Add model variation for RPC attention

[gkielian]

Attempt to replicate the RPC Style of attention, review the following:

https://arxiv.org/abs/2406.13762

--- a/model.py
+++ b/model.py
@@ -7,6 +7,7 @@
 import torch.nn as nn
 from torch.nn import functional as F
 
+from torch.linalg import norm
 from dataclasses import dataclass
 
 @dataclass
@@ -128,7 +129,66 @@
 
         return logits, loss
 
+class PrincipalAttentionPursuit(nn.Module):
+    """
+    Implements the Principal Attention Pursuit (PAP) algorithm for RPC-Attention.
+    """
+
+    def __init__(self, n_iter, lambda_):
+        super().__init__()
+        self.n_iter = n_iter
+        self.lambda_ = lambda_
+
+    def shrinkage_operator(self, x, tau):
+        return torch.sign(x) * torch.relu(torch.abs(x) - tau)
+
+    def forward(self, q, k):
+        # Initialize S and Y
+        B, T, C = k.size()
+        S = torch.zeros_like(k)
+        Y = torch.zeros_like(k)
+        
+        # Calculate mu
+        mu = (B * T) / (4 * norm(k, p=1)) 
+
+        # Ensure mu and lambda_ are on the same device as k
+        mu = mu.to(k.device)
+        lambda_ = self.lambda_.to(k.device)
+
+        # Iterative refinement
+        L = k.clone() # Initialize L to k
+        for _ in range(self.n_iter):
+            S = self.shrinkage_operator(L - (mu**-1) * Y, lambda_ / mu)
+            # Softmax will be applied in the forward pass of the calling CausalSelfAttention
+            # Here we use L in place of K in the attention mechanism
+            # Thus passing (q, L) instead of (q,k) to the forward of CausalSelfAttention
+            L = L # In this implementation L remains unchanged through the iterations
+            Y = Y + mu * (k - L - S)
+
+        return L # Return refined key matrix L
+
 class CausalSelfAttention(nn.Module):
 
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
+        # regularization
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.dropout = config.dropout
+        # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0
+        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
+        if not self.flash:
+            print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")
+            # causal mask to ensure that attention is only applied to the left in the input sequence
+            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
+                                        .view(1, 1, config.block_size, config.block_size))
+        
+        # RPC-Attention parameters
+        self.use_rpc_attention = config.use_rpc_attention
+        if self.use_rpc_attention:
+            self.pap = PrincipalAttentionPursuit(n_iter=config.n_iter, lambda_=torch.tensor(config.lambda_))
+
     def __init__(self, config):
         super().__init__()
         assert config.n_embd % config.n_head == 0
@@ -153,14 +213,20 @@
         # q, k, v = (x.transpose(1, 2) for x in self.c_attn(x).split(self.n_embd, dim=2))
         q, k, v  = self.c_attn(x).split(self.n_embd, dim=2)
         
-        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
-        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
-        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        if self.use_rpc_attention:
+            # Apply PAP to refine the key matrix
+            k = self.pap(q, k)
+            
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)  # (B, nh, T, hs)
 
         # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
         if self.flash:
             # efficient attention using Flash Attention CUDA kernels
-            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout if self.training else 0, is_causal=True)
+            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None,
+                                                                 dropout_p=self.dropout if self.training else 0,
+                                                                 is_causal=True)
         else:
             # manual implementation of attention
             att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
@@ -202,6 +268,7 @@
         super().__init__()
         self.block_size = config.block_size
         self.transformer = nn.ModuleDict(dict(
+            # Removed wte and wpe initializations to only initialize them in the GPT class
             drop = nn.Dropout(config.dropout),
             h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
             ln_f = LayerNorm(config.n_embd, bias=config.bias),
@@ -216,6 +283,28 @@
         self.apply(self._init_weights)
         # apply special scaled init to the residual projections, per GPT-2 paper
         for pn, p in self.named_parameters():
+            if pn.endswith('c_proj.weight'):
+                torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))
+
+    def get_num_params(self, non_embedding=True):
+        """
+        Return the number of parameters in the model.
+        For non-embedding count (default), the position embeddings get subtracted.
+        The token embeddings would too, except due to the parameter sharing these
+        params are actually used as weights in the final layer, so we include them.
+        """
+        n_params = sum(p.numel() for p in self.parameters())
+        if non_embedding:
+            n_params -= self.transformer.wpe.weight.numel()
+        return n_params
+
+    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)
             if pn.endswith('c_proj.weight'):
                 torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))
 
@@ -262,6 +351,8 @@
         self.use_flash = use_flash
         self.lambda_ = lambda_
         self.n_iter = n_iter
+        self.use_rpc_attention = use_rpc_attention
+        
 
     def save(self, filepath):
         """
@@ -285,9 +376,9 @@
             block_size=self.block_size,
             n_layer=self.n_layer,
             n_head=self.n_head,
-            n_embd=self.n_embd,
             dropout=self.dropout,
             bias=self.bias,
+            n_embd=self.n_embd,
             use_flash=self.use_flash,
             lambda_ = self.lambda_,
             n_iter = self.n_iter,
@@ -297,6 +388,10 @@
         """
         # Override the default implementation to initialize the model using the config
         config = self.get_config()
+        
+        # Initialize token and position embeddings here
+        config.transformer.wte = torch.nn.Embedding(config.vocab_size, config.n_embd)
+        config.transformer.wpe = torch.nn.Embedding(config.block_size, config.n_embd)
+        
         model = GPT(config)
         return model