zhangyiss/mlx-examples
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx-examples.git synced 2025-12-16 02:08:55 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

top_p refactor

Browse Source
This commit is contained in:
Neil Mehta
2025-03-08 08:55:49 -05:00
parent d2e02b3aae
commit 58e912966a
2 changed files with 37 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
llms/tests/test_sample_utils.py
View File

@@ -8,31 +8,42 @@ class TestSampleUtils(unittest.TestCase):
def test_top_p_sampling(self):
probs = mx.array([0.9, 0.0, 0.0, 0.1])[None]
logits = mx.log(probs)
temperature = 1.0
token = top_p_sampling(logits, 0.3, temperature).item()
self.assertEqual(token, 0)
actual_logits = top_p_sampling(logits, 0.3)
actual_probs = mx.softmax(actual_logits.squeeze())
self.assertEqual(actual_probs.tolist(), [1.0, 0.0, 0.0, 0.0])
token = top_p_sampling(logits, 0.95, temperature).item()
self.assertTrue(token in (0, 3))
actual_logits = top_p_sampling(logits, 0.95)
actual_probs = mx.softmax(actual_logits.squeeze())
self.assertEqual(probs.squeeze().tolist(), actual_probs.tolist())
probs = mx.array([0.0, 0.5, 0.4, 0.1])[None]
logits = mx.log(probs)
actual_logits = top_p_sampling(logits, 0.4)
actual_probs = mx.softmax(actual_logits.squeeze())
self.assertEqual(actual_probs.tolist(), [0.0, 1.0, 0.0, 0.0])
token = top_p_sampling(logits, 0.4, temperature).item()
self.assertEqual(token, 1)
actual_logits = top_p_sampling(logits, 0.6)
actual_probs = mx.softmax(actual_logits.squeeze())
self.assertEqual(
[round(p, 4) for p in actual_probs.tolist()], [0.0, 0.5556, 0.4444, 0.0]
)
token = top_p_sampling(logits, 0.6, temperature).item()
self.assertTrue(token in (1, 2))
token = top_p_sampling(logits, 0.95, temperature).item()
self.assertTrue(token in (1, 2, 3))
actual_logits = top_p_sampling(logits, 0.95)
actual_probs = mx.softmax(actual_logits.squeeze())
actual_rounded = [round(p, 4) for p in actual_probs.tolist()]
expected_rounded = [0.0, 0.5, 0.4, 0.1]
self.assertEqual(actual_rounded, expected_rounded)
self.assertAlmostEqual(sum(actual_probs.tolist()), 1.0)
# Batch mode works
probs = mx.array([[0.9, 0.0, 0.0, 0.1], [0.0, 0.8, 0.0, 0.1]])
probs = mx.array([[0.9, 0.0, 0.0, 0.1], [0.0, 0.8, 0.1, 0.1]])
logits = mx.log(probs)
tokens = top_p_sampling(logits, 0.5, temperature)
self.assertEqual(tokens.tolist(), [0, 1])
actual_logits = top_p_sampling(logits, 0.5)
actual_probs = mx.softmax(actual_logits, axis=-1)
self.assertEqual(
actual_probs.tolist(), [[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0]]
)
def test_min_p_sampling(self):
probs = mx.array([0.9, 0.0, 0.0, 0.1])[None]