Switch to fast RMS/LN Norm (#603)

* use nn.RMSNorm, use sdpa, cleanup

* bump mlx versions

* minor update

* use fast layer norm

* version bump

* update requirement for whisper

* update requirement for gguf

llms/gguf_llm/models.py

@@ -49,20 +49,6 @@ class ModelArgs:
         )
 
 
-class RMSNorm(nn.Module):
-    def __init__(self, dims: int, eps: float = 1e-5):
-        super().__init__()
-        self.weight = mx.ones((dims,))
-        self.eps = eps
-
-    def _norm(self, x):
-        return x * mx.rsqrt(x.square().mean(-1, keepdims=True) + self.eps)
-
-    def __call__(self, x):
-        output = self._norm(x.astype(mx.float32)).astype(x.dtype)
-        return self.weight * output
-
-
 class Attention(nn.Module):
     def __init__(self, args: ModelArgs):
         super().__init__()
@@ -107,10 +93,6 @@ class Attention(nn.Module):
         keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
         values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
 
-        if self.repeats > 1:
-            keys = mx.repeat(keys, self.repeats, axis=1)
-            values = mx.repeat(values, self.repeats, axis=1)
-
         if cache is not None:
             key_cache, value_cache = cache
             queries = self.rope(queries, offset=key_cache.shape[2])
@@ -121,11 +103,10 @@ class Attention(nn.Module):
             queries = self.rope(queries)
             keys = self.rope(keys)
 
-        scores = (queries * self.scale) @ keys.transpose(0, 1, 3, 2)
-        if mask is not None:
-            scores += mask
-        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(scores.dtype)
-        output = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
+        output = mx.fast.scaled_dot_product_attention(
+            queries, keys, values, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
         return self.o_proj(output), (keys, values)
 
 
@@ -147,8 +128,10 @@ class TransformerBlock(nn.Module):
         self.hidden_size = args.hidden_size
         self.self_attn = Attention(args)
         self.mlp = MLP(args.hidden_size, args.intermediate_size)
-        self.input_layernorm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
-        self.post_attention_layernorm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
         self.args = args
 
     def __call__(
@@ -175,7 +158,7 @@ class LlamaModel(nn.Module):
         self.layers = [
             TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
         ]
-        self.norm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
 
     def __call__(
         self,

llms/gguf_llm/requirements.txt

@@ -1,4 +1,4 @@
-mlx>=0.0.11
+mlx>=0.8
 numpy
 protobuf==3.20.2
 sentencepiece