# Copyright © 2024 Apple Inc. # mlx-lm architecture module for Nanbeige4.2 (looped / recurrent-depth transformer). # # Nanbeige4.2 is Llama-style (GQA attention, SwiGLU MLP, RMSNorm, rotary embeddings) # with one twist: the full decoder stack is executed `num_loops` times. Each loop pass # keeps its own KV-cache slice, and (when skip_loop_final_norm is False) the final # RMSNorm is applied at the end of every loop pass — the normalized output of one loop # feeds the next loop as input. 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 .cache import KVCache from .rope_utils import initialize_rope @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 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 num_loops: int = 1 skip_loop_final_norm: bool = False def __post_init__(self): if self.num_key_value_heads is None: self.num_key_value_heads = self.num_attention_heads if self.num_loops < 1: self.num_loops = 1 class Attention(nn.Module): def __init__(self, args: ModelArgs): super().__init__() dim = args.hidden_size self.n_heads = args.num_attention_heads self.n_kv_heads = args.num_key_value_heads self.head_dim = head_dim = args.head_dim or (dim // self.n_heads) self.scale = head_dim**-0.5 self.q_proj = nn.Linear(dim, self.n_heads * head_dim, bias=args.attention_bias) self.k_proj = nn.Linear(dim, self.n_kv_heads * head_dim, bias=args.attention_bias) self.v_proj = nn.Linear(dim, self.n_kv_heads * head_dim, bias=args.attention_bias) self.o_proj = nn.Linear(self.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=None, cache=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) 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 MLP(nn.Module): def __init__(self, args: ModelArgs): super().__init__() dim, hidden = args.hidden_size, args.intermediate_size self.gate_proj = nn.Linear(dim, hidden, bias=args.mlp_bias) self.down_proj = nn.Linear(hidden, dim, bias=args.mlp_bias) self.up_proj = nn.Linear(dim, hidden, bias=args.mlp_bias) def __call__(self, x) -> mx.array: 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.self_attn = Attention(args) self.mlp = MLP(args) self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps) self.post_attention_layernorm = nn.RMSNorm( args.hidden_size, eps=args.rms_norm_eps ) def __call__(self, x: mx.array, mask=None, cache=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 NanbeigeModel(nn.Module): def __init__(self, args: ModelArgs): super().__init__() self.args = args self.num_hidden_layers = args.num_hidden_layers self.num_loops = args.num_loops self.skip_loop_final_norm = args.skip_loop_final_norm assert args.vocab_size > 0 self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size) self.layers = [TransformerBlock(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, input_embeddings=None): if input_embeddings is not None: h = input_embeddings else: h = self.embed_tokens(inputs) n = self.num_hidden_layers if cache is None: cache = [None] * (n * self.num_loops) # Each loop pass replays the whole stack against its own cache slice. for loop_idx in range(self.num_loops): loop_cache = cache[loop_idx * n : (loop_idx + 1) * n] mask = create_attention_mask(h, loop_cache[0]) for layer, c in zip(self.layers, loop_cache): h = layer(h, mask, cache=c) if not self.skip_loop_final_norm: h = self.norm(h) if self.skip_loop_final_norm: h = self.norm(h) return h class Model(nn.Module): def __init__(self, args: ModelArgs): super().__init__() self.args = args self.model_type = args.model_type self.model = NanbeigeModel(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, input_embeddings=None): out = self.model(inputs, cache, input_embeddings) 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): # Drop non-persistent rotary buffers if present in a checkpoint. weights = { k: v for k, v in weights.items() if "rotary_emb.inv_freq" not in k and ".rope." not in k } if self.args.tie_word_embeddings: weights.pop("lm_head.weight", None) return weights @property def layers(self): return self.model.layers def make_cache(self): # One KV cache per (loop, layer): the stack is executed num_loops times # and each pass must not see the other passes' keys/values. return [ KVCache() for _ in range(self.args.num_hidden_layers * self.args.num_loops) ]