import math import torch import torch.nn as nn import torch.nn.functional as F from transformers import PreTrainedModel from transformers.generation import GenerationMixin from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast from configuration_caca import CacaConfig # --- NORM & MLP --- class CacaRMSNorm(nn.Module): def __init__(self, dim, eps=1e-6): super().__init__() self.eps = eps self.weight = nn.Parameter(torch.zeros(dim)) def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) def forward(self, x): out = self._norm(x.float()) out = out * (1.0 + self.weight.float()) return out.type_as(x) class CacaMLP(nn.Module): def __init__(self, config: CacaConfig, intermediate_size=None): super().__init__() inter = intermediate_size or config.intermediate_size self.gate_proj = nn.Linear(config.hidden_size, inter, bias=False) self.up_proj = nn.Linear(config.hidden_size, inter, bias=False) self.down_proj = nn.Linear(inter, config.hidden_size, bias=False) self.act_fn = nn.SiLU() if config.hidden_activation == "silu" else nn.GELU(approximate="tanh") self.dropout = nn.Dropout(config.hidden_dropout) def forward(self, x): return self.dropout(self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))) # --- ROTARY EMBEDDING — default / linear / dynamic / YaRN --- class CacaRotaryEmbedding(nn.Module): def __init__(self, config: CacaConfig, dim: int, device=None): super().__init__() self.config = config self.dim = dim rope_params = getattr(config, "rope_parameters", None) or {} self.rope_type = rope_params.get("rope_type", "default") self.base = rope_params.get("rope_theta", getattr(config, "rope_theta", 10000.0)) self.factor = rope_params.get("factor", 1.0) self.original_max_pos = rope_params.get( "original_max_position_embeddings", config.max_position_embeddings ) self.beta_fast = rope_params.get("beta_fast", 32) self.beta_slow = rope_params.get("beta_slow", 1) self.mscale = rope_params.get("mscale", 1.0) if self.rope_type == "yarn": inv_freq, self.attention_scaling = self._yarn_inv_freq(device) else: inv_freq = 1.0 / (self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim)) self.attention_scaling = 1.0 self.register_buffer("inv_freq", inv_freq, persistent=False) self.max_seq_len_cached = config.max_position_embeddings def _yarn_find_correction_dim(self, num_rot): return (self.dim * math.log(self.original_max_pos / (num_rot * 2 * math.pi))) / (2 * math.log(self.base)) def _yarn_inv_freq(self, device): dim = self.dim pos_freqs = self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim) inv_freq_extrapolation = 1.0 / pos_freqs inv_freq_interpolation = 1.0 / (self.factor * pos_freqs) low = max(math.floor(self._yarn_find_correction_dim(self.beta_fast)), 0) high = min(math.ceil(self._yarn_find_correction_dim(self.beta_slow)), dim - 1) ramp = torch.linspace(0, 1, dim // 2, device=device) ramp = torch.clamp((ramp * dim - low) / max(high - low, 1e-3), 0, 1) inv_freq_mask = 1.0 - ramp inv_freq = inv_freq_interpolation * (1 - inv_freq_mask) + inv_freq_extrapolation * inv_freq_mask mscale = 0.1 * math.log(self.factor) + 1.0 if self.factor > 1 else 1.0 return inv_freq, mscale @torch.no_grad() def forward(self, x, position_ids): inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) pos_expanded = position_ids[:, None, :].float() freqs = (inv_freq_expanded @ pos_expanded).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() * self.attention_scaling sin = emb.sin() * self.attention_scaling return cos.to(x.dtype), sin.to(x.dtype) def rotate_half(x): x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :] 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 def repeat_kv(x, n_rep): if n_rep == 1: return x b, h, s, d = x.shape x = x[:, :, None, :, :].expand(b, h, n_rep, s, d) return x.reshape(b, h * n_rep, s, d) # --- CACHE — sederhana --- class SimpleCache: def __init__(self): self.entries = {} def update(self, layer_idx, *tensors): if layer_idx not in self.entries: self.entries[layer_idx] = list(tensors) else: self.entries[layer_idx] = [ torch.cat([old, new], dim=-2) for old, new in zip(self.entries[layer_idx], tensors) ] return self.entries[layer_idx] def get_seq_length(self, layer_idx=0): if layer_idx not in self.entries: return 0 return self.entries[layer_idx][0].shape[-2] # --- ATTENTION — GQA (use_mla=False) --- class CacaGQAAttention(nn.Module): def __init__(self, config: CacaConfig, layer_idx: int): super().__init__() self.layer_idx = layer_idx self.head_dim = config.head_dim self.num_heads = config.num_attention_heads self.num_kv_heads = config.num_key_value_heads self.num_kv_groups = self.num_heads // self.num_kv_heads self.scaling = config.query_pre_attn_scalar ** -0.5 self.attn_dropout = config.attention_dropout self.attn_softcap = config.attn_logit_softcapping self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias) self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.attention_bias) self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=config.attention_bias) self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=config.attention_bias) self.use_qk_norm = config.use_qk_norm if self.use_qk_norm: self.q_norm = CacaRMSNorm(self.head_dim, config.rms_norm_eps) self.k_norm = CacaRMSNorm(self.head_dim, config.rms_norm_eps) self.rotary_emb = CacaRotaryEmbedding(config, dim=self.head_dim) def forward(self, hidden_states, attention_mask, position_ids, cache=None, **kwargs): b, seq_len, _ = hidden_states.shape shape = (b, seq_len, -1, self.head_dim) q = self.q_proj(hidden_states).view(shape) k = self.k_proj(hidden_states).view(shape) v = self.v_proj(hidden_states).view(shape).transpose(1, 2) if self.use_qk_norm: q, k = self.q_norm(q), self.k_norm(k) q, k = q.transpose(1, 2), k.transpose(1, 2) cos, sin = self.rotary_emb(hidden_states, position_ids) q, k = apply_rotary_pos_emb(q, k, cos, sin) if cache is not None: k, v = cache.update(self.layer_idx, k, v) k = repeat_kv(k, self.num_kv_groups) v = repeat_kv(v, self.num_kv_groups) attn_weights = torch.matmul(q, k.transpose(2, 3)) * self.scaling if self.attn_softcap is not None: attn_weights = torch.tanh(attn_weights / self.attn_softcap) * self.attn_softcap if attention_mask is not None: attn_weights = attn_weights + attention_mask[:, :, :, : k.shape[-2]] attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype) attn_weights = F.dropout(attn_weights, p=self.attn_dropout, training=self.training) attn_output = torch.matmul(attn_weights, v) attn_output = attn_output.transpose(1, 2).contiguous().reshape(b, seq_len, -1) return self.o_proj(attn_output) # --- ATTENTION — MLA --- class CacaMLAAttention(nn.Module): def __init__(self, config: CacaConfig, layer_idx: int): super().__init__() self.layer_idx = layer_idx self.num_heads = config.num_attention_heads self.q_lora_rank = config.q_lora_rank self.kv_lora_rank = config.kv_lora_rank self.qk_nope_head_dim = config.qk_nope_head_dim self.qk_rope_head_dim = config.qk_rope_head_dim self.v_head_dim = config.v_head_dim self.q_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim self.scaling = self.q_head_dim ** -0.5 self.attn_dropout = config.attention_dropout self.attn_softcap = config.attn_logit_softcapping self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None if self.q_lora_rank > 0: self.q_a_proj = nn.Linear(config.hidden_size, self.q_lora_rank, bias=False) self.q_a_norm = CacaRMSNorm(self.q_lora_rank, config.rms_norm_eps) self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * self.q_head_dim, bias=False) else: self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.q_head_dim, bias=False) self.kv_a_proj_with_mqa = nn.Linear( config.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False ) self.kv_a_norm = CacaRMSNorm(self.kv_lora_rank, config.rms_norm_eps) self.kv_b_proj = nn.Linear( self.kv_lora_rank, self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), bias=False ) self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, config.hidden_size, bias=False) self.use_qk_norm = config.use_qk_norm if self.use_qk_norm: self.q_nope_norm = CacaRMSNorm(self.qk_nope_head_dim, config.rms_norm_eps) self.k_nope_norm = CacaRMSNorm(self.qk_nope_head_dim, config.rms_norm_eps) self.rotary_emb = CacaRotaryEmbedding(config, dim=self.qk_rope_head_dim) def forward(self, hidden_states, attention_mask, position_ids, cache=None, **kwargs): b, seq_len, _ = hidden_states.shape if self.q_lora_rank > 0: q = self.q_b_proj(self.q_a_norm(self.q_a_proj(hidden_states))) else: q = self.q_proj(hidden_states) q = q.view(b, seq_len, self.num_heads, self.q_head_dim).transpose(1, 2) q_nope, q_rope = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1) kv_a = self.kv_a_proj_with_mqa(hidden_states) kv_a, k_rope = kv_a.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1) kv_a = self.kv_a_norm(kv_a) k_rope = k_rope.view(b, seq_len, 1, self.qk_rope_head_dim).transpose(1, 2) if cache is not None: kv_a_seq, k_rope_seq = cache.update(self.layer_idx, kv_a.unsqueeze(1), k_rope) kv_a = kv_a_seq.squeeze(1) else: kv_a_seq, k_rope_seq = kv_a.unsqueeze(1), k_rope kv = self.kv_b_proj(kv_a_seq.squeeze(1) if cache is None else cache.entries[self.layer_idx][0].squeeze(1)) kv_len = kv.shape[1] kv = kv.view(b, kv_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim).transpose(1, 2) k_nope, value = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1) if self.use_qk_norm: q_nope = self.q_nope_norm(q_nope) k_nope = self.k_nope_norm(k_nope) cos, sin = self.rotary_emb(hidden_states, position_ids) q_rope, k_rope_seq = apply_rotary_pos_emb(q_rope, k_rope_seq, cos, sin) k_rope_expanded = k_rope_seq.expand(-1, self.num_heads, -1, -1) q_full = torch.cat([q_nope, q_rope], dim=-1) k_full = torch.cat([k_nope, k_rope_expanded], dim=-1) attn_weights = torch.matmul(q_full, k_full.transpose(2, 3)) * self.scaling if self.attn_softcap is not None: attn_weights = torch.tanh(attn_weights / self.attn_softcap) * self.attn_softcap if attention_mask is not None: attn_weights = attn_weights + attention_mask[:, :, :, :kv_len] attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q_full.dtype) attn_weights = F.dropout(attn_weights, p=self.attn_dropout, training=self.training) attn_output = torch.matmul(attn_weights, value) attn_output = attn_output.transpose(1, 2).contiguous().reshape(b, seq_len, -1) return self.o_proj(attn_output) # --- DECODER LAYER --- class CacaDecoderLayer(nn.Module): def __init__(self, config: CacaConfig, layer_idx: int): super().__init__() self.self_attn = CacaMLAAttention(config, layer_idx) if config.use_mla else CacaGQAAttention(config, layer_idx) self.mlp = CacaMLP(config) self.input_layernorm = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.post_attention_layernorm = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.pre_feedforward_layernorm = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.post_feedforward_layernorm = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.residual_dropout = nn.Dropout(config.hidden_dropout) def forward(self, hidden_states, attention_mask, position_ids, cache=None, **kwargs): residual = hidden_states hidden_states = self.input_layernorm(hidden_states) hidden_states = self.self_attn(hidden_states, attention_mask, position_ids, cache) hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = residual + self.residual_dropout(hidden_states) residual = hidden_states hidden_states = self.pre_feedforward_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = self.post_feedforward_layernorm(hidden_states) hidden_states = residual + self.residual_dropout(hidden_states) return hidden_states # --- MASK UTILS --- def build_attention_mask(attention_mask, seq_len, past_len, sliding_window, dtype, device): min_val = torch.finfo(dtype).min query_pos = torch.arange(past_len, past_len + seq_len, device=device)[:, None] key_pos = torch.arange(past_len + seq_len, device=device)[None, :] causal = key_pos > query_pos mask = torch.zeros((seq_len, past_len + seq_len), dtype=dtype, device=device) mask.masked_fill_(causal, min_val) if sliding_window is not None: too_far = key_pos <= (query_pos - sliding_window) mask.masked_fill_(too_far, min_val) mask = mask[None, None, :, :] if attention_mask is not None: pad = (1.0 - attention_mask[:, None, None, :].to(dtype)) * min_val mask = mask + pad return mask # --- PRETRAINED BASE --- class CacaPreTrainedModel(PreTrainedModel): config_class = CacaConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["CacaDecoderLayer"] 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_() # --- MODEL BODY --- class CacaModel(CacaPreTrainedModel): def __init__(self, config: CacaConfig): super().__init__(config) self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id) self.embedding_dropout = nn.Dropout(config.embedding_dropout) self.layers = nn.ModuleList([CacaDecoderLayer(config, i) for i in range(config.num_hidden_layers)]) self.norm = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.hidden_scale = config.hidden_size ** 0.5 self.post_init() def forward(self, input_ids, attention_mask=None, position_ids=None, cache=None, use_cache=None, **kwargs): use_cache = use_cache if use_cache is not None else self.config.use_cache b, seq_len = input_ids.shape if use_cache and cache is None: cache = SimpleCache() past_len = cache.get_seq_length(0) if cache is not None else 0 if position_ids is None: position_ids = torch.arange(past_len, past_len + seq_len, device=input_ids.device)[None, :].expand(b, -1) hidden_states = self.embed_tokens(input_ids) * self.hidden_scale hidden_states = self.embedding_dropout(hidden_states) full_mask = build_attention_mask(attention_mask, seq_len, past_len, None, hidden_states.dtype, hidden_states.device) sliding_mask = build_attention_mask( attention_mask, seq_len, past_len, self.config.sliding_window, hidden_states.dtype, hidden_states.device ) for layer in self.layers: mask = sliding_mask if layer.self_attn.sliding_window is not None else full_mask if self.gradient_checkpointing and self.training: hidden_states = torch.utils.checkpoint.checkpoint( layer, hidden_states, mask, position_ids, cache, use_reentrant=False ) else: hidden_states = layer(hidden_states, mask, position_ids, cache) hidden_states = self.norm(hidden_states) return BaseModelOutputWithPast(last_hidden_state=hidden_states, past_key_values=cache) # --- MULTI-TOKEN PREDICTION MODULE --- class CacaMTPModule(nn.Module): def __init__(self, config: CacaConfig): super().__init__() self.norm_prev = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.norm_emb = CacaRMSNorm(config.hidden_size, config.rms_norm_eps) self.combine_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False) self.decoder_layer = CacaDecoderLayer(config, layer_idx=0) def forward(self, prev_hidden, target_embeds, attention_mask, position_ids): combined = self.combine_proj(torch.cat([self.norm_prev(prev_hidden), self.norm_emb(target_embeds)], dim=-1)) return self.decoder_layer(combined, attention_mask, position_ids, cache=None) # --- CAUSAL LM HEAD --- class CacaForCausalLM(CacaPreTrainedModel, GenerationMixin): _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"} def __init__(self, config: CacaConfig): super().__init__(config) self.model = CacaModel(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.mtp_modules = nn.ModuleList( [CacaMTPModule(config) for _ in range(config.num_mtp_tokens)] ) if config.num_mtp_tokens > 0 else None 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 forward( self, input_ids, attention_mask=None, position_ids=None, labels=None, cache=None, use_cache=None, logits_to_keep=0, **kwargs, ): outputs = self.model(input_ids, attention_mask, position_ids, cache, use_cache) hidden_states = outputs.last_hidden_state slice_idx = slice(-logits_to_keep, None) if logits_to_keep else slice(None) logits = self.lm_head(hidden_states[:, slice_idx, :]) if self.config.final_logit_softcapping is not None: cap = self.config.final_logit_softcapping logits = torch.tanh(logits / cap) * cap loss = None if labels is not None: shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() main_loss = F.cross_entropy( shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1), ignore_index=-100 ) loss = main_loss if self.mtp_modules is not None: mtp_loss_total = 0.0 prev_hidden = hidden_states b, seq_len = input_ids.shape pos_ids = position_ids if position_ids is not None else torch.arange(seq_len, device=input_ids.device)[None, :].expand(b, -1) for k, mtp in enumerate(self.mtp_modules, start=1): if seq_len - k <= 1: break target_ids = input_ids[:, k:] target_embeds = self.model.embed_tokens(target_ids) * self.model.hidden_scale aligned_prev = prev_hidden[:, : target_ids.shape[1], :] aligned_mask = None mtp_hidden = mtp(aligned_prev, target_embeds, aligned_mask, pos_ids[:, : target_ids.shape[1]]) mtp_logits = self.lm_head(mtp_hidden) mtp_labels = labels[:, k + 1 :] mtp_logits_trimmed = mtp_logits[:, : mtp_labels.shape[1], :] if mtp_labels.shape[1] > 0: mtp_loss = F.cross_entropy( mtp_logits_trimmed.reshape(-1, mtp_logits_trimmed.size(-1)), mtp_labels.reshape(-1), ignore_index=-100, ) mtp_loss_total = mtp_loss_total + mtp_loss prev_hidden = mtp_hidden if isinstance(mtp_loss_total, torch.Tensor): loss = main_loss + self.config.mtp_loss_weight * mtp_loss_total return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values) def prepare_inputs_for_generation(self, input_ids, cache=None, attention_mask=None, **kwargs): if cache is not None and cache.get_seq_length(0) > 0: input_ids = input_ids[:, -1:] return {"input_ids": input_ids, "attention_mask": attention_mask, "cache": cache, "use_cache": True, "logits_to_keep": 1} # --- AUTO-REGISTER --- CacaConfig.register_for_auto_class() CacaModel.register_for_auto_class("AutoModel") CacaForCausalLM.register_for_auto_class("AutoModelForCausalLM")