Create src/architecture/transformer.py

src/architecture/transformer.py

@@ -0,0 +1,167 @@
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+import math
+
+# --- Sovereign Components ---
+
+class RMSNorm(nn.Module):
+    """Faster and more stable normalization for Sovereign AI."""
+    def __init__(self, dim, eps=1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+class SwiGLU(nn.Module):
+    """Advanced activation for deep ecological reasoning."""
+    def __init__(self, dim):
+        super().__init__()
+        self.w1 = nn.Linear(dim, dim * 4, bias=False)
+        self.w2 = nn.Linear(dim, dim * 4, bias=False)
+        self.w3 = nn.Linear(dim * 4, dim, bias=False)
+
+    def forward(self, x):
+        return self.w3(F.silu(self.w1(x)) * self.w2(x))
+
+# --- The Core Block ---
+
+class AravalliBlock(nn.Module):
+    """
+    The fundamental unit of ARAVALLI-1 logic.
+    Each block processes the survival-context of the previous tokens.
+    """
+    def __init__(self, config):
+        super().__init__()
+        self.n_head = config['model_params']['n_head']
+        self.n_embd = config['model_params']['n_embd']
+        
+        # Norms
+        self.attention_norm = RMSNorm(self.n_embd)
+        self.ffn_norm = RMSNorm(self.n_embd)
+        
+        # Self-Attention (Simplified for MVP structure)
+        self.wq = nn.Linear(self.n_embd, self.n_embd, bias=False)
+        self.wk = nn.Linear(self.n_embd, self.n_embd, bias=False)
+        self.wv = nn.Linear(self.n_embd, self.n_embd, bias=False)
+        self.wo = nn.Linear(self.n_embd, self.n_embd, bias=False)
+        
+        # Feed Forward Network
+        self.feed_forward = SwiGLU(self.n_embd)
+
+    def forward(self, x):
+        # 1. Attention with Residual Connection
+        h = x + self.wo(self._self_attention(self.attention_norm(x)))
+        # 2. Feed Forward with Residual Connection
+        out = h + self.feed_forward(self.ffn_norm(h))
+        return out
+
+    def _self_attention(self, x):
+        # Optimized Multi-Head Attention Logic
+        B, T, C = x.size()
+        q = self.wq(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
+        k = self.wk(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
+        v = self.wv(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
+
+        # Scaled Dot-Product Attention
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+        # Apply causal mask (The model cannot see the future)
+        mask = torch.tril(torch.ones(T, T)).to(x.device)
+        att = att.masked_fill(mask == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        
+        y = att @ v
+        y = y.transpose(1, 2).contiguous().view(B, T, C)
+        return y
+class AravalliModel(nn.Module):
+    """
+    The full Sovereign AI Model: ARAVALLI-1.
+    A decoder-only transformer built from scratch for GOEC ecosystem mandates.
+    """
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        params = config['model_params']
+        
+        # 1. Token & Positional Embeddings
+        # We use a standard Embedding layer for tokens
+        self.token_embedding = nn.Embedding(params['vocab_size'], params['n_embd'])
+        
+        # 2. Transformer Blocks (The 'Brain' Layers)
+        self.blocks = nn.ModuleList([
+            AravalliBlock(config) for _ in range(params['n_layer'])
+        ])
+        
+        # 3. Final Normalization
+        self.final_norm = RMSNorm(params['n_embd'])
+        
+        # 4. Language Modeling Head
+        # Projects the 2048-dim embedding back to the 50,257-dim vocab
+        self.lm_head = nn.Linear(params['n_embd'], params['vocab_size'], bias=False)
+        
+        # Weight Tying (Optional but recommended for efficiency)
+        # This shares weights between embedding and lm_head
+        self.token_embedding.weight = self.lm_head.weight
+        
+        # Initialize all weights
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+
+    def forward(self, idx, targets=None):
+        B, T = idx.size()
+        
+        # Token Embeddings
+        x = self.token_embedding(idx) # Shape (B, T, n_embd)
+        
+        # Pass through the stack of AravalliBlocks
+        for block in self.blocks:
+            x = block(x)
+        
+        # Final Norm
+        x = self.final_norm(x)
+        
+        # Compute Logits
+        logits = self.lm_head(x) # Shape (B, T, vocab_size)
+
+        loss = None
+        if targets is not None:
+            # Flatten for CrossEntropyLoss
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+
+        return logits, loss
+
+    @torch.no_grad()
+    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
+        """Simple greedy/sampled generation for the Secretariat Node."""
+        for _ in range(max_new_tokens):
+            # Crop index if it exceeds context window
+            idx_cond = idx if idx.size(1) <= self.config['model_params']['n_positions'] else idx[:, -self.config['model_params']['n_positions']:]
+            
+            # Get logits
+            logits, _ = self(idx_cond)
+            # Focus only on the last time step
+            logits = logits[:, -1, :] / temperature
+            
+            if top_k is not None:
+                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                logits[logits < v[:, [-1]]] = -float('Inf')
+            
+            probs = F.softmax(logits, dim=-1)
+            idx_next = torch.multinomial(probs, num_samples=1)
+            
+            idx = torch.cat((idx, idx_next), dim=1)
+            
+        return idx