# Copyright © 2025 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 mlx_lm.models.activations import swiglu from mlx_lm.models.base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention from mlx_lm.models.cache import CacheList, KVCache from mlx_lm.models.mla import MultiLinear from mlx_lm.models.rope_utils import initialize_rope from mlx_lm.models.switch_layers import SwitchGLU @dataclass class ModelArgs(BaseModelArgs): model_type: str vocab_size: int hidden_size: int index_head_dim: int index_n_heads: int index_topk: int intermediate_size: int moe_intermediate_size: int num_hidden_layers: int num_attention_heads: int num_key_value_heads: int n_shared_experts: Optional[int] n_routed_experts: Optional[int] routed_scaling_factor: float kv_lora_rank: int q_lora_rank: int qk_rope_head_dim: int v_head_dim: int qk_nope_head_dim: int topk_method: str scoring_func: str norm_topk_prob: bool n_group: int topk_group: int num_experts_per_tok: int moe_layer_freq: int first_k_dense_replace: int max_position_embeddings: int rms_norm_eps: float rope_parameters: Dict attention_bias: bool rope_scaling: Dict = None rope_theta: Optional[float] = None indexer_rope_interleave: bool = True rope_interleave: bool = True indexer_types: Optional[list] = None index_topk_freq: int = 1 index_skip_topk_offset: int = 0 index_topk_pattern: Optional[Any] = None num_nextn_predict_layers: int = 0 mlp_layer_types: Optional[list] = None def __post_init__(self): self.rope_scaling = self.rope_parameters self.rope_theta = self.rope_parameters["rope_theta"] if self.indexer_types is None: # Reference schedule: a layer runs its own indexer ("full") iff # max(i - offset + 1, 0) % freq == 0; the others ("shared") reuse the most recent # full layer's top-k selection and carry no indexer weights. f, o = self.index_topk_freq, self.index_skip_topk_offset self.indexer_types = [ "full" if max(i - o + 1, 0) % f == 0 else "shared" for i in range(self.num_hidden_layers) ] class Indexer(nn.Module): def __init__(self, args: ModelArgs): super().__init__() self.dim = args.hidden_size self.n_heads = args.index_n_heads self.head_dim = args.index_head_dim self.rope_head_dim = args.qk_rope_head_dim self.index_topk = args.index_topk self.q_lora_rank = args.q_lora_rank self.wq_b = nn.Linear( self.q_lora_rank, self.n_heads * self.head_dim, bias=False ) self.wk = nn.Linear(self.dim, self.head_dim, bias=False) self.k_norm = nn.LayerNorm(self.head_dim, eps=1e-6) self.weights_proj = nn.Linear(self.dim, self.n_heads, bias=False) self.softmax_scale = self.head_dim**-0.5 self.rope = initialize_rope( dims=args.qk_rope_head_dim, base=args.rope_theta, traditional=args.indexer_rope_interleave, max_position_embeddings=args.max_position_embeddings, scaling_config=args.rope_scaling, ) def __call__( self, x: mx.array, qr: mx.array, mask: Optional[mx.array], cache: Optional[Any] = None, ): # Computes top_k indices for attention b, s, _ = x.shape q = self.wq_b(qr) q = q.reshape(b, s, self.n_heads, self.head_dim).swapaxes(1, 2) k = self.wk(x) k = self.k_norm(k) k = mx.reshape(k, (b, 1, s, self.head_dim)) offset = cache.offset if cache is not None else 0 q = self.rope(q, offset=offset) k = self.rope(k, offset=offset) if cache is not None: k, _ = cache.update_and_fetch(k, mx.zeros([b, 1, s, 0])) if k.shape[2] <= self.index_topk: return None # Scores in float32, as the reference (which also keeps weights_proj in fp32): in bf16 # keys at the top-k boundary flip. scores = q.astype(mx.float32) @ k.astype(mx.float32).swapaxes(-1, -2) scores = mx.maximum(scores, 0) weights = self.weights_proj(x).astype(mx.float32) * (self.n_heads**-0.5 * self.softmax_scale) weights = weights.swapaxes(-1, -2)[..., None] scores = scores * weights scores = scores.sum(axis=1, keepdims=True) if mask is not None: scores = mx.where(mask, scores, -float("inf")) return mx.argpartition(scores, kth=-self.index_topk, axis=-1)[ ..., -self.index_topk : ] class DeepseekV32Attention(nn.Module): def __init__(self, config: ModelArgs, layer_idx: int = 0): 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 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 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=1e-6) 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=1e-6) 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 self.config.rope_scaling is not None: mscale_all_dim = self.config.rope_scaling.get("mscale_all_dim", 0) if mscale_all_dim: scaling_factor = self.config.rope_scaling["factor"] if scaling_factor > 1: s = 0.1 * mscale_all_dim * math.log(scaling_factor) + 1.0 self.scale = self.scale * s * s # "shared" layers carry no indexer weights: they reuse the previous full layer's top-k. self.indexer = Indexer(config) if config.indexer_types[layer_idx] == "full" else None 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=self.config.rope_scaling, ) def __call__( self, x: mx.array, mask: Optional[mx.array] = None, cache: Optional[Any] = None, prev_topk_indices: Optional[mx.array] = None, ): B, L, D = x.shape qr = self.q_a_layernorm(self.q_a_proj(x)) q = self.q_b_proj(qr) 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[0].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[0].update_and_fetch(kv_latent, k_pe) else: cache = [None] * 2 if self.indexer is not None: topk_indices = self.indexer(x, qr, mask, cache=cache[1]) else: topk_indices = prev_topk_indices if topk_indices is not None: if L == 1: idx = topk_indices[:, :, 0, :, None] kv_latent = mx.take_along_axis( kv_latent, mx.broadcast_to(idx, idx.shape[:-1] + (kv_latent.shape[-1],)), axis=2, ) k_pe = mx.take_along_axis( k_pe, mx.broadcast_to(idx, idx.shape[:-1] + (k_pe.shape[-1],)), axis=2, ) if mask is not None: mask = mx.take_along_axis(mask, topk_indices, axis=-1) else: shape = list(topk_indices.shape) shape[-1] = kv_latent.shape[2] sparse_mask = mx.zeros(shape, dtype=mx.bool_) sparse_mask = mx.put_along_axis( sparse_mask, topk_indices, mx.array(True), axis=-1 ) if mask is not None: sparse_mask = sparse_mask & mask mask = sparse_mask # Ensure the indexer cache is evaluated even if the topk_indices are unused # to keep the graph from getting too large if self.indexer is not None and cache is not None and cache[0] is not None: cache[0].keys = mx.depends(cache[0].keys, (cache[1].keys, cache[1].values)) 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), topk_indices class DeepseekV32MLP(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 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.astype(mx.float32) @ self.weight.astype(mx.float32).T, self.e_score_correction_bias, self.top_k, self.n_group, self.topk_group, self.routed_scaling_factor, self.norm_topk_prob, ) class DeepseekV32MoE(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 = DeepseekV32MLP( 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 DeepseekV32DecoderLayer(nn.Module): def __init__(self, config: ModelArgs, layer_idx: int): super().__init__() self.self_attn = DeepseekV32Attention(config, layer_idx) self.mlp = ( DeepseekV32MoE(config) if ( config.n_routed_experts is not None and layer_idx >= config.first_k_dense_replace and layer_idx % config.moe_layer_freq == 0 ) else DeepseekV32MLP(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, prev_topk_indices: Optional[mx.array] = None, ): r, topk_indices = self.self_attn(self.input_layernorm(x), mask, cache, prev_topk_indices) h = x + r r = self.mlp(self.post_attention_layernorm(h)) return h + r, topk_indices class DeepseekV32Model(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 = [ DeepseekV32DecoderLayer(config, idx) for idx in range(config.num_hidden_layers) ] self.start_idx = 0 self.end_idx = len(self.layers) self.num_layers = self.end_idx self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.pipeline_rank = 0 self.pipeline_size = 1 def pipeline(self, group): # Split layers in reverse so rank=0 gets the last layers and # rank=pipeline_size-1 gets the first self.pipeline_rank = group.rank() self.pipeline_size = group.size() layers_per_rank = len(self.layers) // self.pipeline_size extra = len(self.layers) - layers_per_rank * self.pipeline_size if self.pipeline_rank < extra: layers_per_rank += 1 self.start_idx = (self.pipeline_size - self.pipeline_rank - 1) * layers_per_rank self.end_idx = self.start_idx + layers_per_rank self.layers = self.layers[: self.end_idx] self.layers[: self.start_idx] = [None] * self.start_idx self.num_layers = len(self.layers) - self.start_idx 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] * self.num_layers mask = create_attention_mask( h, cache[0][0] if cache[0] else None, 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)) # A "shared" layer reuses the top-k selection of the most recent "full" layer. topk = None for i in range(self.num_layers): h, topk = self.layers[self.start_idx + i](h, mask, cache[i], topk) # 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][0].keys = mx.depends(cache[-1][0].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 = DeepseekV32Model(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): # Remove multi-token prediction layers mpt_layer = self.args.num_hidden_layers new_weights = {} for k, v in weights.items(): parts = k.split(".") if len(parts) >= 3 and parts[1] == "layers" and int(parts[2]) >= mpt_layer: continue new_weights[k] = v weights = new_weights def dequant(weight, scale_inv): dtype = mx.bfloat16 weight = mx.from_fp8(weight, dtype=mx.bfloat16) bs = 128 # block size m, n = weight.shape pad_bottom = (-m) % bs pad_side = (-n) % bs weight = mx.pad(weight, ((0, pad_bottom), (0, pad_side))) weight = weight.reshape( ((m + pad_bottom) // bs, bs, (n + pad_side) // bs, bs) ) weight = (weight * scale_inv[:, None, :, None]).reshape( m + pad_bottom, n + pad_side ) return weight[:m, :n].astype(dtype) # Dequantize new_weights = {} for k, v in weights.items(): if "weight_scale_inv" in k: scale_inv = v wk = k.replace("_scale_inv", "") weight = weights[wk] weight = dequant(weight, scale_inv) new_weights[wk] = weight elif k not in new_weights: 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.model.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() N = group.size() rank = group.rank() for layer in self.model.layers: layer.self_attn.q_b_proj = shard_linear( layer.self_attn.q_b_proj, "all-to-sharded", group=group ) layer.self_attn.o_proj = shard_linear( layer.self_attn.o_proj, "sharded-to-all", 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) # Shard the MLP if isinstance(layer.mlp, DeepseekV32MLP): 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 = group 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.layers[self.model.start_idx : self.model.end_idx] @property def cast_predicate(self): def predicate(k): return "e_score_correction_bias" not in k return predicate @property def quant_predicate(self): def predicate(path, module): # The lightning indexer (0.03% of parameters, reference keeps weights_proj in fp32) and the # MoE router decide *what* is computed; keep them as stored. if ".indexer." in path or path.endswith("mlp.gate"): return False return True return predicate def make_cache(self): # Shared layers have no indexer, so no indexer key cache: an empty second KVCache would # trip mlx-lm's `cache.state` evaluation (keys is None) during generation. return [ CacheList(KVCache(), KVCache()) if layer.self_attn.indexer is not None else CacheList(KVCache()) for layer in self.layers ]