import math import torch import torch.nn as nn import torch.nn.functional as F from transformers import PreTrainedModel from transformers.modeling_outputs import CausalLMOutputWithPast from .configuration_sovythos import SovythosConfig class RMSNorm(nn.Module): def __init__(self, dim, eps=1e-06): super().__init__() self.eps = eps self.weight = nn.Parameter(torch.ones(dim)) def forward(self, x): dtype = x.dtype x = x.float() rms = x.pow(2).mean(-1, keepdim=True) x = x * torch.rsqrt(rms + self.eps) return self.weight * x.to(dtype) def precompute_rope(head_dim, max_seq_len, theta, device, dtype=torch.float32): inv_freq = 1.0 / theta ** (torch.arange(0, head_dim, 2, device=device).float() / head_dim) t = torch.arange(max_seq_len, device=device).float() freqs = torch.outer(t, inv_freq) emb = torch.cat((freqs, freqs), dim=-1) return emb.cos().to(dtype), emb.sin().to(dtype) def rotate_half(x): x1, x2 = x.chunk(2, dim=-1) return torch.cat((-x2, x1), dim=-1) def apply_rope(x, cos, sin, offset=0): T = x.shape[-2] cos = cos[offset:offset + T][None, None, :, :] sin = sin[offset:offset + T][None, None, :, :] return x * cos + rotate_half(x) * sin def repeat_kv(x, n_rep): if n_rep == 1: return x B, Hkv, T, D = x.shape return x[:, :, None, :, :].expand(B, Hkv, n_rep, T, D).reshape(B, Hkv * n_rep, T, D) class Attention(nn.Module): def __init__(self, cfg): super().__init__() self.n_heads = cfg.n_heads self.n_kv_heads = cfg.n_kv_heads self.n_rep = self.n_heads // self.n_kv_heads self.head_dim = cfg.dim // cfg.n_heads self.dropout = cfg.dropout self.q_proj = nn.Linear(cfg.dim, cfg.n_heads * self.head_dim, bias=False) self.k_proj = nn.Linear(cfg.dim, cfg.n_kv_heads * self.head_dim, bias=False) self.v_proj = nn.Linear(cfg.dim, cfg.n_kv_heads * self.head_dim, bias=False) self.o_proj = nn.Linear(cfg.n_heads * self.head_dim, cfg.dim, bias=False) self.q_norm = RMSNorm(self.head_dim, cfg.norm_eps) self.k_norm = RMSNorm(self.head_dim, cfg.norm_eps) def forward(self, x, cos, sin): B, T, C = x.shape q = self.q_proj(x).view(B, T, self.n_heads, self.head_dim).transpose(1, 2) k = self.k_proj(x).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2) v = self.v_proj(x).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2) q, k = self.q_norm(q), self.k_norm(k) q = apply_rope(q, cos, sin) k = apply_rope(k, cos, sin) k = repeat_kv(k, self.n_rep) v = repeat_kv(v, self.n_rep) out = F.scaled_dot_product_attention(q, k, v, is_causal=True, dropout_p=self.dropout if self.training else 0.0) out = out.transpose(1, 2).contiguous().view(B, T, -1) return self.o_proj(out) class FeedForward(nn.Module): def __init__(self, cfg): super().__init__() hidden = int(2 * (4 * cfg.dim) / 3) if cfg.ffn_dim_multiplier is not None: hidden = int(cfg.ffn_dim_multiplier * hidden) hidden = cfg.ffn_multiple_of * ((hidden + cfg.ffn_multiple_of - 1) // cfg.ffn_multiple_of) self.gate_proj = nn.Linear(cfg.dim, hidden, bias=False) self.up_proj = nn.Linear(cfg.dim, hidden, bias=False) self.down_proj = nn.Linear(hidden, cfg.dim, bias=False) def forward(self, x): return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x)) class DecoderBlock(nn.Module): def __init__(self, cfg): super().__init__() self.input_layernorm = RMSNorm(cfg.dim, cfg.norm_eps) self.self_attn = Attention(cfg) self.post_attention_layernorm = RMSNorm(cfg.dim, cfg.norm_eps) self.mlp = FeedForward(cfg) def forward(self, x, cos, sin): x = x + self.self_attn(self.input_layernorm(x), cos, sin) x = x + self.mlp(self.post_attention_layernorm(x)) return x class SovythosPreTrainedModel(PreTrainedModel): config_class = SovythosConfig base_model_prefix = "model" supports_gradient_checkpointing = True def _init_weights(self, module): if isinstance(module, nn.Linear): nn.init.normal_(module.weight, mean=0.0, std=0.02) 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=0.02) class SovythosForCausalLM(SovythosPreTrainedModel): def __init__(self, config): super().__init__(config) self.embed_tokens = nn.Embedding(config.vocab_size, config.dim) self.layers = nn.ModuleList([DecoderBlock(config) for _ in range(config.n_layers)]) self.norm = RMSNorm(config.dim, config.norm_eps) self.lm_head = nn.Linear(config.dim, config.vocab_size, bias=False) if config.tie_embeddings: self.lm_head.weight = self.embed_tokens.weight self.head_dim = config.dim // config.n_heads self._rope_cache = {} self.post_init() def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, value): self.embed_tokens = value def get_output_embeddings(self): return self.lm_head def _rope_for(self, device, dtype): # ملاحظة: بنحسب rope live بدل ما نعتمد على register_buffer، لأن # transformers بيحمل الموديل عادةً عبر meta-device (low_cpu_mem_usage) # وأي buffer معمول persistent=False ومش موجود في checkpoint بيفضل # ذاكرة غير مهيأة (garbage) بدل القيم الحقيقية بعد النقل من meta لـ real. key = str(device) if key not in self._rope_cache: cos, sin = precompute_rope(self.head_dim, self.config.max_seq_len, self.config.rope_theta, device=device) self._rope_cache[key] = (cos, sin) return self._rope_cache[key] def forward(self, input_ids=None, attention_mask=None, labels=None, **kwargs): cos, sin = self._rope_for(input_ids.device, torch.float32) h = self.embed_tokens(input_ids) for layer in self.layers: h = layer(h, cos, sin) h = self.norm(h) logits = self.lm_head(h) 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, shift_logits.size(-1)), shift_labels.view(-1), ignore_index=-100, ) return CausalLMOutputWithPast(loss=loss, logits=logits) def prepare_inputs_for_generation(self, input_ids, **kwargs): return {"input_ids": input_ids}