Add support for Cohere's Command-R (#565)

* initial commit for command-R

* update mlp, layernorm, lm_head and model args

* add custom layernorm

* add default to tie_word_embeddings

* add layernorm weight type and refactor

* update layernorm (bias conditional) in model/layers

* fix layer norm use traditional rope

* add test

---------

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

llms/mlx_lm/models/cohere.py

@@ -0,0 +1,171 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .layers import LayerNorm
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int = 8192
+    num_hidden_layers: int = 40
+    intermediate_size: int = 22528
+    num_attention_heads: int = 64
+    num_key_value_heads: int = 64
+    rope_theta: float = 8000000.0
+    vocab_size: int = 256000
+    layer_norm_eps: float = 1e-05
+    logit_scale: float = 0.0625
+    attention_bias: bool = False
+    layer_norm_bias: bool = False
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.hidden_size // args.num_attention_heads
+        self.scale = head_dim**-0.5
+
+        attetion_bias = args.attention_bias
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attetion_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attetion_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attetion_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attetion_bias)
+
+        self.rope = nn.RoPE(head_dim, traditional=True, base=args.rope_theta)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        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 cache is not None:
+            key_cache, value_cache = cache
+            queries = self.rope(queries, offset=key_cache.shape[2])
+            keys = self.rope(keys, offset=key_cache.shape[2])
+            keys = mx.concatenate([key_cache, keys], axis=2)
+            values = mx.concatenate([value_cache, values], axis=2)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        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)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+
+    def __call__(self, x):
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.n_heads = args.num_attention_heads
+
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        self.input_layernorm = LayerNorm(
+            args.hidden_size, eps=args.layer_norm_eps, bias=args.layer_norm_bias
+        )
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        h = self.input_layernorm(x)
+        attn_h, cache = self.self_attn(h, mask, cache)
+        ff_h = self.mlp(h)
+        return attn_h + ff_h + x, cache
+
+
+class CohereModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = LayerNorm(
+            args.hidden_size, eps=args.layer_norm_eps, bias=args.layer_norm_bias
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        mask = None
+        if h.shape[1] > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(h.shape[1])
+            mask = mask.astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for e, layer in enumerate(self.layers):
+            h, cache[e] = layer(h, mask, cache[e])
+
+        return self.norm(h), cache
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.model_type = args.model_type
+        self.model = CohereModel(args)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out, cache = self.model(inputs, cache)
+        out = out @ self.model.embed_tokens.weight.T
+        out = out * self.model.args.logit_scale
+        return out, cache
+
+    @property
+    def layers(self):
+        return self.model.layers

llms/mlx_lm/models/layers.py

@@ -23,29 +23,58 @@ class RMSNorm(nn.Module):
 
 @partial(mx.compile, shapeless=True)
 def ln_norm(x, eps, weight=None, bias=None):
+    """
+    Layer normalization for input tensor x.
+
+    Args:
+        x (np.ndarray): Input tensor.
+        eps (float, optional): Small value to avoid division by zero.
+        weight (np.ndarray, optional): Weight tensor for normalization.
+        bias (np.ndarray, optional): Bias tensor for normalization.
+
+    Returns:
+        np.ndarray: Normalized tensor.
+    """
     t = x.dtype
     x = x.astype(mx.float32)
+
+    # Compute mean and variance along the last dimension
     means = mx.mean(x, axis=-1, keepdims=True)
     var = mx.var(x, axis=-1, keepdims=True)
+
+    # Normalize the input tensor
     x = (x - means) * mx.rsqrt(var + eps)
     x = x.astype(t)
-    return weight * x + bias if weight is not None else x
+
+    # Apply weight and bias if provided
+    if weight is not None:
+        x = x * weight
+    if bias is not None:
+        x = x + bias
+    return x
 
 
 class LayerNorm(nn.Module):
-    def __init__(self, dims: int, eps: float = 1e-5, affine: bool = True):
+    def __init__(
+        self, dims: int, eps: float = 1e-5, affine: bool = True, bias: bool = True
+    ):
         super().__init__()
-        if affine:
-            self.bias = mx.zeros((dims,))
-            self.weight = mx.ones((dims,))
         self.eps = eps
         self.dims = dims
+        self.affine = affine
+
+        if affine:
+            self.weight = mx.ones((dims,))
+            self.bias = mx.zeros((dims,)) if bias else None
 
     def _extra_repr(self):
         return f"{self.dims}, eps={self.eps}, affine={'weight' in self}"
 
     def __call__(self, x: mx.array) -> mx.array:
-        if "weight" in self:
-            return ln_norm(x, self.eps, self.weight, self.bias)
+        if self.affine:
+            if self.bias is not None:
+                return ln_norm(x, self.eps, self.weight, self.bias)
+            else:
+                return ln_norm(x, self.eps, self.weight)
         else:
             return ln_norm(x, self.eps)

llms/tests/test_models.py

@@ -254,7 +254,6 @@ class TestModels(unittest.TestCase):
         self.assertEqual(sanitized_weights["lm_head.weight"], "some_value")
 
     def test_starcoder2_tie_word_embeddings_with_lm_head_weight(self):
-
         from mlx_lm.models import starcoder2
 
         args = starcoder2.ModelArgs(
@@ -276,6 +275,17 @@ class TestModels(unittest.TestCase):
         self.assertIn("lm_head.weight", sanitized_weights)
         self.assertEqual(sanitized_weights["lm_head.weight"], "existing_value")
 
+    def test_cohere(self):
+        from mlx_lm.models import cohere
+
+        args = cohere.ModelArgs(
+            model_type="cohere",
+        )
+        model = cohere.Model(args)
+        self.model_test_runner(
+            model, args.model_type, args.vocab_size, args.num_hidden_layers
+        )
+
 
 if __name__ == "__main__":
     unittest.main()