, "drgrpo")
    self.assertEqual(drgrpo_config.advantage_estimator, "drgrpo")
    self.assertEqual(drgrpo_config.loss_agg_mode, "sequence-mean-token-scale")
    self.assertEqual(drgrpo_config.num_generations, 5)
    self.assertEqual(drgrpo_config.num_iterations, 3)
    self.assertEqual(drgrpo_config.epsilon, 0.1)
    self.assertEqual(drgrpo_config.beta, 0.123)

  def test_drgrpo_advantage_estimator(self):
    drgrpo_config = drgrpo_lib.DrGRPOConfig()
    grpo_config = grpo_lib.GRPOConfig()

    grpo_advantage_estimator = fr.get_advantage_estimator(
        grpo_config.advantage_estimator
    )
    drgrpo_advantage_estimator = fr.get_advantage_estimator(
        drgrpo_config.advantage_estimator
    )

    # Batch size 3 with group size 2.
    n_generations = 2
    rewards = jnp.array([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])
    num_generations = n_generations
    grpo_advantages = grpo_advantage_estimator(
        rewards=rewards.ravel(), num_generations=num_generations
    )
    drgrpo_advantages = drgrpo_advantage_estimator(
        rewards=rewards.ravel(), num_generations=num_generations
    )
    # Dr. GRPO advantages are not scaled by the standard deviation.
    # Std. across groups above is the same by construction.
    std_factor = jnp.array([1.0, 2.0]).std(ddof=1) + 1e-6
    np.testing.assert_allclose(grpo_advantages * std_factor, drgrpo_advantages)

  def test_drgrpo_loss_fn(self):
    drgrpo_config = drgrpo_lib.DrGRPOConfig()

    drgrpo_loss_fn_impl = fr.default_registry.get(
        "policy_loss_fn", drgrpo_config.policy_loss_fn
    )

    # Create the same input for both functions
    train_example = self.create_train_example()
    pad_id = self.pad_id
    eos_id = self.eos_id
    vocab = tc.MockVocab()
    model = tc.ToyTransformer(
        config=tc.ModelConfig(vocab_size=vocab.GetPieceSize()),
        rngs=nnx.Rngs(0),
    )

    # Call DrGRPO loss function
    drgrpo_loss, drgrpo_aux = drgrpo_loss_fn_impl(
        model, train_example, drgrpo_config, pad_id, eos_id
    )

    self.assertIn("kl", drgrpo_aux)
    self.assertTrue(jnp.isfinite(drgrpo_loss).all())

  def test_compute_advantages(self):
    rewards = jnp.array(
        [[0.57450044, 0.09968603, 0.7419659, 0.8941783, 0.59656656, 0.45325184]]
    )
    advantages = drgrpo_lib.compute_advantages(rewards, num_generations=3)
    expected_array = jnp.array([
        [0.10245, -0.372365, 0.269915, 0.246179, -0.051432, -0.194747],
    ])
    np.testing.assert_allclose(advantages, expected_array, rtol=1e-5, atol=1e-5)


if __name__ == "__main__":
  absltest.main()