[Cohere] Add cohere2_moe model support

mlx_lm/models/cohere2_moe.py

@@ -0,0 +1,350 @@
+# Copyright © 2026 Apple Inc.
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .cache import KVCache, RotatingKVCache
+from .switch_layers import SwitchGLU
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int = 1024
+    head_dim: int = 128
+    num_hidden_layers: int = 36
+    intermediate_size: int = 1024
+    num_attention_heads: int = 64
+    num_key_value_heads: int = 8
+    rope_theta: float = 50000.0
+    vocab_size: int = 256000
+    layer_norm_eps: float = 1e-05
+    logit_scale: float = 0.0625
+    attention_bias: bool = False
+    layer_norm_bias: bool = False
+    sliding_window: int = 4096
+    sliding_window_pattern: int = 4
+    num_experts: int = 128
+    num_experts_per_tok: int = 8
+    norm_topk_prob: bool = True
+    num_shared_experts: Optional[int] = None
+    moe_num_shared_experts: int = 4
+    moe_gate_act: str = "sigmoid"
+    expert_selection_fn: Optional[str] = None
+    shared_expert_combination_strategy: str = "average"
+    rms_norm_eps: Optional[float] = None
+    first_k_dense_replace: int = 0
+    prefix_dense_intermediate_size: Optional[int] = None
+    prefix_dense_sliding_window_pattern: int = 1
+    layer_types: Optional[List[str]] = None
+
+    def __post_init__(self):
+        if self.num_shared_experts is not None:
+            self.moe_num_shared_experts = self.num_shared_experts
+        if self.expert_selection_fn is not None:
+            self.moe_gate_act = self.expert_selection_fn
+        if self.prefix_dense_intermediate_size is None:
+            self.prefix_dense_intermediate_size = self.intermediate_size
+
+
+def is_prefix_dense_layer(args: ModelArgs, layer_idx: int):
+    return layer_idx < args.first_k_dense_replace
+
+
+def is_sliding_layer(args: ModelArgs, layer_idx: int):
+    if is_prefix_dense_layer(args, layer_idx):
+        return False
+    if args.layer_types is not None:
+        return args.layer_types[layer_idx] == "sliding_attention"
+    return (layer_idx + 1) % args.sliding_window_pattern != 0
+
+
+def norm_layer(args: ModelArgs):
+    if args.rms_norm_eps is not None:
+        return nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+    return nn.LayerNorm(
+        args.hidden_size, eps=args.layer_norm_eps, bias=args.layer_norm_bias
+    )
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.args = args
+        self.layer_idx = layer_idx
+
+        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
+        self.head_dim = head_dim = args.head_dim
+        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)
+
+        self.use_sliding_window = is_sliding_layer(args, layer_idx)
+        self.force_rope = (
+            is_prefix_dense_layer(args, layer_idx)
+            and args.prefix_dense_sliding_window_pattern == 1
+        )
+
+    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)
+
+        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)
+
+        # Cohere2Moe applies RoPE to sliding layers and optionally to prefix
+        # dense full-attention layers.
+        if self.use_sliding_window or self.force_rope:
+            if cache is None:
+                queries = self.rope(queries)
+                keys = self.rope(keys)
+            else:
+                queries = self.rope(queries, offset=cache.offset)
+                keys = self.rope(keys, offset=cache.offset)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        sdpa_type = mx.float32 if queries.dtype == mx.float16 else queries.dtype
+        output = scaled_dot_product_attention(
+            queries.astype(sdpa_type),
+            keys,
+            values,
+            cache=cache,
+            scale=self.scale,
+            mask=mask,
+        ).astype(queries.dtype)
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+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 CohereMoeSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        intermediate_size = args.intermediate_size
+
+        self.num_experts = num_experts = args.num_experts
+        self.top_k = args.num_experts_per_tok
+        self.norm_topk_prob = args.norm_topk_prob
+
+        self.gate = nn.Linear(dim, num_experts, bias=False)
+        self.switch_mlp = SwitchGLU(dim, intermediate_size, num_experts)
+
+        if getattr(args, "moe_num_shared_experts", 0) > 0:
+            shared_intermediate_size = (
+                args.intermediate_size * args.moe_num_shared_experts
+            )
+            self.shared_experts = MLP(
+                args.hidden_size, shared_intermediate_size,
+            )
+            self.shared_expert_combination_strategy = \
+                args.shared_expert_combination_strategy
+            assert self.shared_expert_combination_strategy in [
+                "average", "sum"
+            ], "shared_expert_combination_strategy "
+            "must be one of ['average', 'sum']"
+        else:
+            self.shared_experts = None
+            self.shared_expert_combination_strategy = None
+
+        if args.moe_gate_act == "softmax":
+            self.gate_act = nn.Softmax()
+        elif args.moe_gate_act == "sigmoid":
+            self.gate_act = nn.Sigmoid()
+        else:
+            raise ValueError(f"{args.moe_gate_act} is not supported.")
+
+    def __call__(
+        self,
+        x: mx.array,
+    ):
+        gates = self.gate(x)
+        gates = self.gate_act(gates.astype(mx.float32))
+
+        k = self.top_k
+        inds = mx.stop_gradient(mx.argpartition(-gates, kth=k - 1, axis=-1)[..., :k])
+        scores = mx.take_along_axis(gates, inds, axis=-1)
+        if self.norm_topk_prob:
+            scores = scores / mx.maximum(scores.sum(axis=-1, keepdims=True), 1e-12)
+
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+
+        if self.shared_experts is not None:
+            if self.shared_expert_combination_strategy == "average":
+                y = (y + self.shared_experts(x)) / 2
+            else:
+                y = y + self.shared_experts(x)
+
+        return y
+
+
+class CohereMoEDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs, layer_idx: int):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.n_heads = args.num_attention_heads
+
+        self.self_attn = Attention(args, layer_idx)
+        self.mlp = (
+            MLP(args.hidden_size, args.prefix_dense_intermediate_size)
+            if is_prefix_dense_layer(args, layer_idx)
+            else CohereMoeSparseMoeBlock(args)
+        )
+        self.input_layernorm = norm_layer(args)
+        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 = self.self_attn(h, mask, cache)
+        ff_h = self.mlp(h)
+
+        return attn_h + ff_h + x
+
+
+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.window_size = args.sliding_window
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            CohereMoEDecoderLayer(args=args, layer_idx=i)
+            for i in range(args.num_hidden_layers)
+        ]
+        self.norm = norm_layer(args)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for layer, c in zip(self.layers, cache):
+            mask = create_attention_mask(
+                h,
+                c,
+                window_size=(
+                    self.window_size if layer.self_attn.use_sliding_window else None
+                ),
+            )
+
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.model_type = args.model_type
+        self.model = CohereModel(args)
+        self.args = args
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out = self.model(inputs, cache)
+        out = self.model.embed_tokens.as_linear(out)
+        out = out * self.model.args.logit_scale
+        return out
+
+    def make_cache(self):
+        caches = []
+        for i in range(self.args.num_hidden_layers):
+            if is_sliding_layer(self.args, i):
+                caches.append(
+                    RotatingKVCache(max_size=self.args.sliding_window, keep=0)
+                )
+            else:
+                caches.append(KVCache())
+        return caches
+
+    def sanitize(self, weights):
+        for l in range(self.args.num_hidden_layers):
+            if is_prefix_dense_layer(self.args, l):
+                continue
+            prefix = f"model.layers.{l}"
+            for n in ["up_proj", "down_proj", "gate_proj"]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{n}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{n}.{k}")
+                            for e in range(self.args.num_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{n}.{k}"] = mx.stack(to_join)
+
+        for key in list(weights.keys()):
+            if "rotary_emb.inv_freq" in key:
+                weights.pop(key)
+            elif key.endswith(".bias"):
+                if ".mlp." in key:
+                    weights.pop(key)
+                elif ".self_attn." in key and not self.args.attention_bias:
+                    weights.pop(key)
+                elif "layernorm" in key.lower() and not self.args.layer_norm_bias:
+                    weights.pop(key)
+
+        return weights
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, module):
+            if ".self_attn." in path:
+                return False
+            if ".mlp.gate" in path and "gate_proj" not in path:
+                return False
+            return True
+
+        return predicate
+
+    @property
+    def layers(self):
+        return self.model.layers