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

llava/language.py

@@ -45,20 +45,6 @@ class TextConfig:
                 raise ValueError("rope_scaling 'type' currently only supports 'linear'")
 
 
-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, config: TextConfig):
         super().__init__()
@@ -105,10 +91,6 @@ 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])
@@ -119,11 +101,10 @@ 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)
 
 
@@ -145,8 +126,8 @@ class TransformerBlock(nn.Module):
         self.hidden_size = config.hidden_size
         self.self_attn = Attention(config)
         self.mlp = MLP(config.hidden_size, config.intermediate_size)
-        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = RMSNorm(
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
             config.hidden_size, eps=config.rms_norm_eps
         )
         self.config = config
@@ -175,7 +156,7 @@ class Llama(nn.Module):
         self.layers = [
             TransformerBlock(config=config) for _ in range(config.num_hidden_layers)
         ]
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
     def __call__(
         self,

llava/requirements.txt

@@ -1,4 +1,4 @@
-mlx>=0.5.0
+mlx>=0.8.0
 numpy
 transformers
 torch