# Copyright © 2026 Apple Inc. import math from dataclasses import dataclass from typing import Any, Dict, Optional import mlx.core as mx import mlx.nn as nn from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients from .activations import swiglu from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention from .mla import MultiLinear from .pipeline import PipelineMixin from .rope_utils import initialize_rope from .switch_layers import SwitchGLU @dataclass class ModelArgs(BaseModelArgs): model_type: str = "glm4_moe_lite" vocab_size: int = 154880 hidden_size: int = 2048 intermediate_size: int = 10240 moe_intermediate_size: int = 1536 num_hidden_layers: int = 47 num_attention_heads: int = 20 num_key_value_heads: int = 20 n_shared_experts: Optional[int] = 1 n_routed_experts: Optional[int] = 64 routed_scaling_factor: float = 1.8 kv_lora_rank: int = 512 q_lora_rank: int = 768 qk_rope_head_dim: int = 64 qk_nope_head_dim: int = 192 v_head_dim: int = 256 topk_method: str = "noaux_tc" scoring_func: str = "sigmoid" norm_topk_prob: bool = True n_group: int = 1 topk_group: int = 1 num_experts_per_tok: int = 4 moe_layer_freq: int = 1 first_k_dense_replace: int = 1 max_position_embeddings: int = 202752 rms_norm_eps: float = 1e-5 rope_theta: float = 1_000_000.0 rope_scaling: Optional[Dict] = None attention_bias: bool = False attention_dropout: float = 0.0 partial_rotary_factor: float = 1.0 tie_word_embeddings: bool = False num_nextn_predict_layers: int = 1 quantization: Optional[Dict[str, Any]] = None class Glm4MoeLiteAttention(nn.Module): def __init__(self, config: ModelArgs): super().__init__() self.config = config self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.max_position_embeddings = config.max_position_embeddings rope_params = config.rope_scaling self.rope_theta = config.rope_theta self.q_lora_rank = config.q_lora_rank self.qk_rope_head_dim = config.qk_rope_head_dim self.kv_lora_rank = config.kv_lora_rank self.v_head_dim = config.v_head_dim self.qk_nope_head_dim = config.qk_nope_head_dim self.q_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim self.scale = self.q_head_dim**-0.5 if self.q_lora_rank is None: self.q_proj = nn.Linear( self.hidden_size, self.num_heads * self.q_head_dim, bias=False ) else: self.q_a_proj = nn.Linear( self.hidden_size, self.q_lora_rank, bias=config.attention_bias ) self.q_a_layernorm = nn.RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) self.q_b_proj = nn.Linear( self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False ) self.kv_a_proj_with_mqa = nn.Linear( self.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=config.attention_bias, ) self.kv_a_layernorm = nn.RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) head_dim = self.qk_nope_head_dim + self.v_head_dim self.embed_q = MultiLinear( self.qk_nope_head_dim, self.kv_lora_rank, self.num_heads ) self.unembed_out = MultiLinear( self.kv_lora_rank, self.v_head_dim, self.num_heads ) self.o_proj = nn.Linear( self.num_heads * self.v_head_dim, self.hidden_size, bias=config.attention_bias, ) if rope_params is not None: mscale_all_dim = rope_params.get("mscale_all_dim", 0) if mscale_all_dim: scaling_factor = rope_params["factor"] if scaling_factor > 1: s = 0.1 * mscale_all_dim * math.log(scaling_factor) + 1.0 self.scale = self.scale * s * s self.rope = initialize_rope( dims=self.qk_rope_head_dim, base=self.rope_theta, traditional=True, max_position_embeddings=self.max_position_embeddings, scaling_config=rope_params, ) def __call__( self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None, ) -> mx.array: B, L, D = x.shape if self.q_lora_rank is None: q = self.q_proj(x) else: q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(x))) q = q.reshape(B, L, self.num_heads, self.q_head_dim).transpose(0, 2, 1, 3) q_nope, q_pe = mx.split(q, [self.qk_nope_head_dim], axis=-1) compressed_kv = self.kv_a_proj_with_mqa(x) compressed_kv, k_pe = mx.split(compressed_kv, [self.kv_lora_rank], axis=-1) k_pe = k_pe.reshape(B, L, 1, self.qk_rope_head_dim).transpose(0, 2, 1, 3) kv_latent = self.kv_a_layernorm(compressed_kv) offset = cache.offset if cache is not None else 0 q_pe = self.rope(q_pe, offset) k_pe = self.rope(k_pe, offset) kv_latent = mx.expand_dims(kv_latent, axis=1) if cache is not None: kv_latent, k_pe = cache.update_and_fetch(kv_latent, k_pe) pe_scores = (q_pe * self.scale) @ k_pe.swapaxes(-1, -2) if mask is not None: pe_scores = mx.where( mask, pe_scores, mx.array(mx.finfo(pe_scores.dtype).min, pe_scores.dtype), ) if L == 1: q_nope = self.embed_q(q_nope) k = v = kv_latent else: k = self.embed_q(kv_latent, transpose=False) v = self.unembed_out(kv_latent) output = scaled_dot_product_attention( q_nope, k, v, cache=cache, scale=self.scale, mask=pe_scores ) if L == 1: output = self.unembed_out(output) output = output.transpose(0, 2, 1, 3).reshape(B, L, -1) return self.o_proj(output) class Glm4MoeLiteMLP(nn.Module): def __init__( self, config: ModelArgs, hidden_size: int = None, intermediate_size: int = None ): super().__init__() self.config = config self.hidden_size = config.hidden_size if hidden_size is None else hidden_size self.intermediate_size = ( config.intermediate_size if intermediate_size is None else intermediate_size ) self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) def __call__(self, x): down_proj = self.down_proj(swiglu(self.gate_proj(x), self.up_proj(x))) return down_proj @mx.compile def group_expert_select( gates, e_score_correction_bias, top_k, n_group, topk_group, routed_scaling_factor, norm_topk_prob, ): scores = mx.sigmoid(gates.astype(mx.float32)) orig_scores = scores scores = scores + e_score_correction_bias if n_group > 1: scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1)) group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True) k = n_group - topk_group group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :] scores = mx.put_along_axis( scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2 ) scores = mx.flatten(scores, -2, -1) k = top_k inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k] scores = mx.take_along_axis(orig_scores, inds, axis=-1) if top_k > 1 and norm_topk_prob: denominator = scores.sum(axis=-1, keepdims=True) scores = scores / (denominator + 1e-20) scores = scores * routed_scaling_factor return inds, scores class MoEGate(nn.Module): def __init__(self, config: ModelArgs): super().__init__() self.config = config self.top_k = config.num_experts_per_tok self.norm_topk_prob = config.norm_topk_prob self.n_routed_experts = config.n_routed_experts self.routed_scaling_factor = config.routed_scaling_factor self.n_group = config.n_group self.topk_group = config.topk_group self.weight = mx.zeros((self.n_routed_experts, config.hidden_size)) self.e_score_correction_bias = mx.zeros((self.n_routed_experts,)) assert config.topk_method == "noaux_tc", "Unsupported topk method." def __call__(self, x): return group_expert_select( x @ self.weight.T, self.e_score_correction_bias, self.top_k, self.n_group, self.topk_group, self.routed_scaling_factor, self.norm_topk_prob, ) class Glm4MoeLiteMoE(nn.Module): def __init__(self, config: ModelArgs): super().__init__() self.config = config self.num_experts_per_tok = config.num_experts_per_tok self.switch_mlp = SwitchGLU( config.hidden_size, config.moe_intermediate_size, config.n_routed_experts, ) self.gate = MoEGate(config) if config.n_shared_experts is not None: intermediate_size = config.moe_intermediate_size * config.n_shared_experts self.shared_experts = Glm4MoeLiteMLP( config=config, intermediate_size=intermediate_size ) self.sharding_group = None def __call__(self, x): if self.sharding_group is not None: x = sum_gradients(self.sharding_group)(x) inds, scores = self.gate(x) y = self.switch_mlp(x, inds) y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype) if self.config.n_shared_experts is not None: y = y + self.shared_experts(x) if self.sharding_group is not None: y = mx.distributed.all_sum(y, group=self.sharding_group) return y class Glm4MoeLiteDecoderLayer(nn.Module): def __init__(self, config: ModelArgs, layer_idx: int): super().__init__() self.self_attn = Glm4MoeLiteAttention(config) use_moe = ( config.n_routed_experts is not None and layer_idx >= config.first_k_dense_replace and layer_idx % config.moe_layer_freq == 0 ) self.mlp = Glm4MoeLiteMoE(config) if use_moe else Glm4MoeLiteMLP(config) 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 ) def __call__( self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None, ) -> mx.array: r = self.self_attn(self.input_layernorm(x), mask, cache) h = x + r r = self.mlp(self.post_attention_layernorm(h)) return h + r class Glm4MoeLiteModel(PipelineMixin, nn.Module): def __init__(self, config: ModelArgs): super().__init__() self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size) self.layers = [ Glm4MoeLiteDecoderLayer(config, idx) for idx in range(config.num_hidden_layers) ] self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps) def __call__( self, x: mx.array, cache: Optional[Any] = None, ) -> mx.array: h = self.embed_tokens(x) pipeline_rank = self.pipeline_rank pipeline_size = self.pipeline_size if cache is None: cache = [None] * len(self.pipeline_layers) mask = create_attention_mask(h, cache[0], return_array=True) # Receive from the previous process in the pipeline if pipeline_rank < pipeline_size - 1: h = mx.distributed.recv_like(h, (pipeline_rank + 1)) for l, c in zip(self.pipeline_layers, cache): h = l(h, mask, cache=c) # Send to the next process in the pipeline if pipeline_rank != 0: h = mx.distributed.send(h, (pipeline_rank - 1) % pipeline_size) if cache[-1] is not None: cache[-1].keys = mx.depends(cache[-1].keys, h) # Broadcast h while keeping it in the graph if pipeline_size > 1: h = mx.distributed.all_gather(h)[: h.shape[0]] return self.norm(h) class Model(nn.Module): def __init__(self, config: ModelArgs): super().__init__() self.args = config self.model_type = config.model_type self.model = Glm4MoeLiteModel(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) def __call__( self, inputs: mx.array, cache: Optional[Any] = None, ): out = self.model(inputs, cache) return self.lm_head(out) def sanitize(self, weights): def is_mpt_layer(key): subkeys = key.split(".") if len(subkeys) < 3: return False if ( subkeys[1] == "layers" and int(subkeys[2]) >= self.args.num_hidden_layers ): return True return False new_weights = {} for k, v in weights.items(): if is_mpt_layer(k): continue else: new_weights[k] = v weights = new_weights # Stack experts for l in range(self.args.num_hidden_layers): prefix = f"model.layers.{l}" for n, m in [("w1", "gate_proj"), ("w2", "down_proj"), ("w3", "up_proj")]: for k in ["weight", "scales", "biases"]: if f"{prefix}.mlp.experts.0.{m}.{k}" in weights: to_join = [ weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}") for e in range(self.args.n_routed_experts) ] weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join) prefix = f"model.layers.{l}.self_attn" if f"{prefix}.kv_b_proj.weight" in weights: layer = self.layers[l].self_attn.embed_q quantized = f"{prefix}.kv_b_proj.scales" in weights v = weights.pop(f"{prefix}.kv_b_proj.weight") head_dim = self.args.qk_nope_head_dim + self.args.v_head_dim if quantized: dims = self.args.kv_lora_rank scales = weights.pop(f"{prefix}.kv_b_proj.scales") biases = weights.pop(f"{prefix}.kv_b_proj.biases") # Try to infer bits and group size bits = (v.shape[-1] * 32) // dims group_size = dims // scales.shape[-1] v = mx.dequantize( v, scales, biases, bits=bits, group_size=group_size ) num_heads = self.args.num_attention_heads v = v.reshape(num_heads, head_dim, -1) wk = mx.contiguous( v[:, : self.args.qk_nope_head_dim, :].swapaxes(-1, -2) ) wv = mx.contiguous(v[:, self.args.qk_nope_head_dim :, :]) if quantized: wk, wk_scales, wk_biases = mx.quantize( wk, bits=bits, group_size=group_size ) wv, wv_scales, wv_biases = mx.quantize( wv, bits=bits, group_size=group_size ) weights[f"{prefix}.embed_q.scales"] = wk_scales weights[f"{prefix}.unembed_out.scales"] = wv_scales weights[f"{prefix}.embed_q.biases"] = wk_biases weights[f"{prefix}.unembed_out.biases"] = wv_biases weights[f"{prefix}.embed_q.weight"] = wk weights[f"{prefix}.unembed_out.weight"] = wv return weights def shard(self, group: Optional[mx.distributed.Group] = None): group = group or mx.distributed.init() rank = group.rank() N = group.size() for layer in self.model.layers: # Shard the self attention if layer.self_attn.q_lora_rank is None: layer.self_attn.q_proj = shard_linear( layer.self_attn.q_proj, "all-to-sharded", group=group ) else: layer.self_attn.q_b_proj = shard_linear( layer.self_attn.q_b_proj, "all-to-sharded", group=group ) layer.self_attn.num_heads //= N num_heads = layer.self_attn.num_heads sh = rank * num_heads eh = sh + num_heads def shard_heads(w): return w[sh:eh] layer.self_attn.embed_q.apply(shard_heads) layer.self_attn.unembed_out.apply(shard_heads) layer.self_attn.o_proj = shard_linear( layer.self_attn.o_proj, "sharded-to-all", group=group ) # Shard the MLP if isinstance(layer.mlp, Glm4MoeLiteMLP): layer.mlp.gate_proj = shard_linear( layer.mlp.gate_proj, "all-to-sharded", group=group ) layer.mlp.down_proj = shard_linear( layer.mlp.down_proj, "sharded-to-all", group=group ) layer.mlp.up_proj = shard_linear( layer.mlp.up_proj, "all-to-sharded", group=group ) # Shard the MoE. Shard in place since the MoE should be responsible # for aggregating the results. else: layer.mlp.sharding_group = group if getattr(layer.mlp, "shared_experts", None) is not None: shard_inplace( layer.mlp.shared_experts.gate_proj, "all-to-sharded", group=group, ) shard_inplace( layer.mlp.shared_experts.down_proj, "sharded-to-all", group=group, ) shard_inplace( layer.mlp.shared_experts.up_proj, "all-to-sharded", group=group, ) shard_inplace( layer.mlp.switch_mlp.gate_proj, "all-to-sharded", group=group ) shard_inplace( layer.mlp.switch_mlp.down_proj, "sharded-to-all", group=group ) shard_inplace( layer.mlp.switch_mlp.up_proj, "all-to-sharded", group=group ) @property def layers(self): return self.model.pipeline_layers @property def cast_predicate(self): def predicate(k): return "e_score_correction_bias" not in k return predicate