from typing import Optional import mlx.core as mx import mlx.nn as nn from ..base import ( LanguageModelOutput, create_attention_mask, scaled_dot_product_attention, ) from .config import TextConfig class Attention(nn.Module): def __init__(self, config: TextConfig): super().__init__() dim = config.hidden_size self.n_heads = n_heads = config.num_attention_heads self.n_kv_heads = n_kv_heads = config.num_key_value_heads self.repeats = n_heads // n_kv_heads head_dim = config.hidden_size // n_heads self.scale = config.attention_multiplier self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=config.attention_bias) self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=config.attention_bias) self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=config.attention_bias) self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=config.attention_bias) self.rope = nn.RoPE( head_dim, traditional=config.rope_traditional, base=config.rope_theta, ) def __call__(self, x, mask=None, cache=None): 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, scale=self.scale, mask=mask ) output = output.transpose(0, 2, 1, 3).reshape(B, L, -1) return self.o_proj(output) class SharedMLP(nn.Module): def __init__(self, config: TextConfig): super().__init__() # Fused gate + up projection: output is 2 * intermediate_size self.input_linear = nn.Linear( config.hidden_size, config.shared_intermediate_size * 2, bias=config.mlp_bias, ) self.output_linear = nn.Linear( config.shared_intermediate_size, config.hidden_size, bias=config.mlp_bias ) def __call__(self, x) -> mx.array: x = self.input_linear(x) gate, x = mx.split(x, 2, axis=-1) return self.output_linear(nn.silu(gate) * x) class MLP(nn.Module): def __init__(self, config: TextConfig): super().__init__() self.gate_proj = nn.Linear( config.hidden_size, config.intermediate_size, bias=config.mlp_bias ) self.up_proj = nn.Linear( config.hidden_size, config.intermediate_size, bias=config.mlp_bias ) self.down_proj = nn.Linear( config.intermediate_size, config.hidden_size, bias=config.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, config: TextConfig): super().__init__() self.self_attn = Attention(config) self.use_shared_mlp = config.use_shared_mlp if self.use_shared_mlp: self.shared_mlp = SharedMLP(config) else: self.mlp = MLP(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 ) self.residual_multiplier = config.residual_multiplier def __call__(self, x, mask=None, cache=None): r = self.self_attn(self.input_layernorm(x), mask, cache) h = x + r * self.residual_multiplier mlp_fn = self.shared_mlp if self.use_shared_mlp else self.mlp r = mlp_fn(self.post_attention_layernorm(h)) out = h + r * self.residual_multiplier return out class Granite(nn.Module): def __init__(self, config: TextConfig): super().__init__() self.config = config self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size) self.layers = [ TransformerBlock(config=config) for _ in range(config.num_hidden_layers) ] self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.embedding_multiplier = config.embedding_multiplier def __call__( self, inputs: mx.array, inputs_embeds: Optional[mx.array] = None, mask: Optional[mx.array] = None, cache=None, **kwargs, ): # Pop deepstack features (only used during initial prefill) deepstack_visual_embeds = kwargs.pop("deepstack_visual_embeds", None) visual_pos_masks = kwargs.pop("visual_pos_masks", None) # Target layers stored as attribute by get_input_embeddings deepstack_target_layers = getattr(self, "_deepstack_target_layers", None) if inputs_embeds is None: h = self.embed_tokens(inputs) # No deepstack for cached token steps deepstack_visual_embeds = None else: h = inputs_embeds h = h * self.embedding_multiplier if cache is None: cache = [None] * len(self.layers) if mask is None: mask = create_attention_mask( h, cache[0] if cache and cache[0] is not None else cache ) for layer_idx, (layer, c) in enumerate(zip(self.layers, cache)): # Inject deepstack features at target layers if ( deepstack_visual_embeds is not None and deepstack_target_layers is not None and visual_pos_masks is not None ): for feat_idx, target_layer in enumerate(deepstack_target_layers): if layer_idx == target_layer: features = deepstack_visual_embeds[feat_idx] # Add features at image positions h = mx.where( visual_pos_masks[..., None], h + features, h, ) h = layer(h, mask, c) return self.norm(h) class LanguageModel(nn.Module): def __init__(self, config: TextConfig): super().__init__() self.config = config self.model_type = config.model_type self.model = Granite(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.logits_scaling = config.logits_scaling def __call__(self, inputs, inputs_embeds=None, mask=None, cache=None, **kwargs): out = self.model( inputs, mask=mask, cache=cache, inputs_embeds=inputs_embeds, **kwargs ) logits = self.lm_head(out) logits = logits / self.logits_scaling return LanguageModelOutput(logits=logits) @staticmethod def sanitize(weights): return { k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k } @property def layers(self): return self.model.layers @property def head_dim(self): return self.config.hidden_size // self.config.num_attention_heads @property def n_kv_heads(self): return self.config.num_key_value_heads