| """ループドTransformer本体(HF ``PreTrainedModel`` 互換・単一実装)。 |
| |
| 設計 docs/architecture.md §2 の Llama 系レシピ(RMSNorm / RoPE / SwiGLU / |
| bias なし / weight tying)を素の PyTorch で実装。``forward`` は K 回ループする |
| 形で書き、``k=1`` で標準Transformerに厳密に縮退する(``tests/test_k1_equivalence.py``)。 |
| |
| このファイルは ``save_pretrained`` 時に checkpoint へ複製され、公式 |
| evaluation-pipeline(別プロセス・``trust_remote_code=True``)から import される。 |
| そのため **torch / transformers / 標準ライブラリ以外に依存しない**こと。 |
| probing 用のループ毎中間表現は HF 標準 ``hidden_states``(層ごと)と混ぜず、 |
| 別フィールド ``loop_hidden_states`` に格納する。 |
| """ |
|
|
| from __future__ import annotations |
|
|
| import math |
| from dataclasses import dataclass |
|
|
| import torch |
| import torch.nn.functional as F |
| from torch import nn |
| from transformers.modeling_outputs import ModelOutput |
| from transformers.modeling_utils import PreTrainedModel |
|
|
| from .configuration_babyloop import BabyloopConfig |
|
|
|
|
| |
|
|
|
|
| class RMSNorm(nn.Module): |
| def __init__(self, dim: int, eps: float): |
| super().__init__() |
| self.eps = eps |
| self.weight = nn.Parameter(torch.ones(dim)) |
|
|
| def forward(self, x: torch.Tensor) -> torch.Tensor: |
| dtype = x.dtype |
| x = x.float() |
| x = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) |
| return (x.to(dtype)) * self.weight |
|
|
|
|
| def _rotate_half(x: torch.Tensor) -> torch.Tensor: |
| x1, x2 = x.chunk(2, dim=-1) |
| return torch.cat((-x2, x1), dim=-1) |
|
|
|
|
| def _apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor: |
| |
| return x * cos + _rotate_half(x) * sin |
|
|
|
|
| class Attention(nn.Module): |
| """RoPE 付き causal Multi-Head Attention(dropout なし)。""" |
|
|
| def __init__(self, config: BabyloopConfig): |
| super().__init__() |
| self.n_heads = config.n_heads |
| self.head_dim = config.d_model // config.n_heads |
| self.qkv_proj = nn.Linear(config.d_model, 3 * config.d_model, bias=config.bias) |
| self.o_proj = nn.Linear(config.d_model, config.d_model, bias=config.bias) |
|
|
| def forward(self, x, cos, sin, attn_bias): |
| B, T, C = x.shape |
| q, k, v = self.qkv_proj(x).split(C, dim=-1) |
| q = q.view(B, T, self.n_heads, self.head_dim).transpose(1, 2) |
| k = k.view(B, T, self.n_heads, self.head_dim).transpose(1, 2) |
| v = v.view(B, T, self.n_heads, self.head_dim).transpose(1, 2) |
| q = _apply_rope(q, cos, sin) |
| k = _apply_rope(k, cos, sin) |
| out = F.scaled_dot_product_attention( |
| q, k, v, attn_mask=attn_bias, is_causal=attn_bias is None |
| ) |
| out = out.transpose(1, 2).reshape(B, T, C) |
| return self.o_proj(out) |
|
|
|
|
| class SwiGLU(nn.Module): |
| def __init__(self, config: BabyloopConfig): |
| super().__init__() |
| self.gate_proj = nn.Linear(config.d_model, config.ffn_hidden, bias=config.bias) |
| self.up_proj = nn.Linear(config.d_model, config.ffn_hidden, bias=config.bias) |
| self.down_proj = nn.Linear(config.ffn_hidden, config.d_model, bias=config.bias) |
|
|
| def forward(self, x): |
| return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x)) |
|
|
|
|
| class Block(nn.Module): |
| """Pre-Norm 残差ブロック: h += Attn(RMSNorm(h)); h += SwiGLU(RMSNorm(h))。""" |
|
|
| def __init__(self, config: BabyloopConfig): |
| super().__init__() |
| self.attn_norm = RMSNorm(config.d_model, config.rms_eps) |
| self.attn = Attention(config) |
| self.mlp_norm = RMSNorm(config.d_model, config.rms_eps) |
| self.mlp = SwiGLU(config) |
|
|
| def forward(self, h, cos, sin, attn_bias): |
| h = h + self.attn(self.attn_norm(h), cos, sin, attn_bias) |
| h = h + self.mlp(self.mlp_norm(h)) |
| return h |
|
|
|
|
| |
| |
|
|
|
|
| class VisualConnector(nn.Module): |
| """事前計算済み視覚特徴 (B, V, vision_feature_dim) を d_model 空間へ射影する。 |
| |
| - ``mlp``: 2層 MLP(vit_dim→d_model→d_model, GELU。LLaVA-1.5 流)。②の本命・trainable。 |
| - ``identity``: 射影なし(``vit_dim==d_model`` 前提)。射影なし+未事前学習の劣化条件 |
| (ADR-0006 の naive)。学習パラメータを持たない。 |
| """ |
|
|
| def __init__(self, config: BabyloopConfig): |
| super().__init__() |
| self.connector_type = config.connector_type |
| vit_dim = config.vision_feature_dim |
| d = config.d_model |
| if self.connector_type == "mlp": |
| |
| self.fc1 = nn.Linear(vit_dim, d, bias=True) |
| self.act = nn.GELU() |
| self.fc2 = nn.Linear(d, d, bias=True) |
| elif self.connector_type == "identity": |
| if vit_dim != d: |
| raise ValueError( |
| f"connector_type=identity は vision_feature_dim==d_model が前提 " |
| f"({vit_dim} != {d})。ViT-B/14(768) を使うか mlp connector にする。" |
| ) |
| else: |
| raise ValueError(f"unknown connector_type: {self.connector_type}") |
|
|
| def forward(self, vision_features: torch.Tensor) -> torch.Tensor: |
| if self.connector_type == "identity": |
| return vision_features |
| return self.fc2(self.act(self.fc1(vision_features))) |
|
|
|
|
| |
|
|
|
|
| @dataclass |
| class LoopedModelOutput(ModelOutput): |
| last_hidden_state: torch.FloatTensor | None = None |
| hidden_states: tuple[torch.FloatTensor, ...] | None = None |
| loop_hidden_states: tuple[torch.FloatTensor, ...] | None = None |
|
|
|
|
| @dataclass |
| class LoopedCausalLMOutput(ModelOutput): |
| loss: torch.FloatTensor | None = None |
| logits: torch.FloatTensor | None = None |
| hidden_states: tuple[torch.FloatTensor, ...] | None = None |
| loop_hidden_states: tuple[torch.FloatTensor, ...] | None = None |
|
|
|
|
| |
|
|
|
|
| class LoopedPreTrainedModel(PreTrainedModel): |
| config_class = BabyloopConfig |
| base_model_prefix = "model" |
| supports_gradient_checkpointing = False |
|
|
| def _init_weights(self, module): |
| std = 0.02 |
| if isinstance(module, nn.Linear): |
| nn.init.normal_(module.weight, mean=0.0, std=std) |
| if module.bias is not None: |
| nn.init.zeros_(module.bias) |
| elif isinstance(module, nn.Embedding): |
| nn.init.normal_(module.weight, mean=0.0, std=std) |
| elif isinstance(module, RMSNorm): |
| nn.init.ones_(module.weight) |
|
|
|
|
| class LoopedModel(LoopedPreTrainedModel): |
| """重み共有 core ブロックを K 回反復するバックボーン(lm_head なし)。""" |
|
|
| def __init__(self, config: BabyloopConfig): |
| super().__init__(config) |
| self.config = config |
| self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model) |
| self.prelude = nn.ModuleList(Block(config) for _ in range(config.n_prelude)) |
| self.core = nn.ModuleList(Block(config) for _ in range(config.n_core)) |
| self.coda = nn.ModuleList(Block(config) for _ in range(config.n_coda)) |
| self.final_norm = RMSNorm(config.d_model, config.rms_eps) |
|
|
| |
| |
| self.connector = VisualConnector(config) if config.fusion is not None else None |
| if self.connector is not None and config.visual_inject_mode not in ("prefix_refresh", "broadcast"): |
| raise ValueError(f"unknown visual_inject_mode: {config.visual_inject_mode}") |
|
|
| head_dim = config.d_model // config.n_heads |
| inv_freq = 1.0 / ( |
| config.rope_base ** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim) |
| ) |
| self.register_buffer("inv_freq", inv_freq, persistent=False) |
| self.post_init() |
|
|
| def get_input_embeddings(self): |
| return self.embed_tokens |
|
|
| def set_input_embeddings(self, value): |
| self.embed_tokens = value |
|
|
| def _rope(self, T: int, device, dtype): |
| t = torch.arange(T, device=device, dtype=torch.float32) |
| freqs = torch.outer(t, self.inv_freq.to(device)) |
| emb = torch.cat((freqs, freqs), dim=-1) |
| return emb.cos().to(dtype)[None, None], emb.sin().to(dtype)[None, None] |
|
|
| def _attn_bias(self, attention_mask, T, device, dtype): |
| |
| if attention_mask is None or bool((attention_mask == 1).all()): |
| return None |
| causal = torch.ones(T, T, device=device, dtype=torch.bool).triu(1) |
| key_pad = attention_mask.to(device) == 0 |
| mask = causal[None, None] | key_pad[:, None, None, :] |
| bias = torch.zeros(mask.shape, device=device, dtype=dtype) |
| return bias.masked_fill(mask, torch.finfo(dtype).min) |
|
|
| def inject(self, h, visual, loop_index: int): |
| """④: ループ初期の各イテレーション先頭で視覚を再注入(``visual_inject_iters>0``)。 |
| |
| Args: |
| h: 現在の hidden states。``prefix_refresh`` では prefix 済 (B, V+T, d)、 |
| ``broadcast`` では (B, T, d)。 |
| visual: connector 射影済みの視覚特徴 (B, V, d)。n_visual は ``visual.size(1)`` から導出 |
| (pad 量の不一致バグを封じるため別経路で渡さない)。 |
| loop_index: 現在のループ回(0始まり、呼び出し側で ``< inject_iters`` を保証)。 |
| |
| - ``prefix_refresh``(B 主系): prefix 位置(先頭 V)へ視覚を再加算(右ゼロパディングして加算)。 |
| ②と同じ入れ方・同じ特徴で、変換で薄れた視覚信号を補充する([ADR-0006])。 |
| - ``broadcast``(A, ablation): pooled 視覚ベクトルを全位置へブロードキャスト加算。系列長不変。 |
| """ |
| if self.config.visual_inject_mode == "broadcast": |
| return h + visual.mean(dim=1, keepdim=True) |
| |
| pad = h.size(1) - visual.size(1) |
| return h + F.pad(visual, (0, 0, 0, pad)) |
|
|
| def forward( |
| self, |
| input_ids=None, |
| attention_mask=None, |
| inputs_embeds=None, |
| vision_features=None, |
| output_hidden_states=False, |
| **kwargs, |
| ) -> LoopedModelOutput: |
| if inputs_embeds is None: |
| inputs_embeds = self.embed_tokens(input_ids) |
|
|
| |
| |
| |
| |
| inject_iters = self.config.visual_inject_iters |
| mode = self.config.visual_inject_mode |
| visual = None |
| if vision_features is not None and self.connector is not None: |
| visual = self.connector(vision_features.to(inputs_embeds.dtype)).to(inputs_embeds.dtype) |
|
|
| |
| use_prefix = visual is not None and (inject_iters == 0 or mode == "prefix_refresh") |
| n_visual = 0 |
| if use_prefix: |
| n_visual = visual.size(1) |
| inputs_embeds = torch.cat([visual, inputs_embeds], dim=1) |
| if attention_mask is not None: |
| visual_mask = attention_mask.new_ones((attention_mask.size(0), n_visual)) |
| attention_mask = torch.cat([visual_mask, attention_mask], dim=1) |
|
|
| h = inputs_embeds |
| |
| |
| |
| residual_input = inputs_embeds |
| if n_visual > 0: |
| residual_input = F.pad(inputs_embeds[:, n_visual:], (0, 0, n_visual, 0)) |
| B, T, _ = h.shape |
| cos, sin = self._rope(T, h.device, h.dtype) |
| attn_bias = self._attn_bias(attention_mask, T, h.device, h.dtype) |
|
|
| |
| do_inject = visual is not None and inject_iters > 0 |
|
|
| all_hidden = [h] if output_hidden_states else None |
| loop_hidden = [] |
|
|
| for block in self.prelude: |
| h = block(h, cos, sin, attn_bias) |
| if output_hidden_states: |
| all_hidden.append(h) |
| for loop_index in range(self.config.k): |
| |
| if do_inject and loop_index < inject_iters: |
| h = self.inject(h, visual, loop_index) |
| for block in self.core: |
| h = block(h, cos, sin, attn_bias) |
| if output_hidden_states: |
| all_hidden.append(h) |
| if self.config.inject_input: |
| h = h + residual_input |
| loop_hidden.append(h) |
| for block in self.coda: |
| h = block(h, cos, sin, attn_bias) |
| if output_hidden_states: |
| all_hidden.append(h) |
|
|
| h = self.final_norm(h) |
| return LoopedModelOutput( |
| last_hidden_state=h, |
| hidden_states=tuple(all_hidden) if output_hidden_states else None, |
| loop_hidden_states=tuple(loop_hidden), |
| ) |
|
|
|
|
| class LoopedForCausalLM(LoopedPreTrainedModel): |
| """言語モデリングヘッド付き(``AutoModelForCausalLM`` 互換)。""" |
|
|
| _tied_weights_keys = ["lm_head.weight"] |
|
|
| def __init__(self, config: BabyloopConfig): |
| super().__init__(config) |
| self.model = LoopedModel(config) |
| self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False) |
| 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): |
| self.lm_head = new |
|
|
| def forward( |
| self, |
| input_ids=None, |
| attention_mask=None, |
| inputs_embeds=None, |
| vision_features=None, |
| labels=None, |
| output_hidden_states=False, |
| **kwargs, |
| ) -> LoopedCausalLMOutput: |
| out = self.model( |
| input_ids=input_ids, |
| attention_mask=attention_mask, |
| inputs_embeds=inputs_embeds, |
| vision_features=vision_features, |
| output_hidden_states=output_hidden_states, |
| ) |
| logits = self.lm_head(out.last_hidden_state) |
|
|
| loss = None |
| if labels is not None: |
| |
| |
| n_prefix = logits.size(1) - labels.size(1) |
| if n_prefix > 0: |
| pad = labels.new_full((labels.size(0), n_prefix), -100) |
| labels = torch.cat([pad, labels], dim=1) |
| shift_logits = logits[:, :-1, :].contiguous() |
| shift_labels = labels[:, 1:].contiguous() |
| loss = F.cross_entropy( |
| shift_logits.view(-1, shift_logits.size(-1)), |
| shift_labels.view(-1), |
| ignore_index=-100, |
| ) |
|
|
| return LoopedCausalLMOutput( |
| loss=loss, |
| logits=logits, |
| hidden_states=out.hidden_states, |
| loop_hidden_states=out.loop_hidden_states, |
| ) |
|
|
|
|
| |
| |
| BabyloopConfig.register_for_auto_class() |
| LoopedModel.register_for_auto_class("AutoModel") |
| LoopedForCausalLM.register_for_auto_class("AutoModelForCausalLM") |
|
|