diff --git a/llms/mlx_lm/sample_utils.py b/llms/mlx_lm/sample_utils.py index d7049f7d..5ad3d2c5 100644 --- a/llms/mlx_lm/sample_utils.py +++ b/llms/mlx_lm/sample_utils.py @@ -88,7 +88,6 @@ def make_logits_processors( def top_k_sampling( logprobs: mx.array, top_k: int, - temperature=1.0, ) -> mx.array: """ Sample from only the top K tokens ranked by probability. @@ -103,12 +102,11 @@ def top_k_sampling( f"`top_k` has to be an integer in the (0, {vocab_size}] interval," f" but is {top_k}." ) - logprobs = logprobs * (1 / temperature) mask_idx = mx.argpartition(-logprobs, kth=top_k - 1, axis=-1)[..., top_k:] masked_logprobs = mx.put_along_axis( logprobs, mask_idx, mx.array(-float("inf"), logprobs.dtype), axis=-1 ) - return mx.random.categorical(masked_logprobs, axis=-1) + return masked_logprobs @partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state) @@ -116,7 +114,6 @@ def min_p_sampling( logprobs: mx.array, min_p: float, min_tokens_to_keep: int = 1, - temperature=1.0, ) -> mx.array: """ Apply min-p sampling to the logprobs. @@ -144,8 +141,6 @@ def min_p_sampling( ) # reference implementation: https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L531-L605 - logprobs = logprobs * (1 / temperature) - # Indices sorted in decreasing order sorted_indices = mx.argsort(-logprobs, axis=-1) sorted_logprobs = mx.take_along_axis(logprobs, sorted_indices, axis=-1) @@ -163,9 +158,16 @@ def min_p_sampling( # Create pool of tokens with probability less than scaled min_p selected_logprobs = mx.where(tokens_to_remove, -float("inf"), sorted_logprobs) - # Return sampled tokens - sorted_tokens = mx.random.categorical(selected_logprobs, axis=-1)[:, None] - return mx.take_along_axis(sorted_indices, sorted_tokens, axis=-1).squeeze(1) + # Create a mapping to rearrange back to original indices + # Use argsort of sorted_indices to get the inverse permutation + inverse_indices = mx.argsort(sorted_indices, axis=-1) + + # Rearrange selected_logprobs back to original order + original_order_logprobs = mx.take_along_axis( + selected_logprobs, inverse_indices, axis=-1 + ) + + return original_order_logprobs @partial(mx.compile, inputs=mx.random.state, outputs=mx.random.state) diff --git a/llms/tests/test_sample_utils.py b/llms/tests/test_sample_utils.py index 5a3d8847..19b65e4f 100644 --- a/llms/tests/test_sample_utils.py +++ b/llms/tests/test_sample_utils.py @@ -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__":