Aether-6B-11Attn (open weights) — 11 sequence-mixing mechanisms on an 11x11 Latin square

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,83 @@
+---
+license: apache-2.0
+language:
+  - en
+  - ko
+pipeline_tag: text-generation
+library_name: transformers
+tags:
+  - aether
+  - open-weights
+  - heterogeneous-attention
+  - mixture-of-experts
+  - mamba2
+  - state-space-model
+  - hyena
+  - mla
+  - korean
+  - vidraft
+---
+
+# Aether-6B-11Attn
+
+**A 121-layer network that places 11 different sequence-mixing mechanisms on an 11x11 Latin square.**
+
+To our knowledge, this is the first model to integrate **11 heterogeneous sequence-mixing mechanisms** — attention, state-space, convolutional and linear families — inside a single network. Existing hybrids combine two or three (e.g. Mamba + Transformer).
+
+> **Release scope: open weights.** This repository ships the weights plus the minimum code needed to load them. Unlike [Aether-7B-5Attn](https://huggingface.co/FINAL-Bench/Aether-7B-5Attn) — which is fully open (data recipe, training code, logs, checkpoints) — the training data recipe, training code, logs and intermediate checkpoints for this model are **not** released.
+
+## The 11 mechanisms
+
+121 layers = 11 x 11. Each mechanism occupies **exactly 11 layers**, one per row and column of the Latin square, so no mechanism concentrates at any depth.
+
+| # | Mechanism | Family | Distinctive parameters (verifiable in the weights) |
+|---|-----------|--------|---------------------------------------------------|
+| 0 | `nsa` | sparse attention | `compress_mlp`, `gate_logit`, `select_score` |
+| 1 | `differential` | attention | `lambda_q1/k1/q2/k2`, `subln` |
+| 2 | `full` | attention | `q/k/v/o_proj` |
+| 3 | `mla` | latent attention | `kv_a_proj`, `kv_b_proj`, `q_a_proj`, `q_b_proj` |
+| 4 | `sliding` | attention | windowed masking over the full path |
+| 5 | `compress` | attention | `full_attn.*` submodule |
+| 6 | `hybrid` | attention | `nsa.*` + `diff.*` + `gate` + `merge_norm` |
+| 7 | `linear` | linear attention | `gate`, `norm` |
+| 8 | `mamba2` | **state-space** | `A_log`, `D`, `conv1d`, `dt_proj`, `B_proj`, `C_proj` |
+| 9 | `gdn` | gated delta net | `conv1d_q`, `conv1d_k`, `gate_proj` |
+| 10 | `hyena` | **convolutional** | `filter_dense` (implicit filter) |
+
+*Precision matters here: `full` and `sliding` share the same parameters (sliding is masking over the full path), so the weights show **10 distinct parameter signatures for 11 labels**. And several of these are not "attention" in the strict sense — `mamba2`, `hyena`, `gdn` and `linear` are other sequence-mixing families, which is exactly the point of the experiment.*
+
+## Specifications
+
+| | |
+|---|---|
+| Parameters | 5.79B total (MoE) |
+| Layers | 121 (11 x 11 Latin square) |
+| Hidden size | 1024 |
+| Attention heads | 8 query / 2 key-value (GQA), head dim 128 |
+| Experts | 25 routed (top-7) + 1 shared |
+| Vocabulary | 151,936 |
+| Precision | bfloat16 |
+| Checkpoint | training step 174,000 |
+| License | Apache-2.0 |
+
+## Provenance and honest caveats
+
+This model was reconstructed from a sharded FSDP training checkpoint. **The exact training code was not archived**, so the architecture code shipped here was rebuilt and then validated two ways:
+
+1. **Exact parameter match** — all 11,047 checkpoint tensors load with **0 missing / 0 unexpected** keys, and the parameter count matches the checkpoint exactly (5.7885B).
+2. **Perplexity** — the loaded model predicts real text sensibly (PPL 11.8 English / 6.5 Korean on sample sentences). A mis-reconstructed architecture would produce perplexity orders of magnitude worse.
+
+Treat this as a **research artifact**, not a polished product: it is a mid-training checkpoint of an architecture experiment, with no instruction tuning.
+
+## Usage
+
+Custom architecture — load with the bundled `aether_pkg/`.
+
+> **Run at batch_size = 1** (the NSA branches do not consume a padding mask), and run the forward under `torch.autocast(dtype=torch.bfloat16)` — several branches create fp32 intermediates.
+
+## Related
+
+- **Fully-open sibling:** [Aether-7B-5Attn](https://huggingface.co/FINAL-Bench/Aether-7B-5Attn) (weights + data recipe + training code + logs + checkpoints)
+- **Collection:** [Aether Foundation Model](https://huggingface.co/collections/FINAL-Bench/aether-foundation-model)
+
+*Built by VIDRAFT. License: Apache-2.0*

aether_pkg/configuration_aether_v2_11attn.py

@@ -0,0 +1,319 @@
+# coding=utf-8
+"""AETHER-30B-11Attn — Configuration.
+
+Prime Series (5/7/11/13/17/...) — 11×11 Latin Square × 11 attention types.
+
+Family ecosystem:
+  hidden_size:    2048   (V2 family 통일, Qwen3 ecosystem 호환)
+  vocab_size:     151,936 (Qwen3 base)
+  num_layers:     121    (11×11 정통 Latin Square)
+  num_experts:    25 (top-7 active)
+
+Attention types (11종):
+  Transformer 7종: NSA / Differential / Full / MLA / Sliding / Compress / Hybrid
+  Non-Transformer 4종: Linear / Mamba2 / GDN / DeltaNet  (Hyena는 conv 계열)
+
+  실제로는 11종으로 확장:
+  0: nsa            (DeepSeek 2502 NSA)
+  1: differential   (Microsoft 2410 Diff)
+  2: full           (Vaswani 2017 표준)
+  3: mla            (DeepSeek-V2/V3 Multi-head Latent)
+  4: sliding        (Window)
+  5: compress       (NSA 압축 only)
+  6: hybrid         (NSA + Diff)
+  7: linear         (Mamba-style linear)
+  8: mamba2         (SSM)
+  9: gdn            (Gated Delta Network)
+  10: hyena         (Convolutional)
+"""
+from transformers.configuration_utils import PretrainedConfig
+
+
+class AETHERV211AttnConfig(PretrainedConfig):
+    """AETHER-30B-11Attn config.
+
+    11×11 Latin Square × 121 layers
+    25-expert MoE × top-7 active per token
+    5-element cyclic FFN phase (오행)
+    """
+    model_type = "aether_v2_11attn"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        # Vocab
+        vocab_size: int = 151936,
+        pad_token_id: int = 151643,
+
+        # Model dimensions ⭐ V2 family 통일
+        hidden_size: int = 2048,
+        intermediate_size: int = 6144,
+        num_hidden_layers: int = 121,        # 11×11 Latin Square
+
+        # Attention common
+        num_attention_heads: int = 16,
+        num_key_value_heads: int = 4,        # GQA 4:1
+        head_dim: int = 128,
+        attention_dropout: float = 0.0,
+
+        # Attention type-specific
+        sliding_window_size: int = 1024,
+        compress_block_size: int = 32,
+        # NSA
+        nsa_compress_block_size: int = 32,
+        nsa_select_top_k: int = 16,
+        nsa_sliding_window: int = 512,
+        # Differential
+        diff_lambda_init: float = 0.8,
+        diff_lambda_layer_decay: float = 0.6,
+        # MLA (DeepSeek-V2/V3)
+        mla_kv_lora_rank: int = 256,         # hidden=2048에 맞춰 축소 (Pilot 512)
+        mla_q_lora_rank: int = 768,
+        mla_qk_rope_head_dim: int = 64,
+        mla_qk_nope_head_dim: int = 64,
+        mla_v_head_dim: int = 128,
+        # Hybrid
+        hybrid_groupnorm_groups: int = 8,
+        # Linear (Mamba-style)
+        linear_d_state: int = 64,
+        linear_d_conv: int = 4,
+        linear_expand: int = 1,
+        # Mamba2
+        mamba2_d_state: int = 128,
+        mamba2_d_conv: int = 4,
+        mamba2_expand: int = 2,
+        mamba2_headdim: int = 64,
+        mamba2_chunk_size: int = 256,
+        # GDN
+        gdn_d_conv: int = 4,
+        gdn_expand: int = 1,
+        gdn_chunk_size: int = 64,
+        # DeltaNet
+        deltanet_d_state: int = 64,
+        deltanet_chunk_size: int = 128,
+        # Hyena
+        hyena_filter_order: int = 64,
+        hyena_dropout: float = 0.0,
+
+        # MoE
+        num_experts: int = 25,
+        num_experts_per_tok: int = 7,
+        expert_intermediate_size: int = 1280,
+        use_shared_expert: bool = True,
+        router_aux_loss_coef: float = 0.01,
+        z_loss_coef: float = 0.001,
+
+        # Oheng (5-phase cyclic FFN)
+        oheng_enable: bool = True,
+        oheng_num_phases: int = 5,
+
+        # Position
+        max_position_embeddings: int = 8192,
+        rope_theta: float = 1000000.0,
+
+        # Norm + activation
+        rms_norm_eps: float = 1e-6,
+        hidden_act: str = "silu",
+        initializer_range: float = 0.02,
+
+        # DeepNet / training stability (121-layer deep)
+        use_deepnet_init: bool = True,
+        rezero_alpha_init: float = 0.0,
+
+        # Training / inference
+        use_cache: bool = True,
+        output_router_logits: bool = False,
+        tie_word_embeddings: bool = True,    # V2 family 동일
+
+        **kwargs,
+    ):
+        # Save all
+        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
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.attention_dropout = attention_dropout
+
+        self.sliding_window_size = sliding_window_size
+        self.compress_block_size = compress_block_size
+        self.nsa_compress_block_size = nsa_compress_block_size
+        self.nsa_select_top_k = nsa_select_top_k
+        self.nsa_sliding_window = nsa_sliding_window
+        self.diff_lambda_init = diff_lambda_init
+        self.diff_lambda_layer_decay = diff_lambda_layer_decay
+        self.mla_kv_lora_rank = mla_kv_lora_rank
+        self.mla_q_lora_rank = mla_q_lora_rank
+        self.mla_qk_rope_head_dim = mla_qk_rope_head_dim
+        self.mla_qk_nope_head_dim = mla_qk_nope_head_dim
+        self.mla_v_head_dim = mla_v_head_dim
+        self.hybrid_groupnorm_groups = hybrid_groupnorm_groups
+        self.linear_d_state = linear_d_state
+        self.linear_d_conv = linear_d_conv
+        self.linear_expand = linear_expand
+        self.mamba2_d_state = mamba2_d_state
+        self.mamba2_d_conv = mamba2_d_conv
+        self.mamba2_expand = mamba2_expand
+        self.mamba2_headdim = mamba2_headdim
+        self.mamba2_chunk_size = mamba2_chunk_size
+        self.gdn_d_conv = gdn_d_conv
+        self.gdn_expand = gdn_expand
+        self.gdn_chunk_size = gdn_chunk_size
+        self.deltanet_d_state = deltanet_d_state
+        self.deltanet_chunk_size = deltanet_chunk_size
+        self.hyena_filter_order = hyena_filter_order
+        self.hyena_dropout = hyena_dropout
+
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.expert_intermediate_size = expert_intermediate_size
+        self.use_shared_expert = use_shared_expert
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.z_loss_coef = z_loss_coef
+        self.oheng_enable = oheng_enable
+        self.oheng_num_phases = oheng_num_phases
+
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.rms_norm_eps = rms_norm_eps
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.use_deepnet_init = use_deepnet_init
+        self.rezero_alpha_init = rezero_alpha_init
+
+        self.use_cache = use_cache
+        self.output_router_logits = output_router_logits
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+# =============================================================================
+# Pre-defined configs (nano / mini / full)
+# =============================================================================
+
+# Nano: forward test 검증용 (~150M total)
+nano_11attn_cfg = AETHERV211AttnConfig(
+    hidden_size=512,
+    intermediate_size=1536,
+    expert_intermediate_size=192,
+    num_hidden_layers=121,
+    num_attention_heads=8,
+    num_key_value_heads=2,
+    head_dim=64,
+    sliding_window_size=128,
+    compress_block_size=16,
+    nsa_compress_block_size=16,
+    nsa_sliding_window=64,
+    mla_kv_lora_rank=128,
+    mla_q_lora_rank=384,
+    mla_qk_rope_head_dim=32,
+    mla_qk_nope_head_dim=32,
+    mla_v_head_dim=64,
+    linear_d_state=32,
+    mamba2_d_state=64,
+    mamba2_headdim=32,
+    deltanet_d_state=32,
+    hyena_filter_order=32,
+)
+
+# Mini: PoC 학습 (~3B total / 0.8B active)
+mini_11attn_cfg = AETHERV211AttnConfig(
+    hidden_size=1024,
+    intermediate_size=3072,
+    expert_intermediate_size=512,
+    num_hidden_layers=121,
+    num_attention_heads=8,
+    num_key_value_heads=2,
+    head_dim=128,
+)
+
+# Full ⭐ AETHER-30B-11Attn target
+full_11attn_cfg = AETHERV211AttnConfig(
+    hidden_size=2048,
+    intermediate_size=6144,
+    expert_intermediate_size=1536,    # 30B target fit
+    num_hidden_layers=121,
+    num_attention_heads=16,
+    num_key_value_heads=4,
+    head_dim=128,
+    sliding_window_size=2048,
+)
+
+
+def estimate_params_11attn(cfg: AETHERV211AttnConfig) -> dict:
+    """Estimate AETHER-30B-11Attn parameter count."""
+    h = cfg.hidden_size
+    n_layers = cfg.num_hidden_layers
+    n_heads = cfg.num_attention_heads
+    n_kv = cfg.num_key_value_heads
+    head_dim = cfg.head_dim
+    n_experts = cfg.num_experts
+    e_inter = cfg.expert_intermediate_size
+    vocab = cfg.vocab_size
+
+    # Embed (tied with LM head)
+    embed = vocab * h
+
+    # Per-layer (average across 11 attention types)
+    norm = 2 * h
+
+    # Attention avg (대략):
+    # Full/Sliding: 4 × h² × heads/(heads+kv) ≈ h × (heads+2kv) × head_dim
+    # MLA: lower (low-rank), but not significantly
+    # Mamba2/GDN/Linear: smaller
+    attn_avg = h * (n_heads + 2 * n_kv) * head_dim + h * h  # Q,K,V,O
+
+    # MoE
+    moe_gate = h * n_experts
+    moe_experts = n_experts * 3 * h * e_inter  # gate_proj, up_proj, down_proj
+    moe_shared = 3 * h * e_inter
+
+    per_layer = norm + attn_avg + moe_gate + moe_experts + moe_shared
+
+    total = embed + n_layers * per_layer
+
+    # Active per token (top-k experts + shared)
+    active_per_layer = norm + attn_avg + moe_gate + cfg.num_experts_per_tok * 3 * h * e_inter + moe_shared
+    active_total = embed + n_layers * active_per_layer
+
+    return {
+        "total": total,
+        "total_M": total / 1e6,
+        "total_B": total / 1e9,
+        "active": active_total,
+        "active_M": active_total / 1e6,
+        "active_B": active_total / 1e9,
+        "per_layer_M": per_layer / 1e6,
+    }
+
+
+__all__ = [
+    "AETHERV211AttnConfig",
+    "nano_11attn_cfg",
+    "mini_11attn_cfg",
+    "full_11attn_cfg",
+    "estimate_params_11attn",
+]
+
+
+if __name__ == "__main__":
+    print("=" * 70)
+    print("AETHER-30B-11Attn Param Estimation")
+    print("=" * 70)
+    for name, cfg in [("nano", nano_11attn_cfg), ("mini", mini_11attn_cfg), ("full", full_11attn_cfg)]:
+        info = estimate_params_11attn(cfg)
+        print(f"\n[{name}]")
+        print(f"  hidden:      {cfg.hidden_size}")
+        print(f"  layers:      {cfg.num_hidden_layers}")
+        print(f"  experts:     {cfg.num_experts} (top-{cfg.num_experts_per_tok})")
+        print(f"  expert_int:  {cfg.expert_intermediate_size}")
+        print(f"  Total:       {info['total_B']:.2f}B  ({info['total_M']:.0f}M)")
+        print(f"  Active:      {info['active_B']:.2f}B  ({info['active_M']:.0f}M)")
+        print(f"  Per layer:   {info['per_layer_M']:.0f}M")

aether_pkg/modeling_aether_v2_11way.py

@@ -0,0 +1,842 @@
+# coding=utf-8
+# Copyright 2026 VIDRAFT (비드래프트). All rights reserved.
+#
+# AETHER-V2-7way: 7-aware attention + 7×7 Latin Square 49-layer MoE
+# Built upon HuggingFace Transformers conventions.
+#
+# Architecture:
+#   - 49 layers organized as 7×7 Latin Square
+#   - 7 distinct attention types (NSA, Differential, Full, Linear, Sliding, Compress, Hybrid)
+#   - 25 experts per layer, top-7 active per token
+#   - Each row of latin square = 1 cycle of 7 attention types
+#   - Each column = different ordering (Latin square property)
+#
+# Layer index → (row, col) → attention type via LATIN_SQUARE_7x7
+#
+"""PyTorch AETHER-V2-7way model."""
+
+from __future__ import annotations
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+
+from .configuration_aether_v2_11attn import AETHERV211AttnConfig
+
+# 7-aware attention modules (already authored, in v2_attentions/)
+from .v2_attentions.nsa import NSAAttention
+from .v2_attentions.differential import DifferentialAttention
+
+logger = logging.get_logger(__name__)
+
+
+# =============================================================================
+# 7×7 Latin Square — Layer → (attention_type, ffn_phase) 매핑
+# =============================================================================
+# Latin Square property: each row & column has each of {0..6} exactly once.
+# row = layer // 7   (0..6)
+# col = layer % 7    (0..6)
+# attention_type = LATIN_SQUARE_7x7[row][col]
+#
+# 5-element cyclic FFN phase (오행 상생 — Q2B OhengGate inspired):
+# ffn_phase = layer % 5
+#
+# Cyclic Latin Square 11x11 (prime number — Prime Series)
+LATIN_SQUARE_11x11 = [
+    [(i + j) % 11 for j in range(11)] for i in range(11)
+]
+
+# Backward compat alias
+LATIN_SQUARE_7x7 = LATIN_SQUARE_11x11  # for v2_7way imports
+
+# Attention type names (0..6)
+ATTN_TYPES = [
+    "nsa",            # 0: Native Sparse Attention (3-branch)  [Transformer]
+    "differential",   # 1: Differential Attention (λ-gated)    [Transformer]
+    "full",           # 2: Full Attention (standard)           [Transformer]
+    "mla",            # 3: Multi-head Latent (DeepSeek-V2/V3)  [Transformer]
+    "sliding",        # 4: Sliding Window Attention            [Transformer]
+    "compress",       # 5: Compress-only (NSA subset)          [Transformer]
+    "hybrid",         # 6: NSA + Differential combined         [Transformer]
+    "linear",         # 7: Linear Attention (Mamba-style)      [Non-Transformer]
+    "mamba2",         # 8: Mamba2 (SSM)                        [Non-Transformer]
+    "gdn",            # 9: Gated Delta Network                 [Non-Transformer]
+    "hyena",          # 10: Hyena (Convolutional)              [Non-Transformer]
+]
+
+
+def get_attention_type(layer_idx: int) -> str:
+    """Layer index → attention type via 11x11 Latin Square."""
+    row = layer_idx // 11
+    col = layer_idx % 11
+    type_idx = LATIN_SQUARE_11x11[row][col]
+    return ATTN_TYPES[type_idx]
+
+
+def get_ffn_phase(layer_idx: int) -> int:
+    """Layer index → 5-element cyclic phase (오행)."""
+    return layer_idx % 5
+
+
+# =============================================================================
+# Rotary Position Embedding (RoPE)
+# =============================================================================
+class AETHERV211AttnRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, max_pos: int = 4096, base: float = 10000.0, device=None):
+        super().__init__()
+        self.dim = dim
+        self.max_pos = max_pos
+        self.base = base
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._build_cos_sin_cache(max_pos, device or torch.device("cpu"))
+
+    def _build_cos_sin_cache(self, seq_len: int, device, dtype=torch.float32):
+        t = torch.arange(seq_len, device=device, dtype=torch.float32)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat([freqs, freqs], dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x: torch.Tensor, position_ids: torch.Tensor):
+        if position_ids.max() >= self.cos_cached.size(0):
+            self._build_cos_sin_cache(int(position_ids.max() + 1), x.device, x.dtype)
+        cos = self.cos_cached[position_ids].to(x.dtype)
+        sin = self.sin_cached[position_ids].to(x.dtype)
+        return cos, sin
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    x1, x2 = x.chunk(2, dim=-1)
+    return torch.cat([-x2, x1], dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+# =============================================================================
+# RMSNorm
+# =============================================================================
+class AETHERV211AttnRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        in_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+        return self.weight * hidden_states.to(in_dtype)
+
+
+# =============================================================================
+# Standard Multi-Head Attention (Full Attention type)
+# =============================================================================
+class FullAttention(nn.Module):
+    """Standard multi-head attention with GQA support."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = getattr(config, "num_key_value_heads", config.num_attention_heads)
+        self.head_dim = config.head_dim
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.rotary = AETHERV211AttnRotaryEmbedding(
+            self.head_dim, config.max_position_embeddings, config.rope_theta,
+        )
+
+    def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
+        if self.num_kv_groups == 1:
+            return x
+        bsz, n_kv, seq, dim = x.shape
+        return x[:, :, None, :, :].expand(bsz, n_kv, self.num_kv_groups, seq, dim).reshape(
+            bsz, n_kv * self.num_kv_groups, seq, dim,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[Cache]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        q = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary(v, position_ids)
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        if past_key_value is not None:
+            k, v = past_key_value.update(k, v, self.layer_idx)
+
+        k = self._repeat_kv(k)
+        v = self._repeat_kv(v)
+
+        attn_out = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=(attention_mask.to(q.dtype) if attention_mask is not None else None),
+            dropout_p=0.0 if not self.training else self.config.attention_dropout,
+            is_causal=(attention_mask is None and q_len > 1),
+        )
+        attn_out = attn_out.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(attn_out), past_key_value
+
+
+# =============================================================================
+# Linear Attention (Mamba-style, simplified)
+# =============================================================================
+class LinearAttention(nn.Module):
+    """Linear attention (Mamba/RWKV-inspired) for long-context efficiency."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = config.head_dim
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.gate = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.norm = AETHERV211AttnRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[Cache]]:
+        bsz, q_len, _ = hidden_states.size()
+        q = F.elu(self.q_proj(hidden_states), alpha=1.0).view(bsz, q_len, self.num_heads, self.head_dim) + 1
+        k = F.elu(self.k_proj(hidden_states), alpha=1.0).view(bsz, q_len, self.num_heads, self.head_dim) + 1
+        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
+        g = self.gate(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).sigmoid()
+
+        # Linear attention: O(n*d^2) instead of O(n^2*d)
+        kv = torch.einsum("bnhd,bnhe->bhde", k, v)  # (bsz, h, d, e)
+        out = torch.einsum("bnhd,bhde->bnhe", q, kv)
+        z = torch.einsum("bnhd,bhd->bnh", q, k.sum(dim=1)).unsqueeze(-1).clamp_min(1e-6)
+        out = (out / z) * g
+        out = self.norm(out).reshape(bsz, q_len, -1)
+        return self.o_proj(out), past_key_value
+
+
+# =============================================================================
+# Sliding Window Attention
+# =============================================================================
+class SlidingWindowAttention(FullAttention):
+    """Standard MHA but limited to local window for efficiency."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.window_size = getattr(config, "sliding_window_size", 512)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, _ = hidden_states.size()
+        if attention_mask is None and q_len > self.window_size:
+            mask = torch.ones(q_len, q_len, dtype=torch.bool, device=hidden_states.device)
+            mask = torch.tril(mask) & torch.triu(mask, diagonal=-self.window_size)
+            attention_mask = torch.where(mask, 0.0, float("-inf")).unsqueeze(0).unsqueeze(0)
+        return super().forward(hidden_states, attention_mask, position_ids, past_key_value, use_cache, **kwargs)
+
+
+# =============================================================================
+# Compress Attention (NSA-subset, just compress branch)
+# =============================================================================
+class CompressAttention(nn.Module):
+    """Compress branch: reduce KV cache via local average, then full attention on compressed."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.compress_block = getattr(config, "compress_block_size", 16)
+        self.full_attn = FullAttention(config, layer_idx)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, dim = hidden_states.size()
+        # Compress along seq dim by averaging blocks
+        if q_len % self.compress_block == 0 and q_len > self.compress_block:
+            n_blocks = q_len // self.compress_block
+            x_comp = hidden_states.view(bsz, n_blocks, self.compress_block, dim).mean(dim=2)
+            pos_comp = position_ids[:, ::self.compress_block] if position_ids is not None else None
+            out_comp, kv = self.full_attn(x_comp, None, pos_comp, past_key_value, use_cache)
+            # Up-sample by repeat
+            out = out_comp.unsqueeze(2).expand(bsz, n_blocks, self.compress_block, dim).reshape(bsz, q_len, dim)
+            return out, kv
+        return self.full_attn(hidden_states, attention_mask, position_ids, past_key_value, use_cache)
+
+
+# =============================================================================
+# Hybrid Attention (NSA + Differential combined)
+# =============================================================================
+class HybridAttention(nn.Module):
+    """Combine NSA + Differential outputs via learnable gate + final norm (stable).
+
+    Fix v2 (2026-05-05): added post-merge GroupNorm + gate init=0 (sigmoid(0)=0.5 exact balance)
+    + lightly scaled output to prevent 49-layer cumulative divergence.
+    """
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.nsa = NSAAttention(config, layer_idx)
+        self.diff = DifferentialAttention(config, layer_idx)
+        # Per-channel gate (richer than scalar), init to 0 → sigmoid(0)=0.5 exact balance
+        self.gate = nn.Parameter(torch.zeros(config.hidden_size))
+        # Post-merge GroupNorm to stabilize 49-layer stack
+        n_groups = 8 if config.hidden_size % 8 == 0 else 1
+        self.merge_norm = AETHERV211AttnRMSNorm(config.hidden_size, eps=getattr(config,'rms_norm_eps',1e-6))
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        nsa_out, kv1 = self.nsa(hidden_states, attention_mask, position_ids, past_key_value, use_cache, **kwargs)
+        diff_out, kv2 = self.diff(hidden_states, attention_mask, position_ids, past_key_value, use_cache, **kwargs)
+        # Per-channel learnable mix: g (sigmoid) per channel
+        g = torch.sigmoid(self.gate)  # shape (hidden_size,)
+        out = g * nsa_out + (1.0 - g) * diff_out
+        # Stabilize via GroupNorm: (B, S, H) → (B, H, S) → norm → (B, S, H)
+        out = self.merge_norm(out)
+        return out, kv1 if kv1 is not None else kv2
+
+
+# =============================================================================
+# 7-aware Attention Dispatcher
+# =============================================================================
+def build_attention(config: AETHERV211AttnConfig, layer_idx: int) -> nn.Module:
+    """Pick attention type based on Latin Square index."""
+    attn_type = get_attention_type(layer_idx)
+    if attn_type == "nsa":
+        return NSAAttention(config, layer_idx)
+    elif attn_type == "differential":
+        return DifferentialAttention(config, layer_idx)
+    elif attn_type == "full":
+        return FullAttention(config, layer_idx)
+    elif attn_type == "linear":
+        return LinearAttention(config, layer_idx)
+    elif attn_type == "sliding":
+        return SlidingWindowAttention(config, layer_idx)
+    elif attn_type == "compress":
+        return CompressAttention(config, layer_idx)
+    elif attn_type == "hybrid":
+        return HybridAttention(config, layer_idx)
+    elif attn_type == "mla":
+        from aether_pkg.v2_attentions.mla import MLAAttention
+        return MLAAttention(config, layer_idx)
+    elif attn_type == "mamba2":
+        from aether_pkg.v2_attentions.mamba2 import Mamba2Attention
+        return Mamba2Attention(config, layer_idx)
+    elif attn_type == "gdn":
+        from aether_pkg.v2_attentions.gdn import GDNAttention
+        return GDNAttention(config, layer_idx)
+    elif attn_type == "hyena":
+        from aether_pkg.v2_attentions.hyena import HyenaAttention
+        return HyenaAttention(config, layer_idx)
+    raise ValueError(f"Unknown attention type: {attn_type}")
+
+
+# =============================================================================
+# MoE Block: 25 experts, top-7 active per token
+# =============================================================================
+class AETHERV211AttnMLP(nn.Module):
+    """Single expert MLP (SwiGLU)."""
+
+    def __init__(self, config: AETHERV211AttnConfig, intermediate_size: Optional[int] = None):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size or config.expert_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)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class AETHERV211AttnSparseMoE(nn.Module):
+    """25-expert MoE with top-7 active routing.
+
+    Each layer has a 5-phase cyclic FFN bias to encode 오행 (Five Elements) cycle.
+    """
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.ffn_phase = get_ffn_phase(layer_idx)  # 0..4 (5-element cycle)
+
+        # Router: hidden → num_experts logits
+        self.gate = nn.Linear(self.hidden_size, self.num_experts, bias=False)
+
+        # 25 experts (each is a SwiGLU MLP)
+        self.experts = nn.ModuleList([
+            AETHERV211AttnMLP(config) for _ in range(self.num_experts)
+        ])
+
+        # 오행 cyclic phase bias (learnable, 5 phases)
+        self.phase_bias = nn.Parameter(torch.zeros(5, self.num_experts))
+
+        # Optional shared expert (always active, optional)
+        self.use_shared_expert = getattr(config, "use_shared_expert", True)
+        if self.use_shared_expert:
+            self.shared_expert = AETHERV211AttnMLP(
+                config, intermediate_size=config.expert_intermediate_size,
+            )
+            self.shared_expert_gate = nn.Linear(self.hidden_size, 1, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        bsz, seq_len, dim = hidden_states.shape
+        x = hidden_states.view(-1, dim)  # (bsz*seq, dim)
+
+        # Routing
+        router_logits = self.gate(x)  # (bsz*seq, num_experts)
+        # Add 5-phase cyclic bias (오행)
+        router_logits = router_logits + self.phase_bias[self.ffn_phase].unsqueeze(0)
+
+        # Top-k selection
+        routing_weights, selected_experts = torch.topk(router_logits, self.top_k, dim=-1)
+        routing_weights = F.softmax(routing_weights, dim=-1)
+
+        # Initialize output
+        final_out = torch.zeros_like(x)
+
+        # Per-expert dispatch (loop is OK for prototype; can be optimized with grouped GEMM later)
+        for e in range(self.num_experts):
+            mask = (selected_experts == e)
+            if not mask.any():
+                continue
+            token_idx, k_idx = mask.nonzero(as_tuple=True)
+            expert_in = x[token_idx]
+            expert_out = self.experts[e](expert_in)
+            weight = routing_weights[token_idx, k_idx].unsqueeze(-1).to(expert_out.dtype)
+            final_out.index_add_(0, token_idx, (expert_out * weight).to(final_out.dtype))
+
+        # Shared expert
+        if self.use_shared_expert:
+            shared_out = self.shared_expert(x)
+            shared_gate = torch.sigmoid(self.shared_expert_gate(x))
+            final_out = final_out + (shared_out * shared_gate).to(final_out.dtype)
+
+        final_out = final_out.view(bsz, seq_len, dim)
+        return final_out, router_logits.view(bsz, seq_len, self.num_experts)
+
+
+# =============================================================================
+# Decoder Layer: Attention + MoE FFN with 7-aware + 5-phase logic
+# =============================================================================
+class AETHERV211AttnDecoderLayer(nn.Module):
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.attn_type = get_attention_type(layer_idx)
+        self.ffn_phase = get_ffn_phase(layer_idx)
+
+        # 7-aware attention (1 of 7 types based on Latin square)
+        self.self_attn = build_attention(config, layer_idx)
+
+        # MoE FFN with 5-phase cyclic bias
+        self.mlp = AETHERV211AttnSparseMoE(config, layer_idx)
+
+        # Norms
+        self.input_layernorm = AETHERV211AttnRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = AETHERV211AttnRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_router_logits: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[Cache], Optional[torch.Tensor]]:
+        # Self-attention with residual
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states, kv = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # MoE FFN with residual
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states, router_logits = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, kv)
+        if output_router_logits:
+            outputs = outputs + (router_logits,)
+        else:
+            outputs = outputs + (None,)
+        return outputs
+
+
+# =============================================================================
+# Pretrained base
+# =============================================================================
+class AETHERV211AttnPreTrainedModel(PreTrainedModel):
+    config_class = AETHERV211AttnConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["AETHERV211AttnDecoderLayer"]
+    _supports_cache_class = True
+    _supports_static_cache = False
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, AETHERV211AttnRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+# =============================================================================
+# Main Model
+# =============================================================================
+class AETHERV211AttnModel(AETHERV211AttnPreTrainedModel):
+    """49-layer decoder-only model with 7-aware attention + MoE."""
+
+    def __init__(self, config: AETHERV211AttnConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([
+            AETHERV211AttnDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)
+        ])
+        self.norm = AETHERV211AttnRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else False
+        output_router_logits = output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("Cannot specify both input_ids and inputs_embeds")
+        if input_ids is not None:
+            bsz, seq_len = input_ids.shape
+        elif inputs_embeds is not None:
+            bsz, seq_len, _ = inputs_embeds.shape
+        else:
+            raise ValueError("Either input_ids or inputs_embeds must be provided")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        past_seen = past_key_values.get_seq_length() if past_key_values is not None else 0
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_seen, past_seen + seq_len, device=inputs_embeds.device,
+            ).unsqueeze(0)
+
+        hidden_states = inputs_embeds
+
+        all_hidden_states = () if output_hidden_states else None
+        all_router_logits = () if output_router_logits else None
+
+        for layer_idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_out = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states, attention_mask, position_ids,
+                    past_key_values, output_router_logits, use_cache,
+                )
+            else:
+                layer_out = decoder_layer(
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_out[0]
+            if output_router_logits:
+                all_router_logits += (layer_out[2],)
+
+        hidden_states = self.norm(hidden_states)
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [
+                hidden_states, past_key_values, all_hidden_states, None, all_router_logits,
+            ] if v is not None)
+
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=None,
+            router_logits=all_router_logits,
+        )
+
+
+# =============================================================================
+# Causal LM Wrapper
+# =============================================================================
+class AETHERV211AttnForCausalLM(AETHERV211AttnPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: AETHERV211AttnConfig):
+        super().__init__(config)
+        self.model = AETHERV211AttnModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = getattr(config, "router_aux_loss_coef", 0.001)
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        output_router_logits = output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=True,
+        )
+        hidden_states = outputs.last_hidden_state
+        logits = self.lm_head(hidden_states).float()
+
+        loss = None
+        aux_loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss = F.cross_entropy(
+                shift_logits.view(-1, self.vocab_size),
+                shift_labels.view(-1),
+                ignore_index=-100,
+            )
+
+        if output_router_logits and outputs.router_logits is not None:
+            aux_loss = self._compute_router_aux_loss(outputs.router_logits, attention_mask)
+            if loss is not None:
+                loss = loss + self.router_aux_loss_coef * aux_loss
+
+        if not return_dict:
+            output = (logits,) + tuple(v for v in [
+                outputs.past_key_values, outputs.hidden_states, None, outputs.router_logits, aux_loss,
+            ] if v is not None)
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=None,
+            router_logits=outputs.router_logits,
+        )
+
+    def _compute_router_aux_loss(self, router_logits: Tuple[torch.Tensor, ...], attention_mask=None):
+        """Standard switch-transformer auxiliary loss for load balancing."""
+        if router_logits is None or len(router_logits) == 0:
+            return None
+        # Each router_logits[i] shape: (bsz, seq, num_experts) → flatten to (n_tokens, num_experts)
+        flat = []
+        for r in router_logits:
+            if r is None:
+                continue
+            flat.append(r.reshape(-1, self.num_experts))
+        if not flat:
+            return None
+        all_router_logits = torch.cat(flat, dim=0)  # (total_tokens, num_experts)
+
+        routing_weights = F.softmax(all_router_logits.float(), dim=-1)
+        _, selected_experts = torch.topk(routing_weights, self.num_experts_per_tok, dim=-1)
+
+        # Expert mask: (n_tokens, top_k, num_experts)
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts).float()
+        # Tokens-per-expert frequency: average over (n_tokens, top_k) dims → (num_experts,)
+        tokens_per_expert = expert_mask.mean(dim=(0, 1))
+        # Router prob per expert: (num_experts,)
+        router_prob_per_expert = routing_weights.mean(dim=0)
+        # aux_loss = num_experts * sum(token_freq * prob)
+        return self.num_experts * torch.sum(tokens_per_expert * router_prob_per_expert)
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values is not None:
+                position_ids = position_ids[:, -input_ids.shape[1]:]
+        return {
+            "input_ids": input_ids,
+            "position_ids": position_ids,
+            "past_key_values": past_key_values,
+            "use_cache": kwargs.get("use_cache"),
+            "attention_mask": attention_mask,
+        }
+
+
+# =============================================================================
+# Helper: Latin Square Layer Map (for analysis / debugging)
+# =============================================================================
+def print_layer_map(num_layers: int = 49):
+    """Print the Latin Square attention type map."""
+    print(f"=== AETHER-V2-7way Layer Map ({num_layers} layers) ===")
+    for L in range(num_layers):
+        attn = get_attention_type(L)
+        phase = get_ffn_phase(L)
+        row = L // 7
+        col = L % 7
+        print(f"  L{L:02d} (row={row} col={col}): attn={attn:12s} ffn_phase={phase}")
+
+
+__all__ = [
+    "AETHERV211AttnConfig",
+    "AETHERV211AttnModel",
+    "AETHERV211AttnForCausalLM",
+    "AETHERV211AttnPreTrainedModel",
+    "AETHERV211AttnDecoderLayer",
+    "AETHERV211AttnSparseMoE",
+    "build_attention",
+    "get_attention_type",
+    "get_ffn_phase",
+    "LATIN_SQUARE_7x7",
+    "ATTN_TYPES",
+    "print_layer_map",
+]

aether_pkg/modeling_aether_v2_11way.py.groupnorm_backup

@@ -0,0 +1,842 @@
+# coding=utf-8
+# Copyright 2026 VIDRAFT (비드래프트). All rights reserved.
+#
+# AETHER-V2-7way: 7-aware attention + 7×7 Latin Square 49-layer MoE
+# Built upon HuggingFace Transformers conventions.
+#
+# Architecture:
+#   - 49 layers organized as 7×7 Latin Square
+#   - 7 distinct attention types (NSA, Differential, Full, Linear, Sliding, Compress, Hybrid)
+#   - 25 experts per layer, top-7 active per token
+#   - Each row of latin square = 1 cycle of 7 attention types
+#   - Each column = different ordering (Latin square property)
+#
+# Layer index → (row, col) → attention type via LATIN_SQUARE_7x7
+#
+"""PyTorch AETHER-V2-7way model."""
+
+from __future__ import annotations
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+
+from .configuration_aether_v2_11attn import AETHERV211AttnConfig
+
+# 7-aware attention modules (already authored, in v2_attentions/)
+from .v2_attentions.nsa import NSAAttention
+from .v2_attentions.differential import DifferentialAttention
+
+logger = logging.get_logger(__name__)
+
+
+# =============================================================================
+# 7×7 Latin Square — Layer → (attention_type, ffn_phase) 매핑
+# =============================================================================
+# Latin Square property: each row & column has each of {0..6} exactly once.
+# row = layer // 7   (0..6)
+# col = layer % 7    (0..6)
+# attention_type = LATIN_SQUARE_7x7[row][col]
+#
+# 5-element cyclic FFN phase (오행 상생 — Q2B OhengGate inspired):
+# ffn_phase = layer % 5
+#
+# Cyclic Latin Square 11x11 (prime number — Prime Series)
+LATIN_SQUARE_11x11 = [
+    [(i + j) % 11 for j in range(11)] for i in range(11)
+]
+
+# Backward compat alias
+LATIN_SQUARE_7x7 = LATIN_SQUARE_11x11  # for v2_7way imports
+
+# Attention type names (0..6)
+ATTN_TYPES = [
+    "nsa",            # 0: Native Sparse Attention (3-branch)  [Transformer]
+    "differential",   # 1: Differential Attention (λ-gated)    [Transformer]
+    "full",           # 2: Full Attention (standard)           [Transformer]
+    "mla",            # 3: Multi-head Latent (DeepSeek-V2/V3)  [Transformer]
+    "sliding",        # 4: Sliding Window Attention            [Transformer]
+    "compress",       # 5: Compress-only (NSA subset)          [Transformer]
+    "hybrid",         # 6: NSA + Differential combined         [Transformer]
+    "linear",         # 7: Linear Attention (Mamba-style)      [Non-Transformer]
+    "mamba2",         # 8: Mamba2 (SSM)                        [Non-Transformer]
+    "gdn",            # 9: Gated Delta Network                 [Non-Transformer]
+    "hyena",          # 10: Hyena (Convolutional)              [Non-Transformer]
+]
+
+
+def get_attention_type(layer_idx: int) -> str:
+    """Layer index → attention type via 11x11 Latin Square."""
+    row = layer_idx // 11
+    col = layer_idx % 11
+    type_idx = LATIN_SQUARE_11x11[row][col]
+    return ATTN_TYPES[type_idx]
+
+
+def get_ffn_phase(layer_idx: int) -> int:
+    """Layer index → 5-element cyclic phase (오행)."""
+    return layer_idx % 5
+
+
+# =============================================================================
+# Rotary Position Embedding (RoPE)
+# =============================================================================
+class AETHERV211AttnRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, max_pos: int = 4096, base: float = 10000.0, device=None):
+        super().__init__()
+        self.dim = dim
+        self.max_pos = max_pos
+        self.base = base
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._build_cos_sin_cache(max_pos, device or torch.device("cpu"))
+
+    def _build_cos_sin_cache(self, seq_len: int, device, dtype=torch.float32):
+        t = torch.arange(seq_len, device=device, dtype=torch.float32)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat([freqs, freqs], dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x: torch.Tensor, position_ids: torch.Tensor):
+        if position_ids.max() >= self.cos_cached.size(0):
+            self._build_cos_sin_cache(int(position_ids.max() + 1), x.device, x.dtype)
+        cos = self.cos_cached[position_ids].to(x.dtype)
+        sin = self.sin_cached[position_ids].to(x.dtype)
+        return cos, sin
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    x1, x2 = x.chunk(2, dim=-1)
+    return torch.cat([-x2, x1], dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+# =============================================================================
+# RMSNorm
+# =============================================================================
+class AETHERV211AttnRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.eps = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        in_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+        return self.weight * hidden_states.to(in_dtype)
+
+
+# =============================================================================
+# Standard Multi-Head Attention (Full Attention type)
+# =============================================================================
+class FullAttention(nn.Module):
+    """Standard multi-head attention with GQA support."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = getattr(config, "num_key_value_heads", config.num_attention_heads)
+        self.head_dim = config.head_dim
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        self.rotary = AETHERV211AttnRotaryEmbedding(
+            self.head_dim, config.max_position_embeddings, config.rope_theta,
+        )
+
+    def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
+        if self.num_kv_groups == 1:
+            return x
+        bsz, n_kv, seq, dim = x.shape
+        return x[:, :, None, :, :].expand(bsz, n_kv, self.num_kv_groups, seq, dim).reshape(
+            bsz, n_kv * self.num_kv_groups, seq, dim,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[Cache]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        q = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+
+        cos, sin = self.rotary(v, position_ids)
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        if past_key_value is not None:
+            k, v = past_key_value.update(k, v, self.layer_idx)
+
+        k = self._repeat_kv(k)
+        v = self._repeat_kv(v)
+
+        attn_out = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=(attention_mask.to(q.dtype) if attention_mask is not None else None),
+            dropout_p=0.0 if not self.training else self.config.attention_dropout,
+            is_causal=(attention_mask is None and q_len > 1),
+        )
+        attn_out = attn_out.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(attn_out), past_key_value
+
+
+# =============================================================================
+# Linear Attention (Mamba-style, simplified)
+# =============================================================================
+class LinearAttention(nn.Module):
+    """Linear attention (Mamba/RWKV-inspired) for long-context efficiency."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = config.head_dim
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.gate = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.norm = AETHERV211AttnRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[Cache]]:
+        bsz, q_len, _ = hidden_states.size()
+        q = F.elu(self.q_proj(hidden_states), alpha=1.0).view(bsz, q_len, self.num_heads, self.head_dim) + 1
+        k = F.elu(self.k_proj(hidden_states), alpha=1.0).view(bsz, q_len, self.num_heads, self.head_dim) + 1
+        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
+        g = self.gate(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).sigmoid()
+
+        # Linear attention: O(n*d^2) instead of O(n^2*d)
+        kv = torch.einsum("bnhd,bnhe->bhde", k, v)  # (bsz, h, d, e)
+        out = torch.einsum("bnhd,bhde->bnhe", q, kv)
+        z = torch.einsum("bnhd,bhd->bnh", q, k.sum(dim=1)).unsqueeze(-1).clamp_min(1e-6)
+        out = (out / z) * g
+        out = self.norm(out).reshape(bsz, q_len, -1)
+        return self.o_proj(out), past_key_value
+
+
+# =============================================================================
+# Sliding Window Attention
+# =============================================================================
+class SlidingWindowAttention(FullAttention):
+    """Standard MHA but limited to local window for efficiency."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.window_size = getattr(config, "sliding_window_size", 512)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, _ = hidden_states.size()
+        if attention_mask is None and q_len > self.window_size:
+            mask = torch.ones(q_len, q_len, dtype=torch.bool, device=hidden_states.device)
+            mask = torch.tril(mask) & torch.triu(mask, diagonal=-self.window_size)
+            attention_mask = torch.where(mask, 0.0, float("-inf")).unsqueeze(0).unsqueeze(0)
+        return super().forward(hidden_states, attention_mask, position_ids, past_key_value, use_cache, **kwargs)
+
+
+# =============================================================================
+# Compress Attention (NSA-subset, just compress branch)
+# =============================================================================
+class CompressAttention(nn.Module):
+    """Compress branch: reduce KV cache via local average, then full attention on compressed."""
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.compress_block = getattr(config, "compress_block_size", 16)
+        self.full_attn = FullAttention(config, layer_idx)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, dim = hidden_states.size()
+        # Compress along seq dim by averaging blocks
+        if q_len % self.compress_block == 0 and q_len > self.compress_block:
+            n_blocks = q_len // self.compress_block
+            x_comp = hidden_states.view(bsz, n_blocks, self.compress_block, dim).mean(dim=2)
+            pos_comp = position_ids[:, ::self.compress_block] if position_ids is not None else None
+            out_comp, kv = self.full_attn(x_comp, None, pos_comp, past_key_value, use_cache)
+            # Up-sample by repeat
+            out = out_comp.unsqueeze(2).expand(bsz, n_blocks, self.compress_block, dim).reshape(bsz, q_len, dim)
+            return out, kv
+        return self.full_attn(hidden_states, attention_mask, position_ids, past_key_value, use_cache)
+
+
+# =============================================================================
+# Hybrid Attention (NSA + Differential combined)
+# =============================================================================
+class HybridAttention(nn.Module):
+    """Combine NSA + Differential outputs via learnable gate + final norm (stable).
+
+    Fix v2 (2026-05-05): added post-merge GroupNorm + gate init=0 (sigmoid(0)=0.5 exact balance)
+    + lightly scaled output to prevent 49-layer cumulative divergence.
+    """
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.nsa = NSAAttention(config, layer_idx)
+        self.diff = DifferentialAttention(config, layer_idx)
+        # Per-channel gate (richer than scalar), init to 0 → sigmoid(0)=0.5 exact balance
+        self.gate = nn.Parameter(torch.zeros(config.hidden_size))
+        # Post-merge GroupNorm to stabilize 49-layer stack
+        n_groups = 8 if config.hidden_size % 8 == 0 else 1
+        self.merge_norm = nn.GroupNorm(num_groups=n_groups, num_channels=config.hidden_size, eps=1e-5)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        nsa_out, kv1 = self.nsa(hidden_states, attention_mask, position_ids, past_key_value, use_cache, **kwargs)
+        diff_out, kv2 = self.diff(hidden_states, attention_mask, position_ids, past_key_value, use_cache, **kwargs)
+        # Per-channel learnable mix: g (sigmoid) per channel
+        g = torch.sigmoid(self.gate)  # shape (hidden_size,)
+        out = g * nsa_out + (1.0 - g) * diff_out
+        # Stabilize via GroupNorm: (B, S, H) → (B, H, S) → norm → (B, S, H)
+        out = self.merge_norm(out.transpose(1, 2)).transpose(1, 2)
+        return out, kv1 if kv1 is not None else kv2
+
+
+# =============================================================================
+# 7-aware Attention Dispatcher
+# =============================================================================
+def build_attention(config: AETHERV211AttnConfig, layer_idx: int) -> nn.Module:
+    """Pick attention type based on Latin Square index."""
+    attn_type = get_attention_type(layer_idx)
+    if attn_type == "nsa":
+        return NSAAttention(config, layer_idx)
+    elif attn_type == "differential":
+        return DifferentialAttention(config, layer_idx)
+    elif attn_type == "full":
+        return FullAttention(config, layer_idx)
+    elif attn_type == "linear":
+        return LinearAttention(config, layer_idx)
+    elif attn_type == "sliding":
+        return SlidingWindowAttention(config, layer_idx)
+    elif attn_type == "compress":
+        return CompressAttention(config, layer_idx)
+    elif attn_type == "hybrid":
+        return HybridAttention(config, layer_idx)
+    elif attn_type == "mla":
+        from aether_pkg.v2_attentions.mla import MLAAttention
+        return MLAAttention(config, layer_idx)
+    elif attn_type == "mamba2":
+        from aether_pkg.v2_attentions.mamba2 import Mamba2Attention
+        return Mamba2Attention(config, layer_idx)
+    elif attn_type == "gdn":
+        from aether_pkg.v2_attentions.gdn import GDNAttention
+        return GDNAttention(config, layer_idx)
+    elif attn_type == "hyena":
+        from aether_pkg.v2_attentions.hyena import HyenaAttention
+        return HyenaAttention(config, layer_idx)
+    raise ValueError(f"Unknown attention type: {attn_type}")
+
+
+# =============================================================================
+# MoE Block: 25 experts, top-7 active per token
+# =============================================================================
+class AETHERV211AttnMLP(nn.Module):
+    """Single expert MLP (SwiGLU)."""
+
+    def __init__(self, config: AETHERV211AttnConfig, intermediate_size: Optional[int] = None):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size or config.expert_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)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class AETHERV211AttnSparseMoE(nn.Module):
+    """25-expert MoE with top-7 active routing.
+
+    Each layer has a 5-phase cyclic FFN bias to encode 오행 (Five Elements) cycle.
+    """
+
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.ffn_phase = get_ffn_phase(layer_idx)  # 0..4 (5-element cycle)
+
+        # Router: hidden → num_experts logits
+        self.gate = nn.Linear(self.hidden_size, self.num_experts, bias=False)
+
+        # 25 experts (each is a SwiGLU MLP)
+        self.experts = nn.ModuleList([
+            AETHERV211AttnMLP(config) for _ in range(self.num_experts)
+        ])
+
+        # 오행 cyclic phase bias (learnable, 5 phases)
+        self.phase_bias = nn.Parameter(torch.zeros(5, self.num_experts))
+
+        # Optional shared expert (always active, optional)
+        self.use_shared_expert = getattr(config, "use_shared_expert", True)
+        if self.use_shared_expert:
+            self.shared_expert = AETHERV211AttnMLP(
+                config, intermediate_size=config.expert_intermediate_size,
+            )
+            self.shared_expert_gate = nn.Linear(self.hidden_size, 1, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        bsz, seq_len, dim = hidden_states.shape
+        x = hidden_states.view(-1, dim)  # (bsz*seq, dim)
+
+        # Routing
+        router_logits = self.gate(x)  # (bsz*seq, num_experts)
+        # Add 5-phase cyclic bias (오행)
+        router_logits = router_logits + self.phase_bias[self.ffn_phase].unsqueeze(0)
+
+        # Top-k selection
+        routing_weights, selected_experts = torch.topk(router_logits, self.top_k, dim=-1)
+        routing_weights = F.softmax(routing_weights, dim=-1)
+
+        # Initialize output
+        final_out = torch.zeros_like(x)
+
+        # Per-expert dispatch (loop is OK for prototype; can be optimized with grouped GEMM later)
+        for e in range(self.num_experts):
+            mask = (selected_experts == e)
+            if not mask.any():
+                continue
+            token_idx, k_idx = mask.nonzero(as_tuple=True)
+            expert_in = x[token_idx]
+            expert_out = self.experts[e](expert_in)
+            weight = routing_weights[token_idx, k_idx].unsqueeze(-1).to(expert_out.dtype)
+            final_out.index_add_(0, token_idx, (expert_out * weight).to(final_out.dtype))
+
+        # Shared expert
+        if self.use_shared_expert:
+            shared_out = self.shared_expert(x)
+            shared_gate = torch.sigmoid(self.shared_expert_gate(x))
+            final_out = final_out + (shared_out * shared_gate).to(final_out.dtype)
+
+        final_out = final_out.view(bsz, seq_len, dim)
+        return final_out, router_logits.view(bsz, seq_len, self.num_experts)
+
+
+# =============================================================================
+# Decoder Layer: Attention + MoE FFN with 7-aware + 5-phase logic
+# =============================================================================
+class AETHERV211AttnDecoderLayer(nn.Module):
+    def __init__(self, config: AETHERV211AttnConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.attn_type = get_attention_type(layer_idx)
+        self.ffn_phase = get_ffn_phase(layer_idx)
+
+        # 7-aware attention (1 of 7 types based on Latin square)
+        self.self_attn = build_attention(config, layer_idx)
+
+        # MoE FFN with 5-phase cyclic bias
+        self.mlp = AETHERV211AttnSparseMoE(config, layer_idx)
+
+        # Norms
+        self.input_layernorm = AETHERV211AttnRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = AETHERV211AttnRMSNorm(self.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_router_logits: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[Cache], Optional[torch.Tensor]]:
+        # Self-attention with residual
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states, kv = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # MoE FFN with residual
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states, router_logits = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, kv)
+        if output_router_logits:
+            outputs = outputs + (router_logits,)
+        else:
+            outputs = outputs + (None,)
+        return outputs
+
+
+# =============================================================================
+# Pretrained base
+# =============================================================================
+class AETHERV211AttnPreTrainedModel(PreTrainedModel):
+    config_class = AETHERV211AttnConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["AETHERV211AttnDecoderLayer"]
+    _supports_cache_class = True
+    _supports_static_cache = False
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, AETHERV211AttnRMSNorm):
+            module.weight.data.fill_(1.0)
+
+
+# =============================================================================
+# Main Model
+# =============================================================================
+class AETHERV211AttnModel(AETHERV211AttnPreTrainedModel):
+    """49-layer decoder-only model with 7-aware attention + MoE."""
+
+    def __init__(self, config: AETHERV211AttnConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([
+            AETHERV211AttnDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)
+        ])
+        self.norm = AETHERV211AttnRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = output_hidden_states if output_hidden_states is not None else False
+        output_router_logits = output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("Cannot specify both input_ids and inputs_embeds")
+        if input_ids is not None:
+            bsz, seq_len = input_ids.shape
+        elif inputs_embeds is not None:
+            bsz, seq_len, _ = inputs_embeds.shape
+        else:
+            raise ValueError("Either input_ids or inputs_embeds must be provided")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        past_seen = past_key_values.get_seq_length() if past_key_values is not None else 0
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_seen, past_seen + seq_len, device=inputs_embeds.device,
+            ).unsqueeze(0)
+
+        hidden_states = inputs_embeds
+
+        all_hidden_states = () if output_hidden_states else None
+        all_router_logits = () if output_router_logits else None
+
+        for layer_idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_out = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states, attention_mask, position_ids,
+                    past_key_values, output_router_logits, use_cache,
+                )
+            else:
+                layer_out = decoder_layer(
+                    hidden_states=hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_out[0]
+            if output_router_logits:
+                all_router_logits += (layer_out[2],)
+
+        hidden_states = self.norm(hidden_states)
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [
+                hidden_states, past_key_values, all_hidden_states, None, all_router_logits,
+            ] if v is not None)
+
+        return MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=None,
+            router_logits=all_router_logits,
+        )
+
+
+# =============================================================================
+# Causal LM Wrapper
+# =============================================================================
+class AETHERV211AttnForCausalLM(AETHERV211AttnPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: AETHERV211AttnConfig):
+        super().__init__(config)
+        self.model = AETHERV211AttnModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.router_aux_loss_coef = getattr(config, "router_aux_loss_coef", 0.001)
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
+        output_router_logits = output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_hidden_states=output_hidden_states,
+            output_router_logits=output_router_logits,
+            return_dict=True,
+        )
+        hidden_states = outputs.last_hidden_state
+        logits = self.lm_head(hidden_states).float()
+
+        loss = None
+        aux_loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss = F.cross_entropy(
+                shift_logits.view(-1, self.vocab_size),
+                shift_labels.view(-1),
+                ignore_index=-100,
+            )
+
+        if output_router_logits and outputs.router_logits is not None:
+            aux_loss = self._compute_router_aux_loss(outputs.router_logits, attention_mask)
+            if loss is not None:
+                loss = loss + self.router_aux_loss_coef * aux_loss
+
+        if not return_dict:
+            output = (logits,) + tuple(v for v in [
+                outputs.past_key_values, outputs.hidden_states, None, outputs.router_logits, aux_loss,
+            ] if v is not None)
+            return (loss,) + output if loss is not None else output
+
+        return MoeCausalLMOutputWithPast(
+            loss=loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=None,
+            router_logits=outputs.router_logits,
+        )
+
+    def _compute_router_aux_loss(self, router_logits: Tuple[torch.Tensor, ...], attention_mask=None):
+        """Standard switch-transformer auxiliary loss for load balancing."""
+        if router_logits is None or len(router_logits) == 0:
+            return None
+        # Each router_logits[i] shape: (bsz, seq, num_experts) → flatten to (n_tokens, num_experts)
+        flat = []
+        for r in router_logits:
+            if r is None:
+                continue
+            flat.append(r.reshape(-1, self.num_experts))
+        if not flat:
+            return None
+        all_router_logits = torch.cat(flat, dim=0)  # (total_tokens, num_experts)
+
+        routing_weights = F.softmax(all_router_logits.float(), dim=-1)
+        _, selected_experts = torch.topk(routing_weights, self.num_experts_per_tok, dim=-1)
+
+        # Expert mask: (n_tokens, top_k, num_experts)
+        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts).float()
+        # Tokens-per-expert frequency: average over (n_tokens, top_k) dims → (num_experts,)
+        tokens_per_expert = expert_mask.mean(dim=(0, 1))
+        # Router prob per expert: (num_experts,)
+        router_prob_per_expert = routing_weights.mean(dim=0)
+        # aux_loss = num_experts * sum(token_freq * prob)
+        return self.num_experts * torch.sum(tokens_per_expert * router_prob_per_expert)
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        **kwargs,
+    ):
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values is not None:
+                position_ids = position_ids[:, -input_ids.shape[1]:]
+        return {
+            "input_ids": input_ids,
+            "position_ids": position_ids,
+            "past_key_values": past_key_values,
+            "use_cache": kwargs.get("use_cache"),
+            "attention_mask": attention_mask,
+        }
+
+
+# =============================================================================
+# Helper: Latin Square Layer Map (for analysis / debugging)
+# =============================================================================
+def print_layer_map(num_layers: int = 49):
+    """Print the Latin Square attention type map."""
+    print(f"=== AETHER-V2-7way Layer Map ({num_layers} layers) ===")
+    for L in range(num_layers):
+        attn = get_attention_type(L)
+        phase = get_ffn_phase(L)
+        row = L // 7
+        col = L % 7
+        print(f"  L{L:02d} (row={row} col={col}): attn={attn:12s} ffn_phase={phase}")
+
+
+__all__ = [
+    "AETHERV211AttnConfig",
+    "AETHERV211AttnModel",
+    "AETHERV211AttnForCausalLM",
+    "AETHERV211AttnPreTrainedModel",
+    "AETHERV211AttnDecoderLayer",
+    "AETHERV211AttnSparseMoE",
+    "build_attention",
+    "get_attention_type",
+    "get_ffn_phase",
+    "LATIN_SQUARE_7x7",
+    "ATTN_TYPES",
+    "print_layer_map",
+]

aether_pkg/v2_attentions/__init__.py

@@ -0,0 +1,10 @@
+# coding=utf-8
+"""
+V2 Attentions module — 7-aware attention implementations.
+
+⚠️ 동생 작성 placeholder
+"""
+from .nsa import NSAAttention
+from .differential import DifferentialAttention
+
+__all__ = ["NSAAttention", "DifferentialAttention"]

aether_pkg/v2_attentions/differential.py

@@ -0,0 +1,127 @@
+"""
+Differential Transformer (PRODUCTION - causal-safe)
+====================================================
+
+Microsoft 2410.05258 (ICLR 2025)
+Cancels noise via dual-softmax attention map subtraction.
+
+Core formula:
+  DiffAttn(X) = (softmax(Q1 K1^T / sqrt(d)) - λ × softmax(Q2 K2^T / sqrt(d))) V
+  λ = exp(λ_q1 · λ_k1) - exp(λ_q2 · λ_k2) + λ_init
+
+5/7 fixes applied to placeholder bugs:
+  - GroupNorm (sequence-spanning) → LayerNorm(head_dim) (per-token, causal-safe)
+  - Causal triu mask explicit (was relying on SDPA, but split + softmax direct now)
+  - forward signature: layer_idx kwarg + (Tensor, past_kv) tuple return for modeling compat
+
+Hyperparameter (configuration_aether_v2_7way.py):
+  - diff_lambda_init = 0.8
+"""
+from __future__ import annotations
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class DifferentialAttention(nn.Module):
+    """λ-gated dual-softmax differential attention with causal mask + LayerNorm."""
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.h = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = config.num_key_value_heads
+        self.head_dim = config.head_dim
+        # Differential splits head_dim in half
+        self.diff_head_dim = self.head_dim // 2
+
+        # Layer-depth lambda init (paper Eq. 3): λ_init = 0.8 - 0.6 * exp(-0.3 * (depth - 1))
+        # Use config base + depth decay if provided.
+        base = getattr(config, "diff_lambda_init", 0.8)
+        decay = getattr(config, "diff_lambda_layer_decay", 0.6)
+        # Simple per-layer init (placeholder used 0.8 - 0.6*exp(-0.3*idx))
+        self.lambda_init = base - decay * math.exp(-0.3 * max(layer_idx - 1, 0))
+
+        # Q/K/V/O projections
+        self.q_proj = nn.Linear(self.h, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.h, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.h, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.h, bias=False)
+
+        # Lambda parameters per-head
+        self.lambda_q1 = nn.Parameter(torch.zeros(self.num_heads, self.diff_head_dim).normal_(0, 0.1))
+        self.lambda_k1 = nn.Parameter(torch.zeros(self.num_heads, self.diff_head_dim).normal_(0, 0.1))
+        self.lambda_q2 = nn.Parameter(torch.zeros(self.num_heads, self.diff_head_dim).normal_(0, 0.1))
+        self.lambda_k2 = nn.Parameter(torch.zeros(self.num_heads, self.diff_head_dim).normal_(0, 0.1))
+
+        # ★ FIX 5/7: LayerNorm(head_dim) — per-token norm, causal-safe.
+        # Was nn.GroupNorm(num_groups=num_heads, num_channels=num_heads*head_dim) which
+        # normalized over sequence dim and leaked future info.
+        self.subln = nn.LayerNorm(self.head_dim, eps=getattr(config, "rms_norm_eps", 1e-6))
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value=None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[object]]:
+        B, S, _ = hidden_states.shape
+
+        q = self.q_proj(hidden_states).view(B, S, self.num_heads, self.head_dim)
+        k = self.k_proj(hidden_states).view(B, S, self.num_kv_heads, self.head_dim)
+        v = self.v_proj(hidden_states).view(B, S, self.num_kv_heads, self.head_dim)
+
+        # Split Q, K halves
+        q1, q2 = q.split(self.diff_head_dim, dim=-1)  # [B, S, H, d/2]
+        k1, k2 = k.split(self.diff_head_dim, dim=-1)  # [B, S, kv_h, d/2]
+
+        # GQA repeat
+        repeat = self.num_heads // self.num_kv_heads
+        k1 = k1.repeat_interleave(repeat, dim=2)
+        k2 = k2.repeat_interleave(repeat, dim=2)
+        v = v.repeat_interleave(repeat, dim=2)
+
+        # Transpose to [B, H, S, d]
+        q1 = q1.transpose(1, 2)
+        q2 = q2.transpose(1, 2)
+        k1 = k1.transpose(1, 2)
+        k2 = k2.transpose(1, 2)
+        v = v.transpose(1, 2)
+
+        scale = 1.0 / math.sqrt(self.diff_head_dim)
+        scores1 = torch.matmul(q1, k1.transpose(-2, -1)) * scale
+        scores2 = torch.matmul(q2, k2.transpose(-2, -1)) * scale
+
+        # ★ FIX 5/7: causal mask (triu(1) blocks future)
+        causal = torch.ones(S, S, device=q1.device, dtype=torch.bool).triu(1)
+        scores1 = scores1.masked_fill(causal, float("-inf"))
+        scores2 = scores2.masked_fill(causal, float("-inf"))
+
+        attn1 = F.softmax(scores1, dim=-1)
+        attn2 = F.softmax(scores2, dim=-1)
+
+        # Lambda per-head (paper Eq. 3)
+        lam_q1k1 = (self.lambda_q1 * self.lambda_k1).sum(dim=-1)  # [H]
+        lam_q2k2 = (self.lambda_q2 * self.lambda_k2).sum(dim=-1)
+        lam = torch.exp(lam_q1k1) - torch.exp(lam_q2k2) + self.lambda_init  # [H]
+        lam = lam.view(1, -1, 1, 1)
+
+        diff_attn = attn1 - lam * attn2
+
+        out = torch.matmul(diff_attn, v)  # [B, H, S, head_dim]
+        # ★ FIX 5/7: per-token LayerNorm (was sequence-spanning GroupNorm = leak)
+        out = self.subln(out)
+        out = out * (1 - self.lambda_init)
+
+        out = out.transpose(1, 2).reshape(B, S, -1)
+        return self.o_proj(out), past_key_value
+
+
+__all__ = ["DifferentialAttention"]

aether_pkg/v2_attentions/gdn.py

@@ -0,0 +1,201 @@
+# coding=utf-8
+# AETHER-30B-11Attn — Auto-extracted from new_attentions.py
+import math
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# =============================================================================
+# 1. MLA (Multi-head Latent Attention) — DeepSeek-V2/V3
+# =============================================================================
+class MLAAttention(nn.Module):
+    """Multi-head Latent Attention (low-rank Q/KV decomposition).
+
+    DeepSeek-V2 paper: Q/KV을 low-rank로 분해 → KV cache 1/14 절감.
+
+    Q: hidden → q_lora_rank → split (rope + nope) → multi-head
+    KV: hidden → kv_lora_rank → split (rope + nope) → K (rope+nope), V (nope)
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.q_lora_rank = getattr(config, "mla_q_lora_rank", 768)
+        self.kv_lora_rank = getattr(config, "mla_kv_lora_rank", 256)
+        self.qk_rope_head_dim = getattr(config, "mla_qk_rope_head_dim", 64)
+        self.qk_nope_head_dim = getattr(config, "mla_qk_nope_head_dim", 64)
+        self.v_head_dim = getattr(config, "mla_v_head_dim", 128)
+        qk_head_dim = self.qk_rope_head_dim + self.qk_nope_head_dim
+
+        # Q low-rank
+        self.q_a_proj = nn.Linear(self.hidden_size, self.q_lora_rank, bias=False)
+        self.q_a_norm = nn.LayerNorm(self.q_lora_rank, eps=1e-5)
+        self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * qk_head_dim, bias=False)
+
+        # KV low-rank (compressed)
+        self.kv_a_proj = nn.Linear(self.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False)
+        self.kv_a_norm = nn.LayerNorm(self.kv_lora_rank, eps=1e-5)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        # Output
+        self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, self.hidden_size, bias=False)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, _ = hidden_states.shape
+        qk_head_dim = self.qk_rope_head_dim + self.qk_nope_head_dim
+
+        # Q
+        q = self.q_a_proj(hidden_states)
+        q = self.q_a_norm(q)
+        q = self.q_b_proj(q).view(bsz, q_len, self.num_heads, qk_head_dim).transpose(1, 2)
+
+        # KV
+        kv = self.kv_a_proj(hidden_states)
+        kv_compressed, k_rope = kv.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        kv_compressed = self.kv_a_norm(kv_compressed)
+        kv_b = self.kv_b_proj(kv_compressed).view(
+            bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim
+        ).transpose(1, 2)
+        k_nope, v = kv_b.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        # Combine k (broadcast k_rope to all heads)
+        k_rope_expanded = k_rope.view(bsz, q_len, 1, self.qk_rope_head_dim).expand(
+            -1, -1, self.num_heads, -1
+        ).transpose(1, 2)
+        k = torch.cat([k_nope, k_rope_expanded], dim=-1)  # (bsz, n_h, seq, qk_head_dim)
+
+        # SDPA
+        attn_out = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=(attention_mask.to(q.dtype) if attention_mask is not None else None),
+            dropout_p=0.0, is_causal=(attention_mask is None and q_len > 1),
+        )
+        attn_out = attn_out.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(attn_out), past_key_value
+
+
+# =============================================================================
+# 2. Mamba2 — State Space Model (placeholder simplification)
+# =============================================================================
+class Mamba2Attention(nn.Module):
+    """Mamba2 SSM placeholder (simplified S6 selective scan).
+
+    True Mamba2: selective_scan_fn from mamba_ssm package.
+    Here: simplified GRU-style mixing as placeholder.
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.d_state = getattr(config, "mamba2_d_state", 128)
+        self.d_conv = getattr(config, "mamba2_d_conv", 4)
+        self.expand = getattr(config, "mamba2_expand", 2)
+        self.d_inner = self.expand * self.hidden_size
+
+        self.in_proj = nn.Linear(self.hidden_size, 2 * self.d_inner, bias=False)
+        self.conv1d = nn.Conv1d(self.d_inner, self.d_inner, kernel_size=self.d_conv,
+                                groups=self.d_inner, padding=self.d_conv - 1, bias=False)
+        # SSM params (simplified)
+        self.A_log = nn.Parameter(torch.zeros(self.d_inner, self.d_state))
+        self.D = nn.Parameter(torch.zeros(self.d_inner))
+        self.dt_proj = nn.Linear(self.d_inner, self.d_inner, bias=True)
+        self.B_proj = nn.Linear(self.d_inner, self.d_state, bias=False)
+        self.C_proj = nn.Linear(self.d_inner, self.d_state, bias=False)
+
+        self.out_proj = nn.Linear(self.d_inner, self.hidden_size, bias=False)
+        self.norm = nn.LayerNorm(self.d_inner, eps=1e-5)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, seq, _ = hidden_states.shape
+        # Project
+        xz = self.in_proj(hidden_states)
+        x, z = xz.chunk(2, dim=-1)  # (bsz, seq, d_inner) each
+
+        # 1D conv (causal)
+        x = x.transpose(1, 2)  # (bsz, d_inner, seq)
+        x = self.conv1d(x)[:, :, :seq]  # causal trim
+        x = F.silu(x).transpose(1, 2)  # (bsz, seq, d_inner)
+
+        # Simplified SSM: scan as cumulative recurrent
+        # For placeholder, we use the simple gated linear approximation
+        dt = F.softplus(self.dt_proj(x))  # (bsz, seq, d_inner)
+        out = x * dt + self.D[None, None, :] * x  # ultra-simplified
+        out = self.norm(out)
+        out = out * F.silu(z)  # gating
+        out = self.out_proj(out)
+        return out, past_key_value
+
+
+# =============================================================================
+# 3. GDN (Gated Delta Network) — RWKV/RetNet style
+# =============================================================================
+class GDNAttention(nn.Module):
+    """Gated Delta Network — RWKV/RetNet style linear attention with delta gating.
+
+    Placeholder simplified: linear attention + gating.
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.expand = getattr(config, "gdn_expand", 1)
+        self.d_inner = self.expand * self.hidden_size
+        self.d_conv = getattr(config, "gdn_d_conv", 4)
+
+        self.q_proj = nn.Linear(self.hidden_size, self.d_inner, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.d_inner, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.d_inner, bias=False)
+        self.gate_proj = nn.Linear(self.hidden_size, self.d_inner, bias=False)
+
+        self.conv1d_q = nn.Conv1d(self.d_inner, self.d_inner, kernel_size=self.d_conv,
+                                  groups=self.d_inner, padding=self.d_conv - 1, bias=False)
+        self.conv1d_k = nn.Conv1d(self.d_inner, self.d_inner, kernel_size=self.d_conv,
+                                  groups=self.d_inner, padding=self.d_conv - 1, bias=False)
+
+        self.out_proj = nn.Linear(self.d_inner, self.hidden_size, bias=False)
+        self.norm = nn.LayerNorm(self.d_inner, eps=1e-5)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, seq, _ = hidden_states.shape
+        q = self.q_proj(hidden_states)
+        k = self.k_proj(hidden_states)
+        v = self.v_proj(hidden_states)
+        g = F.silu(self.gate_proj(hidden_states))
+
+        # Causal conv on q, k
+        q = self.conv1d_q(q.transpose(1, 2))[:, :, :seq].transpose(1, 2)
+        k = self.conv1d_k(k.transpose(1, 2))[:, :, :seq].transpose(1, 2)
+
+        # Linear attention: (Q · K^T) → causal cumsum → · V
+        q = F.elu(q) + 1
+        k = F.elu(k) + 1
+
+        # Simplified linear attention (per-token)
+        # For truly causal: cumulative outer products. We use simplified scan.
+        kv = k * v  # element-wise (placeholder)
+        kv_cum = torch.cumsum(kv, dim=1)
+        out = q * kv_cum
+        out = self.norm(out) * g
+        out = self.out_proj(out)
+        return out, past_key_value
+
+
+# =============================================================================
+# 4. Hyena — Convolutional alternative (no attention)
+
+__all__ = ['GDNAttention']

aether_pkg/v2_attentions/hyena.py

@@ -0,0 +1,62 @@
+# coding=utf-8
+# AETHER-30B-11Attn — Auto-extracted from new_attentions.py
+import math
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class HyenaAttention(nn.Module):
+    """Hyena — convolutional sequence mixer (Stanford 2023).
+
+    Long-range attention replaced with implicit conv filters.
+    Placeholder simplified.
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.filter_order = getattr(config, "hyena_filter_order", 64)
+        self.dropout = getattr(config, "hyena_dropout", 0.0)
+
+        # Input projections (3 streams: u, v, x)
+        self.in_proj = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
+
+        # Implicit filter (positional encoding-based)
+        self.filter_dense = nn.Sequential(
+            nn.Linear(1, self.filter_order),
+            nn.GELU(),
+            nn.Linear(self.filter_order, self.hidden_size),
+        )
+
+        # Output gate
+        self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.norm = nn.LayerNorm(self.hidden_size, eps=1e-5)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, seq, h = hidden_states.shape
+        # Project to 3 streams
+        proj = self.in_proj(hidden_states)
+        u, v, x = proj.chunk(3, dim=-1)
+
+        # Build implicit filter (length seq)
+        positions = torch.linspace(0, 1, seq, device=hidden_states.device).unsqueeze(-1)  # (seq, 1)
+        filt = self.filter_dense(positions)  # (seq, h)
+        filt = filt.transpose(0, 1).unsqueeze(0)  # (1, h, seq)
+
+        # 1D depthwise conv with implicit filter (placeholder: simple multiply)
+        # True Hyena uses FFT-based long conv. Placeholder:
+        u_t = u.transpose(1, 2)  # (bsz, h, seq)
+        # cumulative average + gate (simplified)
+        u_cum = torch.cumsum(u_t, dim=-1) / torch.arange(1, seq + 1, device=u.device, dtype=u.dtype).view(1, 1, -1)
+        out = u_cum.transpose(1, 2) * v + x
+        out = self.norm(out)
+        out = self.out_proj(out)
+        return out, past_key_value
+
+
+
+__all__ = ['HyenaAttention']

aether_pkg/v2_attentions/mamba2.py

@@ -0,0 +1,144 @@
+# coding=utf-8
+# AETHER-30B-11Attn — Auto-extracted from new_attentions.py
+import math
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# =============================================================================
+# 1. MLA (Multi-head Latent Attention) — DeepSeek-V2/V3
+# =============================================================================
+class MLAAttention(nn.Module):
+    """Multi-head Latent Attention (low-rank Q/KV decomposition).
+
+    DeepSeek-V2 paper: Q/KV을 low-rank로 분해 → KV cache 1/14 절감.
+
+    Q: hidden → q_lora_rank → split (rope + nope) → multi-head
+    KV: hidden → kv_lora_rank → split (rope + nope) → K (rope+nope), V (nope)
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.q_lora_rank = getattr(config, "mla_q_lora_rank", 768)
+        self.kv_lora_rank = getattr(config, "mla_kv_lora_rank", 256)
+        self.qk_rope_head_dim = getattr(config, "mla_qk_rope_head_dim", 64)
+        self.qk_nope_head_dim = getattr(config, "mla_qk_nope_head_dim", 64)
+        self.v_head_dim = getattr(config, "mla_v_head_dim", 128)
+        qk_head_dim = self.qk_rope_head_dim + self.qk_nope_head_dim
+
+        # Q low-rank
+        self.q_a_proj = nn.Linear(self.hidden_size, self.q_lora_rank, bias=False)
+        self.q_a_norm = nn.LayerNorm(self.q_lora_rank, eps=1e-5)
+        self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * qk_head_dim, bias=False)
+
+        # KV low-rank (compressed)
+        self.kv_a_proj = nn.Linear(self.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False)
+        self.kv_a_norm = nn.LayerNorm(self.kv_lora_rank, eps=1e-5)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        # Output
+        self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, self.hidden_size, bias=False)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, _ = hidden_states.shape
+        qk_head_dim = self.qk_rope_head_dim + self.qk_nope_head_dim
+
+        # Q
+        q = self.q_a_proj(hidden_states)
+        q = self.q_a_norm(q)
+        q = self.q_b_proj(q).view(bsz, q_len, self.num_heads, qk_head_dim).transpose(1, 2)
+
+        # KV
+        kv = self.kv_a_proj(hidden_states)
+        kv_compressed, k_rope = kv.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        kv_compressed = self.kv_a_norm(kv_compressed)
+        kv_b = self.kv_b_proj(kv_compressed).view(
+            bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim
+        ).transpose(1, 2)
+        k_nope, v = kv_b.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        # Combine k (broadcast k_rope to all heads)
+        k_rope_expanded = k_rope.view(bsz, q_len, 1, self.qk_rope_head_dim).expand(
+            -1, -1, self.num_heads, -1
+        ).transpose(1, 2)
+        k = torch.cat([k_nope, k_rope_expanded], dim=-1)  # (bsz, n_h, seq, qk_head_dim)
+
+        # SDPA
+        attn_out = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=(attention_mask.to(q.dtype) if attention_mask is not None else None),
+            dropout_p=0.0, is_causal=(attention_mask is None and q_len > 1),
+        )
+        attn_out = attn_out.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(attn_out), past_key_value
+
+
+# =============================================================================
+# 2. Mamba2 — State Space Model (placeholder simplification)
+# =============================================================================
+class Mamba2Attention(nn.Module):
+    """Mamba2 SSM placeholder (simplified S6 selective scan).
+
+    True Mamba2: selective_scan_fn from mamba_ssm package.
+    Here: simplified GRU-style mixing as placeholder.
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.d_state = getattr(config, "mamba2_d_state", 128)
+        self.d_conv = getattr(config, "mamba2_d_conv", 4)
+        self.expand = getattr(config, "mamba2_expand", 2)
+        self.d_inner = self.expand * self.hidden_size
+
+        self.in_proj = nn.Linear(self.hidden_size, 2 * self.d_inner, bias=False)
+        self.conv1d = nn.Conv1d(self.d_inner, self.d_inner, kernel_size=self.d_conv,
+                                groups=self.d_inner, padding=self.d_conv - 1, bias=False)
+        # SSM params (simplified)
+        self.A_log = nn.Parameter(torch.zeros(self.d_inner, self.d_state))
+        self.D = nn.Parameter(torch.zeros(self.d_inner))
+        self.dt_proj = nn.Linear(self.d_inner, self.d_inner, bias=True)
+        self.B_proj = nn.Linear(self.d_inner, self.d_state, bias=False)
+        self.C_proj = nn.Linear(self.d_inner, self.d_state, bias=False)
+
+        self.out_proj = nn.Linear(self.d_inner, self.hidden_size, bias=False)
+        self.norm = nn.LayerNorm(self.d_inner, eps=1e-5)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, seq, _ = hidden_states.shape
+        # Project
+        xz = self.in_proj(hidden_states)
+        x, z = xz.chunk(2, dim=-1)  # (bsz, seq, d_inner) each
+
+        # 1D conv (causal)
+        x = x.transpose(1, 2)  # (bsz, d_inner, seq)
+        x = self.conv1d(x)[:, :, :seq]  # causal trim
+        x = F.silu(x).transpose(1, 2)  # (bsz, seq, d_inner)
+
+        # Simplified SSM: scan as cumulative recurrent
+        # For placeholder, we use the simple gated linear approximation
+        dt = F.softplus(self.dt_proj(x))  # (bsz, seq, d_inner)
+        out = x * dt + self.D[None, None, :] * x  # ultra-simplified
+        out = self.norm(out)
+        out = out * F.silu(z)  # gating
+        out = self.out_proj(out)
+        return out, past_key_value
+
+
+# =============================================================================
+# 3. GDN (Gated Delta Network) — RWKV/RetNet style
+
+__all__ = ['Mamba2Attention']

aether_pkg/v2_attentions/mla.py

@@ -0,0 +1,89 @@
+# coding=utf-8
+# AETHER-30B-11Attn — Auto-extracted from new_attentions.py
+import math
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# =============================================================================
+# 1. MLA (Multi-head Latent Attention) — DeepSeek-V2/V3
+# =============================================================================
+class MLAAttention(nn.Module):
+    """Multi-head Latent Attention (low-rank Q/KV decomposition).
+
+    DeepSeek-V2 paper: Q/KV을 low-rank로 분해 → KV cache 1/14 절감.
+
+    Q: hidden → q_lora_rank → split (rope + nope) → multi-head
+    KV: hidden → kv_lora_rank → split (rope + nope) → K (rope+nope), V (nope)
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.q_lora_rank = getattr(config, "mla_q_lora_rank", 768)
+        self.kv_lora_rank = getattr(config, "mla_kv_lora_rank", 256)
+        self.qk_rope_head_dim = getattr(config, "mla_qk_rope_head_dim", 64)
+        self.qk_nope_head_dim = getattr(config, "mla_qk_nope_head_dim", 64)
+        self.v_head_dim = getattr(config, "mla_v_head_dim", 128)
+        qk_head_dim = self.qk_rope_head_dim + self.qk_nope_head_dim
+
+        # Q low-rank
+        self.q_a_proj = nn.Linear(self.hidden_size, self.q_lora_rank, bias=False)
+        self.q_a_norm = nn.LayerNorm(self.q_lora_rank, eps=1e-5)
+        self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * qk_head_dim, bias=False)
+
+        # KV low-rank (compressed)
+        self.kv_a_proj = nn.Linear(self.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False)
+        self.kv_a_norm = nn.LayerNorm(self.kv_lora_rank, eps=1e-5)
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+
+        # Output
+        self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, self.hidden_size, bias=False)
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None,
+                past_key_value=None, use_cache=False, **kwargs):
+        bsz, q_len, _ = hidden_states.shape
+        qk_head_dim = self.qk_rope_head_dim + self.qk_nope_head_dim
+
+        # Q
+        q = self.q_a_proj(hidden_states)
+        q = self.q_a_norm(q)
+        q = self.q_b_proj(q).view(bsz, q_len, self.num_heads, qk_head_dim).transpose(1, 2)
+
+        # KV
+        kv = self.kv_a_proj(hidden_states)
+        kv_compressed, k_rope = kv.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        kv_compressed = self.kv_a_norm(kv_compressed)
+        kv_b = self.kv_b_proj(kv_compressed).view(
+            bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim
+        ).transpose(1, 2)
+        k_nope, v = kv_b.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+
+        # Combine k (broadcast k_rope to all heads)
+        k_rope_expanded = k_rope.view(bsz, q_len, 1, self.qk_rope_head_dim).expand(
+            -1, -1, self.num_heads, -1
+        ).transpose(1, 2)
+        k = torch.cat([k_nope, k_rope_expanded], dim=-1)  # (bsz, n_h, seq, qk_head_dim)
+
+        # SDPA
+        attn_out = F.scaled_dot_product_attention(
+            q, k, v,
+            attn_mask=(attention_mask.to(q.dtype) if attention_mask is not None else None),
+            dropout_p=0.0, is_causal=(attention_mask is None and q_len > 1),
+        )
+        attn_out = attn_out.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(attn_out), past_key_value
+
+
+# =============================================================================
+# 2. Mamba2 — State Space Model (placeholder simplification)
+
+__all__ = ['MLAAttention']

aether_pkg/v2_attentions/nsa.py

@@ -0,0 +1,149 @@
+# coding=utf-8
+"""
+Native Sparse Attention (NSA) — 3-branch design.
+
+⚠️ 동생 작성 placeholder (사용자님 본 PC 정식 버전 도착 시 교체)
+
+Branches:
+  1. Compress: average over blocks → KV cache reduction
+  2. Select: top-k token selection per query
+  3. Sliding: local window attention
+  4. Gate: learnable combination of 3 branches
+"""
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class NSAAttention(nn.Module):
+    """3-branch Native Sparse Attention.
+
+    For testing purposes. Production version 사용자님 본 PC.
+    """
+
+    def __init__(self, config, layer_idx: int = 0):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = getattr(config, "num_key_value_heads", config.num_attention_heads)
+        self.head_dim = config.head_dim
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+        self.compress_block = getattr(config, "compress_block_size", 16)
+        self.sliding_window = getattr(config, "sliding_window_size", 256)
+        self.select_topk = getattr(config, "nsa_select_topk", 16)
+
+        # Shared Q/K/V projections (3 branches share base projections)
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+        # Compress branch: pooling MLP
+        self.compress_mlp = nn.Linear(self.head_dim, self.head_dim, bias=False)
+
+        # Select branch: scoring head (project to scalar score per token)
+        self.select_score = nn.Linear(self.head_dim, 1, bias=False)
+
+        # 3-branch gate (initialized at 0.5 each, will be normalized)
+        self.gate_logit = nn.Parameter(torch.zeros(3))  # sigmoid → ~0.5
+
+    def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
+        """GQA: repeat KV heads to match Q heads."""
+        if self.num_kv_groups == 1:
+            return x
+        bsz, n_kv, seq, dim = x.shape
+        return (x[:, :, None, :, :]
+                .expand(bsz, n_kv, self.num_kv_groups, seq, dim)
+                .reshape(bsz, n_kv * self.num_kv_groups, seq, dim))
+
+    def _compress_branch(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+        """Compress branch: average KV in blocks, then full attention on compressed."""
+        bsz, n_h, seq, dim = q.shape
+        if seq < self.compress_block * 2:
+            return self._full_attention(q, k, v)
+
+        n_blocks = seq // self.compress_block
+        # Compress K, V by averaging blocks (only complete blocks)
+        usable = n_blocks * self.compress_block
+        k_comp = k[:, :, :usable, :].view(bsz, n_h, n_blocks, self.compress_block, dim).mean(dim=3)
+        v_comp = v[:, :, :usable, :].view(bsz, n_h, n_blocks, self.compress_block, dim).mean(dim=3)
+        # Apply MLP to compressed K
+        k_comp = self.compress_mlp(k_comp)
+        # Full attention against compressed
+        return F.scaled_dot_product_attention(q, k_comp, v_comp, dropout_p=0.0, is_causal=False)
+
+    def _select_branch(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+        """Select branch: top-k token selection per query position."""
+        bsz, n_h, seq, dim = q.shape
+        if seq <= self.select_topk:
+            return self._full_attention(q, k, v)
+
+        # Score each KV position via select_score on K
+        scores = self.select_score(k).squeeze(-1)  # (bsz, n_h, seq)
+        topk = min(self.select_topk, seq)
+        _, top_idx = scores.topk(topk, dim=-1)  # (bsz, n_h, topk)
+
+        # Gather top-k K, V (causal mask: only attend to past)
+        idx_expand = top_idx.unsqueeze(-1).expand(-1, -1, -1, dim)
+        k_sel = k.gather(2, idx_expand)
+        v_sel = v.gather(2, idx_expand)
+
+        return F.scaled_dot_product_attention(q, k_sel, v_sel, dropout_p=0.0, is_causal=False)
+
+    def _sliding_branch(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+        """Sliding window: only attend to local window."""
+        bsz, n_h, seq, dim = q.shape
+        if seq <= self.sliding_window:
+            return self._full_attention(q, k, v)
+
+        # Build sliding window mask
+        idx = torch.arange(seq, device=q.device)
+        rel = idx.unsqueeze(0) - idx.unsqueeze(1)  # (seq, seq)
+        mask = (rel >= 0) & (rel <= self.sliding_window)  # local + causal
+        attn_mask = torch.where(mask, 0.0, float("-inf")).unsqueeze(0).unsqueeze(0).to(q.dtype)
+        return F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=0.0)
+
+    def _full_attention(self, q, k, v):
+        return F.scaled_dot_product_attention(q, k, v, dropout_p=0.0, is_causal=True)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value=None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[object]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        q = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+
+        # GQA: repeat KV to match Q heads
+        k = self._repeat_kv(k)
+        v = self._repeat_kv(v)
+
+        # 3 branches
+        out_compress = self._compress_branch(q, k, v)
+        out_select = self._select_branch(q, k, v)
+        out_sliding = self._sliding_branch(q, k, v)
+
+        # Gate (softmax over 3 branches)
+        gate = F.softmax(self.gate_logit, dim=0)
+        out = (gate[0] * out_compress
+               + gate[1] * out_select
+               + gate[2] * out_sliding)
+
+        # Reshape and project
+        out = out.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(out), past_key_value
+
+
+__all__ = ["NSAAttention"]

config.json

@@ -0,0 +1,59 @@
+{
+  "attention_dropout": 0.0,
+  "compress_block_size": 32,
+  "deltanet_chunk_size": 128,
+  "deltanet_d_state": 64,
+  "diff_lambda_init": 0.8,
+  "diff_lambda_layer_decay": 0.6,
+  "expert_intermediate_size": 512,
+  "gdn_chunk_size": 64,
+  "gdn_d_conv": 4,
+  "gdn_expand": 1,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "hybrid_groupnorm_groups": 8,
+  "hyena_dropout": 0.0,
+  "hyena_filter_order": 64,
+  "initializer_range": 0.02,
+  "intermediate_size": 512,
+  "linear_d_conv": 4,
+  "linear_d_state": 64,
+  "linear_expand": 1,
+  "mamba2_chunk_size": 256,
+  "mamba2_d_conv": 4,
+  "mamba2_d_state": 128,
+  "mamba2_expand": 2,
+  "mamba2_headdim": 64,
+  "max_position_embeddings": 8192,
+  "mla_kv_lora_rank": 256,
+  "mla_q_lora_rank": 768,
+  "mla_qk_nope_head_dim": 64,
+  "mla_qk_rope_head_dim": 64,
+  "mla_v_head_dim": 128,
+  "model_type": "aether_v2_11attn",
+  "nsa_compress_block_size": 32,
+  "nsa_select_top_k": 16,
+  "nsa_sliding_window": 512,
+  "num_attention_heads": 8,
+  "num_experts": 25,
+  "num_experts_per_tok": 7,
+  "num_hidden_layers": 121,
+  "num_key_value_heads": 2,
+  "oheng_enable": true,
+  "oheng_num_phases": 5,
+  "output_router_logits": false,
+  "pad_token_id": 151643,
+  "rezero_alpha_init": 0.0,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "router_aux_loss_coef": 0.01,
+  "sliding_window_size": 1024,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.3.0",
+  "use_cache": false,
+  "use_deepnet_init": true,
+  "use_shared_expert": true,
+  "vocab_size": 151936,
+  "z_loss_coef": 0.001
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6181fbdcf6df034880135bcd7f697aa01563727e1a75b43b51a1f2c2be2264d
+size 11578322830

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650

tokenizer_config.json

@@ -0,0 +1,29 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}