# Copyright © 2026 # MLX port of allenai/Emo (EmoForCausalLM). # Architecture: OLMoE-like — pre-norm decoder, MHA with RoPE, sparse MoE FFN. # Differences from OLMoE: # - no q_norm / k_norm # - shared experts: the last `num_shared_experts` experts of the gate are # routed via a separate softmax + top-k (their topk is num_shared_experts) # and concatenated to the standard top-k experts. Indices are offset by # (num_experts - num_shared_experts). # - layernorm names: pre_attention_layernorm / pre_feedforward_layernorm # (vs input_layernorm / post_attention_layernorm in OLMoE). from dataclasses import dataclass from typing import Any, Dict, Optional, Union import mlx.core as mx import mlx.nn as nn from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention from .rope_utils import initialize_rope from .switch_layers import SwitchGLU @dataclass class ModelArgs(BaseModelArgs): model_type: str hidden_size: int num_hidden_layers: int intermediate_size: int num_attention_heads: int rms_norm_eps: float vocab_size: int num_experts: int num_experts_per_tok: int num_shared_experts: int = 0 norm_topk_prob: bool = False head_dim: Optional[int] = None max_position_embeddings: Optional[int] = None num_key_value_heads: Optional[int] = None attention_bias: bool = False mlp_bias: bool = False rope_theta: float = 10000.0 rope_traditional: bool = False rope_scaling: Optional[Dict[str, Union[float, str]]] = None tie_word_embeddings: bool = False def __post_init__(self): if self.num_key_value_heads is None: self.num_key_value_heads = self.num_attention_heads class Attention(nn.Module): def __init__(self, args: ModelArgs): super().__init__() 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 or args.hidden_size // n_heads self.scale = head_dim**-0.5 self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias) self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias) self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias) self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=args.attention_bias) self.rope = initialize_rope( self.head_dim, args.rope_theta, args.rope_traditional, args.rope_scaling, args.max_position_embeddings, ) def __call__( self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None, ) -> mx.array: B, L, _ = 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) if cache is not None: queries = self.rope(queries, offset=cache.offset) keys = self.rope(keys, offset=cache.offset) keys, values = cache.update_and_fetch(keys, values) else: queries = self.rope(queries) keys = self.rope(keys) output = scaled_dot_product_attention( queries, keys, values, cache=cache, scale=self.scale, mask=mask ) output = output.transpose(0, 2, 1, 3).reshape(B, L, -1) return self.o_proj(output) class EmoSparseMoeBlock(nn.Module): def __init__(self, args: ModelArgs): super().__init__() self.num_experts = args.num_experts self.num_shared = args.num_shared_experts self.top_k = args.num_experts_per_tok self.norm_topk_prob = args.norm_topk_prob # Number of "routed" (non-shared) experts and how many to pick from them. self.num_standard = self.num_experts - self.num_shared self.routed_top_k = self.top_k - self.num_shared self.gate = nn.Linear(args.hidden_size, self.num_experts, bias=False) self.switch_mlp = SwitchGLU( args.hidden_size, args.intermediate_size, self.num_experts, bias=args.mlp_bias, ) def __call__(self, x: mx.array) -> mx.array: B, L, D = x.shape x_flat = x.reshape(-1, D) router_logits = self.gate(x_flat) # (T, num_experts) if self.num_shared > 0: # Split routing: shared = last num_shared cols, standard = the rest. std_logits = router_logits[..., : self.num_standard] sh_logits = router_logits[..., self.num_standard :] std_weights = mx.softmax(std_logits, axis=-1, precise=True) sh_weights = mx.softmax(sh_logits, axis=-1, precise=True) k = self.routed_top_k std_indices = mx.stop_gradient( mx.argpartition(-std_weights, kth=k - 1, axis=-1)[..., :k] ) std_scores = mx.take_along_axis(std_weights, std_indices, axis=-1) ks = self.num_shared if ks == 1: sh_indices = mx.argmax(sh_weights, axis=-1, keepdims=True) else: sh_indices = mx.stop_gradient( mx.argpartition(-sh_weights, kth=ks - 1, axis=-1)[..., :ks] ) sh_scores = mx.take_along_axis(sh_weights, sh_indices, axis=-1) # Order matches the HF impl: [standard_indices, shared_indices + offset] indices = mx.concatenate( [std_indices, sh_indices + self.num_standard], axis=-1 ) scores = mx.concatenate([std_scores, sh_scores], axis=-1) else: weights = mx.softmax(router_logits, axis=-1, precise=True) k = self.top_k indices = mx.stop_gradient( mx.argpartition(-weights, kth=k - 1, axis=-1)[..., :k] ) scores = mx.take_along_axis(weights, indices, axis=-1) if self.norm_topk_prob: scores = scores / scores.sum(axis=-1, keepdims=True) y = self.switch_mlp(x_flat, indices) y = (y * scores[..., None]).sum(axis=-2) return y.reshape(B, L, D) class TransformerBlock(nn.Module): def __init__(self, args: ModelArgs): super().__init__() self.self_attn = Attention(args) self.mlp = EmoSparseMoeBlock(args) self.pre_attention_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps) self.pre_feedforward_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps) def __call__( self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None, ) -> mx.array: x = x + self.self_attn(self.pre_attention_layernorm(x), mask, cache) x = x + self.mlp(self.pre_feedforward_layernorm(x)) return x class EmoModel(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 = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps) def __call__(self, inputs: mx.array, cache=None): h = self.embed_tokens(inputs) if cache is None: cache = [None] * len(self.layers) mask = create_attention_mask(h, cache[0]) for layer, c in zip(self.layers, cache): h = layer(h, mask, cache=c) return self.norm(h) class Model(nn.Module): def __init__(self, args: ModelArgs): super().__init__() self.args = args self.model_type = args.model_type self.model = EmoModel(args) if not args.tie_word_embeddings: self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False) def __call__(self, inputs: mx.array, cache=None): out = self.model(inputs, cache) if self.args.tie_word_embeddings: out = self.model.embed_tokens.as_linear(out) else: out = self.lm_head(out) return out def sanitize(self, weights): # Pack per-expert linears into the SwitchGLU stacked tensors. if "model.layers.0.mlp.experts.0.up_proj.weight" not in weights: return weights for l in range(self.args.num_hidden_layers): 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) return weights @property def layers(self): return self.model.layers