# AetherMind — configuration_aethermind.py # Copyright 2026 AetherMind Project. Apache License 2.0. """AetherMind model configuration. AetherMind is a fully open, trainable family of decoder-only Transformer language models with a modern architecture: * RMSNorm pre-normalization * Rotary Position Embeddings (RoPE, with optional linear scaling) * SwiGLU feed-forward blocks * Grouped-Query Attention (GQA) * Scaled-dot-product / Flash attention (via torch sdpa on CUDA) * Dynamic KV cache for fast generation * Optional Mixture-of-Experts (MoE) feed-forward for very large configs * Optional gradient checkpointing (enable via ``model.gradient_checkpointing_enable()``) The same code path scales from the CPU-trainable ``tiny`` prototype up to 70B-class configs; every hyper-parameter lives here. """ from __future__ import annotations from typing import Any, Dict, Optional from transformers import PretrainedConfig # ---------------------------------------------------------------------- # Named size presets (scaling path: tiny -> 70B) # ---------------------------------------------------------------------- SIZE_PRESETS: Dict[str, Dict[str, Any]] = { # CPU-trainable prototype used to validate the full pipeline. "tiny": dict( hidden_size=256, intermediate_size=768, num_hidden_layers=6, num_attention_heads=8, num_key_value_heads=4, head_dim=32, max_position_embeddings=1024, tie_word_embeddings=True, ), "1B": dict( hidden_size=2048, intermediate_size=5632, num_hidden_layers=24, num_attention_heads=16, num_key_value_heads=4, head_dim=128, max_position_embeddings=32768, vocab_size=65536, ), "3B": dict( hidden_size=3200, intermediate_size=8704, num_hidden_layers=32, num_attention_heads=25, num_key_value_heads=5, head_dim=128, max_position_embeddings=32768, vocab_size=65536, ), "7B": dict( hidden_size=4096, intermediate_size=11008, num_hidden_layers=32, num_attention_heads=32, num_key_value_heads=8, head_dim=128, max_position_embeddings=32768, vocab_size=65536, ), "14B": dict( hidden_size=5120, intermediate_size=13824, num_hidden_layers=40, num_attention_heads=40, num_key_value_heads=8, head_dim=128, max_position_embeddings=32768, vocab_size=65536, ), # From here up, sparse MoE feed-forward keeps FLOPs per token bounded. "32B": dict( hidden_size=5120, intermediate_size=13824, num_hidden_layers=48, num_attention_heads=40, num_key_value_heads=8, head_dim=128, max_position_embeddings=32768, vocab_size=65536, use_moe=True, num_experts=64, num_experts_per_tok=6, moe_intermediate_size=1408, moe_layers_freq=1, ), "70B": dict( hidden_size=8192, intermediate_size=28672, num_hidden_layers=80, num_attention_heads=64, num_key_value_heads=8, head_dim=128, max_position_embeddings=32768, vocab_size=65536, use_moe=True, num_experts=128, num_experts_per_tok=8, moe_intermediate_size=2048, moe_layers_freq=1, rope_scaling={"type": "linear", "factor": 4.0}, # 32K -> 128K effective ), } class AetherMindConfig(PretrainedConfig): r"""Configuration class for :class:`~model.modeling_aethermind.AetherMindModel`.""" model_type = "aethermind" keys_to_ignore_at_inference = ["past_key_values"] def __init__( self, vocab_size: int = 32768, hidden_size: int = 512, intermediate_size: int = 1376, num_hidden_layers: int = 8, num_attention_heads: int = 8, num_key_value_heads: int = 4, head_dim: Optional[int] = None, hidden_act: str = "silu", max_position_embeddings: int = 4096, initializer_range: float = 0.02, rms_norm_eps: float = 1e-6, use_cache: bool = True, pad_token_id: Optional[int] = None, bos_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, tie_word_embeddings: bool = False, rope_theta: float = 10000.0, rope_scaling: Optional[Dict[str, Any]] = None, attention_bias: bool = False, attention_dropout: float = 0.0, # ---- Mixture of Experts (for large-scale configs) ---- use_moe: bool = False, num_experts: int = 8, num_experts_per_tok: int = 2, moe_intermediate_size: Optional[int] = None, moe_layers_freq: int = 1, moe_aux_loss_coeff: float = 0.01, **kwargs, ) -> None: self.vocab_size = vocab_size self.hidden_size = hidden_size self.intermediate_size = intermediate_size self.num_hidden_layers = num_hidden_layers self.num_attention_heads = num_attention_heads if num_key_value_heads is None: num_key_value_heads = num_attention_heads self.num_key_value_heads = num_key_value_heads self.head_dim = head_dim if head_dim is not None else hidden_size // num_attention_heads self.hidden_act = hidden_act self.max_position_embeddings = max_position_embeddings self.initializer_range = initializer_range self.rms_norm_eps = rms_norm_eps self.use_cache = use_cache self.rope_theta = rope_theta self.rope_scaling = rope_scaling if rope_scaling is not None and rope_scaling.get("type", rope_scaling.get("rope_type")) not in (None, "linear"): raise ValueError("AetherMind currently supports rope_scaling type='linear' (or None).") self.attention_bias = attention_bias self.attention_dropout = attention_dropout self.use_moe = use_moe self.num_experts = num_experts self.num_experts_per_tok = num_experts_per_tok self.moe_intermediate_size = moe_intermediate_size or intermediate_size self.moe_layers_freq = moe_layers_freq self.moe_aux_loss_coeff = moe_aux_loss_coeff super().__init__( pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs, ) @classmethod def from_preset( cls, name: str = "tiny", vocab_size: Optional[int] = None, **overrides: Any, ) -> "AetherMindConfig": """Build a config from a named preset (tiny, 1B, 3B, 7B, 14B, 32B, 70B).""" if name not in SIZE_PRESETS: raise KeyError(f"Unknown preset '{name}'. Available: {sorted(SIZE_PRESETS)}") params: Dict[str, Any] = dict(SIZE_PRESETS[name]) if vocab_size is not None: params["vocab_size"] = vocab_size params.update(overrides) return cls(**params) __all__ = ["AetherMindConfig", "SIZE_PRESETS"]