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_k and min_p refactor

Browse Source
This commit is contained in:
Neil Mehta
2025-03-08 10:12:28 -05:00
parent 58e912966a
commit 932b7c0510
2 changed files with 49 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
llms/tests/test_sample_utils.py
View File

@@ -9,28 +9,28 @@ class TestSampleUtils(unittest.TestCase):
probs = mx.array([0.9, 0.0, 0.0, 0.1])[None]
logits = mx.log(probs)
actual_logits = top_p_sampling(logits, 0.3)
actual_probs = mx.softmax(actual_logits.squeeze())
new_logits = top_p_sampling(logits, 0.3)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(actual_probs.tolist(), [1.0, 0.0, 0.0, 0.0])
actual_logits = top_p_sampling(logits, 0.95)
actual_probs = mx.softmax(actual_logits.squeeze())
new_logits = top_p_sampling(logits, 0.95)
actual_probs = mx.softmax(new_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())
new_logits = top_p_sampling(logits, 0.4)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(actual_probs.tolist(), [0.0, 1.0, 0.0, 0.0])
actual_logits = top_p_sampling(logits, 0.6)
actual_probs = mx.softmax(actual_logits.squeeze())
new_logits = top_p_sampling(logits, 0.6)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(
[round(p, 4) for p in actual_probs.tolist()], [0.0, 0.5556, 0.4444, 0.0]
)
actual_logits = top_p_sampling(logits, 0.95)
actual_probs = mx.softmax(actual_logits.squeeze())
new_logits = top_p_sampling(logits, 0.95)
actual_probs = mx.softmax(new_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)
@@ -39,8 +39,8 @@ class TestSampleUtils(unittest.TestCase):
# Batch mode works
probs = mx.array([[0.9, 0.0, 0.0, 0.1], [0.0, 0.8, 0.1, 0.1]])
logits = mx.log(probs)
actual_logits = top_p_sampling(logits, 0.5)
actual_probs = mx.softmax(actual_logits, axis=-1)
new_logits = top_p_sampling(logits, 0.5)
actual_probs = mx.softmax(new_logits, axis=-1)
self.assertEqual(
actual_probs.tolist(), [[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0]]
)
@@ -48,43 +48,50 @@ class TestSampleUtils(unittest.TestCase):
def test_min_p_sampling(self):
probs = mx.array([0.9, 0.0, 0.0, 0.1])[None]
logits = mx.log(probs)
temperature = 1.0
token = min_p_sampling(logits, 0.8)
self.assertEqual(token, 0)
new_logits = min_p_sampling(logits, 0.8)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(actual_probs.tolist(), [1.0, 0.0, 0.0, 0.0])
probs = mx.array([0.9, 0.0, 0.0, 0.1])[None]
logits = mx.log(probs)
temperature = 1.0
for _ in range(5):
token = min_p_sampling(logits, 0.05)
self.assertTrue(token in (0, 3))
new_logits = min_p_sampling(logits, 0.05)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(actual_probs.tolist(), mx.squeeze(probs).tolist())
# Batch mode works
probs = mx.array([[0.9, 0.0, 0.0, 0.1], [0.0, 0.8, 0.0, 0.1]])
logits = mx.log(probs)
tokens = min_p_sampling(logits, 0.7)
self.assertEqual(tokens.tolist(), [0, 1])
new_logits = min_p_sampling(logits, 0.7)
actual_probs = mx.softmax(new_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_top_k_sampling(self):
probs = mx.array([0.9, 0.0, 0.0, 0.1])[None]
logits = mx.log(probs)
token = top_k_sampling(logits, 1).item()
self.assertEqual(token, 0)
new_logits = top_k_sampling(logits, 1)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(actual_probs.tolist(), [1.0, 0.0, 0.0, 0.0])
probs = mx.array([0.5, 0.0, 0.0, 0.5])[None]
tokens = set()
for _ in range(100):
token = top_k_sampling(logits, 2)
tokens.add(token.item())
self.assertEqual(tokens, {0, 3})
probs = mx.array([0.6, 0.0, 0.1, 0.3])[None]
logits = mx.log(probs)
new_logits = top_k_sampling(logits, 2)
actual_probs = mx.softmax(new_logits.squeeze())
self.assertEqual(
[round(p, 4) for p in actual_probs.tolist()], [0.6667, 0.0, 0.0, 0.3333]
)
# Batch mode works
probs = mx.array([[0.9, 0.0, 0.0, 0.1], [0.0, 0.8, 0.0, 0.1]])
logits = mx.log(probs)
tokens = top_k_sampling(logits, 1)
self.assertEqual(tokens.tolist(), [0, 1])
new_logits = top_k_sampling(logits, 1)
actual_probs = mx.softmax(new_logits, axis=-1)
self.assertEqual(
actual_probs.tolist(), [[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0]]
)
if __name__ == "__main__":