chore(mlx-lm): fix the top_p implementation. (#602)

* chore(mlx-lm): clean up the top p imp

* chore: clean up

* chore: add test

* chore: address comments

* chore: clean up docs string

* chore: clean up test

llms/mlx_lm/sample_utils.py

@@ -0,0 +1,39 @@
+import mlx.core as mx
+
+
+def top_p_sampling(logits: mx.array, top_p: float, temperature: float) -> mx.array:
+    """
+    Apply top-p (nucleus) sampling to logits.
+
+    Args:
+        logits: The logits from the model's output.
+        top_p: The cumulative probability threshold for top-p filtering.
+        temperature: Temperature parameter for softmax distribution reshaping.
+    Returns:
+        token selected based on the top-p criterion.
+    """
+    if (
+        logits.dtype == mx.bfloat16
+    ):  # workaround for unable to load kernel contiguous_scan_inclusive_sum_bfloat16_bfloat16
+        logits = logits.astype(mx.float32)
+
+    # referenced implementation from https://github.com/huggingface/transformers/blob/main/src/transformers/generation/logits_process.py#L449-L460
+    probs = mx.softmax(logits / temperature, axis=-1)
+
+    # sort probs in ascending order
+    sorted_indices = mx.argsort(probs, axis=-1)
+    sorted_probs = probs[..., sorted_indices]
+
+    cumulative_probs = mx.cumsum(sorted_probs, axis=-1)
+
+    # select tokens with cumulative probs below threshold
+    top_probs = mx.where(
+        cumulative_probs > 1 - top_p,
+        sorted_probs,
+        mx.zeros_like(sorted_probs),
+    )
+
+    sorted_token = mx.random.categorical(mx.log(top_probs))
+    token = sorted_indices.squeeze(0)[sorted_token]
+
+    return token

llms/mlx_lm/utils.py

@@ -17,6 +17,8 @@ from huggingface_hub import snapshot_download
 from mlx.utils import tree_flatten
 from transformers import AutoConfig, AutoTokenizer, PreTrainedTokenizer
 
+from .sample_utils import top_p_sampling
+
 # Local imports
 from .tuner.utils import apply_lora_layers
 from .tuner.utils import dequantize as dequantize_model
@@ -144,23 +146,7 @@ def generate_step(
             token = mx.argmax(logits, axis=-1)
         else:
             if top_p > 0 and top_p < 1.0:
-                if (
-                    logits.dtype == mx.bfloat16
-                ):  # workdaround for unable to load kernel contiguous_scan_inclusive_sum_bfloat16_bfloat16
-                    logits = logits.astype(mx.float32)
-                probs = mx.softmax(logits / temp, axis=-1)
-
-                sorted_probs = mx.sort(probs)[::-1]
-                sorted_indices = mx.argsort(probs)[::-1]
-                cumulative_probs = mx.cumsum(sorted_probs, axis=-1)
-
-                top_probs = mx.where(
-                    cumulative_probs > 1 - top_p,
-                    sorted_probs,
-                    mx.zeros_like(sorted_probs),
-                )
-                sorted_token = mx.random.categorical(mx.log(top_probs))
-                token = sorted_indices.squeeze(0)[sorted_token]
+                token = top_p_sampling(logits, top_p, temp)
             else:
                 token = mx.random.categorical(logits * (1 / temp))