code/model_cheap_qa_minimal.py

"""
Minimal single-variant GPT — cheap_qa champion ONLY.

Distilled from the 19-variant code/model_canon_HEAD.py in this repo. Every other gate branch has been
removed; the ONLY conditioning path is the cheap_qa gate:

    cheap_qa  ==  qv_variant "dynamic_qa_conditioned"
    gate = sigmoid( Linear(2*head_dim -> head_dim) ( cat(q, y) ) )
    y = y * gate                      # applied at G1: post-SDPA, pre-W_O (c_proj)

Champion result (NanoGPT 124M, OpenWebText, 10k iters, seeds 42/1337/123):
    baseline 4.4708 -> cheap_qa 4.3897  (-0.0811 CE, -1.81%), 98,304 params (+0.079%).

>>> THE G1 SEAM IS MARKED BELOW BY A "G1 SEAM" BANNER COMMENT.
    That block is the entire port surface. Resize the gate for a different attention
    by changing head_dim (hs); the gate is always Linear(2*hs -> hs) on cat(q, y).
"""

import math
import inspect
from dataclasses import dataclass
import torch
import torch.nn as nn
from torch.nn import functional as F


class LayerNorm(nn.Module):
    def __init__(self, ndim, bias):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(ndim))
        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None

    def forward(self, input):
        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)


class CausalSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        assert config.n_embd % config.n_head == 0
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)   # <- W_O

        hs = config.n_embd // config.n_head   # head_dim (64 for the 124M config)

        # === cheap_qa gate parameters ===
        # cat(q, y) has width 2*hs (=128 when hs=64); gate maps it back to hs (=64).
        # This is the ONLY learned gate in this minimal model.
        self.qa_gate_proj = nn.Linear(hs * 2, hs, bias=False)   # Linear(128 -> 64)

        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
        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
        if not self.flash:
            self.register_buffer(
                "bias",
                torch.tril(torch.ones(config.block_size, config.block_size)).view(
                    1, 1, config.block_size, config.block_size
                ),
            )

    def forward(self, x):
        B, T, C = x.size()
        hs = C // self.n_head

        # Project to q, k, v and split into heads. q and v are (B, n_head, T, hs).
        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
        k = k.view(B, T, self.n_head, hs).transpose(1, 2)
        q = q.view(B, T, self.n_head, hs).transpose(1, 2)
        v = v.view(B, T, self.n_head, hs).transpose(1, 2)

        # Scaled dot-product attention -> y is the attention output (the "A" in Q+A).
        if self.flash:
            y = torch.nn.functional.scaled_dot_product_attention(
                q, k, v, is_causal=True, dropout_p=self.dropout if self.training else 0
            )
        else:
            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
            att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float('-inf'))
            att = F.softmax(att, dim=-1)
            att = self.attn_dropout(att)
            y = att @ v

        # =========================== G1 SEAM ============================
        # cheap_qa gate. Inputs available here: q and y, both (B, n_head, T, hs).
        # Location: strictly AFTER SDPA, strictly BEFORE the W_O projection (c_proj).
        # To port to a different attention: keep this exact 4-line body; only hs changes.
        gate_in = torch.cat([q, y], dim=-1)      # (B, n_head, T, 2*hs)  == cat(query, attn_output)
        gate_logit = self.qa_gate_proj(gate_in)  # Linear(2*hs -> hs)
        gate = torch.sigmoid(gate_logit)         # (B, n_head, T, hs) in (0, 1)
        y = y * gate                             # elementwise gate on the attention output
        # ========================= END G1 SEAM ==========================

        # Merge heads back and apply the output projection W_O.
        y = y.transpose(1, 2).contiguous().view(B, T, C)
        y = self.resid_dropout(self.c_proj(y))   # c_proj == W_O
        return y


class MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
        self.gelu = nn.GELU()
        self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x):
        x = self.c_fc(x)
        x = self.gelu(x)
        x = self.c_proj(x)
        x = self.dropout(x)
        return x


class Block(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
        self.mlp = MLP(config)

    def forward(self, x):
        x = x + self.attn(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x


@dataclass
class GPTConfig:
    block_size: int = 1024
    vocab_size: int = 50304   # NOTE: 50304 (GPT-2 50257 rounded up for efficiency)
    n_layer: int = 12
    n_head: int = 12
    n_embd: int = 768         # -> head_dim = 768 / 12 = 64, so the gate is Linear(128 -> 64)
    dropout: float = 0.0
    bias: bool = True
    # qv_variant is fixed to cheap_qa in this minimal model; no field needed.


class GPT(nn.Module):
    def __init__(self, config):
        super().__init__()
        assert config.vocab_size is not None
        assert config.block_size is not None
        self.config = config
        self.transformer = nn.ModuleDict(dict(
            wte=nn.Embedding(config.vocab_size, config.n_embd),
            wpe=nn.Embedding(config.block_size, config.n_embd),
            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),
        ))
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        self.transformer.wte.weight = self.lm_head.weight   # weight tying
        self.apply(self._init_weights)
        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))
        print("number of parameters: %.2fM" % (self.get_num_params() / 1e6,))

    def get_num_params(self, non_embedding=True):
        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)

    def forward(self, idx, targets=None):
        device = idx.device
        b, t = idx.size()
        assert t <= self.config.block_size
        pos = torch.arange(0, t, dtype=torch.long, device=device)
        tok_emb = self.transformer.wte(idx)
        pos_emb = self.transformer.wpe(pos)
        x = self.transformer.drop(tok_emb + pos_emb)
        for block in self.transformer.h:
            x = block(x)
        x = self.transformer.ln_f(x)
        if targets is not None:
            logits = self.lm_head(x)
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
        else:
            logits = self.lm_head(x[:, [-1], :])
            loss = None
        return logits, loss

    def crop_block_size(self, block_size):
        assert block_size <= self.config.block_size
        self.config.block_size = block_size
        self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:block_size])
        for block in self.transformer.h:
            if hasattr(block.attn, 'bias'):
                block.attn.bias = block.attn.bias[:, :, :block_size, :block_size]

    def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
        param_dict = {pn: p for pn, p in self.named_parameters() if p.requires_grad}
        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
        optim_groups = [
            {'params': decay_params, 'weight_decay': weight_decay},
            {'params': nodecay_params, 'weight_decay': 0.0},
        ]
        num_decay_params = sum(p.numel() for p in decay_params)
        num_nodecay_params = sum(p.numel() for p in nodecay_params)
        print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
        print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device_type == 'cuda'
        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, fused=use_fused)
        print(f"using fused AdamW: {use_fused}")
        return optimizer

    def estimate_mfu(self, fwdbwd_per_iter, dt):
        N = self.get_num_params()
        cfg = self.config
        L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd // cfg.n_head, cfg.block_size
        flops_per_token = 6 * N + 12 * L * H * Q * T
        flops_per_fwdbwd = flops_per_token * T
        flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter
        flops_achieved = flops_per_iter * (1.0 / dt)
        flops_promised = 312e12
        return flops_achieved / flops_promised

    @torch.no_grad()
    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
        for _ in range(max_new_tokens):
            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
            logits, _ = self(idx_cond)
            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