Su-RoPE(Rotary Position Embedding) for Phi-3 (#813)

* Su-RoPE

* nits

* Update su_rope.py

* Update su_rope.py

Per GPT4: "The error TypeError: 'type' object is not subscriptable is caused by using the type hint list[float] in a version of Python that does not support it. This syntax is only available in Python 3.9 and later."

* Ran isort

---------

Co-authored-by: Awni Hannun <awni@apple.com>

llms/mlx_lm/models/phi3.py

@@ -5,6 +5,7 @@ import mlx.core as mx
 import mlx.nn as nn
 
 from .base import BaseModelArgs
+from .su_rope import SuScaledRotaryEmbedding
 
 
 @dataclass
@@ -20,6 +21,8 @@ class ModelArgs(BaseModelArgs):
     rope_theta: float = 10000
     rope_traditional: bool = False
     rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    max_position_embeddings: int = 131072
+    original_max_position_embeddings: int = 4096
 
     def __post_init__(self):
         if self.num_key_value_heads is None:
@@ -30,9 +33,9 @@ class ModelArgs(BaseModelArgs):
             if not all(key in self.rope_scaling for key in required_keys):
                 raise ValueError(f"rope_scaling must contain keys {required_keys}")
 
-            if self.rope_scaling["type"] != "linear":
+            if self.rope_scaling["type"] not in ["su", "linear"]:
                 print(
-                    "[WARNING] rope_scaling 'type' currently only supports 'linear' setting rope scaling to false."
+                    "[WARNING] rope_scaling 'type' currently only supports 'linear' and 'su'; setting rope scaling to false."
                 )
                 self.rope_scaling = None
 
@@ -53,11 +56,21 @@ class Attention(nn.Module):
         self.qkv_proj = nn.Linear(dim, op_size, bias=False)
         self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
 
-        rope_scale = (
-            1 / args.rope_scaling["factor"]
-            if args.rope_scaling is not None and args.rope_scaling["type"] == "linear"
-            else 1
+        rope_scale = 1.0
+        if args.rope_scaling and args.rope_scaling["type"] == "su":
+            self.rope = SuScaledRotaryEmbedding(
+                head_dim,
+                traditional=False,
+                base=args.rope_theta,
+                scale=rope_scale,
+                max_position_embeddings=args.max_position_embeddings,
+                original_max_position_embeddings=args.original_max_position_embeddings,
+                short_factor=args.rope_scaling["short_factor"],
+                long_factor=args.rope_scaling["long_factor"],
             )
+        else:
+            if args.rope_scaling and args.rope_scaling["type"] == "linear":
+                rope_scale = 1 / args.rope_scaling["factor"]
             self.rope = nn.RoPE(
                 head_dim,
                 traditional=args.rope_traditional,

llms/mlx_lm/models/su_rope.py

@@ -0,0 +1,79 @@
+import math
+from typing import List, Union
+
+import mlx.core as mx
+
+
+class SuScaledRotaryEmbedding:
+    def __init__(
+        self,
+        dims: int,
+        traditional: bool = False,
+        base: float = 10000.0,
+        scale: float = 1.0,
+        max_position_embeddings: int = 131072,
+        original_max_position_embeddings: int = 4096,
+        short_factor: Union[List[float], float] = 1.0,
+        long_factor: Union[List[float], float] = 1.0,
+    ):
+        """
+        Phi3Su Scaled Rotary Embedding layer for Phi-3 models.
+
+        Args:
+            dims (int): The feature dimensions to be rotated.
+            traditional (bool, optional): Unused. Default: ``False``.
+            base (int, optional): Base for the exponential scaling.
+            scale (float, optional): The scale used to scale the positions.
+              Default: ``1.0``.
+            max_position_embeddings (int, optional): The maximum sequence
+              length that this model was trained with. This is used to determine
+              the size of the original RoPE embeddings when using long scaling.
+              Default: ``131072``.
+            original_max_position_embeddings (int, optional): The maximum
+              sequence length that this model was trained with. This is used to
+              determine the size of the original RoPE embeddings when using long
+              scaling. Default: ``4096``.
+            short_factor (float or list[float], optional): List of scaling
+              factors for sequences of length lesser than
+              ``original_max_position_embeddings``. Default: ``1.0``.
+            long_factor (float or list[float], optional): List of scaling
+              factors for sequences of length greater than
+              ``original_max_position_embeddings``.  Default: ``1.0``.
+        """
+        self.inv_freq_short = 1.0 / (
+            mx.array(short_factor, dtype=mx.float32)
+            * base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
+        )
+        self.inv_freq_long = 1.0 / (
+            scale
+            * mx.array(long_factor, dtype=mx.float32)
+            * base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
+        )
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.scaling_factor = math.sqrt(
+            1
+            + math.log(max_position_embeddings / original_max_position_embeddings)
+            / math.log(original_max_position_embeddings)
+        )
+
+    def _get_cos_sin(self, offset, L):
+        position_ids = mx.arange(offset, offset + L, dtype=mx.float32)
+        inv_freq = (
+            self.inv_freq_long
+            if (offset + L) > self.original_max_position_embeddings
+            else self.inv_freq_short
+        )
+        freqs = position_ids[:, None] * inv_freq[None, :]
+        emb = mx.concatenate([freqs, freqs], axis=-1)
+        cos = mx.cos(emb) * self.scaling_factor
+        sin = mx.sin(emb) * self.scaling_factor
+        return cos, sin
+
+    def __call__(self, x, offset: int = 0):
+        def _rotate_half(_x):
+            midpoint = _x.shape[-1] // 2
+            x1, x2 = _x[..., :midpoint], _x[..., midpoint:]
+            return mx.concatenate([-x2, x1], axis=-1)
+
+        cos, sin = self._get_cos_sin(offset, x.shape[2])
+        return (x * cos) + (_rotate_half(x) * sin)