source/JiRack_H12_L18_V50257_D768_MSL8192_FF768x4.py

# Copyright (c) 2025 CMS Manhattan
# All rights reserved.
# Author: Konstantin Vladimirovich Grabko
# Email: grabko@cmsmanhattan.com
# Phone: +1(516)777-0945
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, version 3 of the License.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <https://www.gnu.org/licenses/>.
#
# Additional terms:
# Any commercial use or distribution of this software or derivative works
# requires explicit written permission from the copyright holder.

import os
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple, List
import math

# ========================================
# Model Configuration (GPT-2 Medium Style)
# ========================================
VOCAB_SIZE = 50257
MODEL_DIM = 768
NUM_HEADS = 12
NUM_LAYERS = 18       # Increased depth
MAX_SEQ_LEN = 8192
FFN_HIDDEN_DIM = 4 * MODEL_DIM
HEAD_DIM = MODEL_DIM // NUM_HEADS

# ИСПРАВЛЕНИЕ: ROCm/HIP-совместимая проверка устройства
if torch.cuda.is_available():
    device = torch.device("cuda")
elif hasattr(torch, 'hip') and torch.hip.is_available():
    device = torch.device("cuda") 
else:
    device = torch.device("cpu")

# -------------------------------
# Learned Positional Embedding
# -------------------------------
class LearnedPositionalEmbedding(nn.Module):
    def __init__(self, max_seq_len: int, embed_dim: int):
        super().__init__()
        self.pos_emb = nn.Parameter(torch.zeros(max_seq_len, embed_dim))

    def forward(self, x: torch.Tensor, pos_offset: int = 0) -> torch.Tensor:
        seq_len = x.size(1)
        if pos_offset + seq_len > self.pos_emb.size(0):
            # Проверка на выход за пределы MAX_SEQ_LEN
            raise ValueError("Sequence length exceeds MAX_SEQ_LEN defined in position embedding.")
        
        pos = self.pos_emb[pos_offset : pos_offset + seq_len]
        return x + pos.unsqueeze(0)


# -------------------------------
# MultiHeadAttention (MHA)
# -------------------------------
class MultiHeadAttention(nn.Module):
    def __init__(self):
        super().__init__()
        self.q_proj = nn.Linear(MODEL_DIM, MODEL_DIM, bias=False)
        self.k_proj = nn.Linear(MODEL_DIM, MODEL_DIM, bias=False)
        self.v_proj = nn.Linear(MODEL_DIM, MODEL_DIM, bias=False)
        self.out_proj = nn.Linear(MODEL_DIM, MODEL_DIM, bias=False)
        self.scale = HEAD_DIM ** -0.5

    def forward(self, x: torch.Tensor, past_kv: Optional[Tuple[torch.Tensor, torch.Tensor]] = None):
        B, T, D = x.shape
        
        q = self.q_proj(x).view(B, T, NUM_HEADS, HEAD_DIM).transpose(1, 2)
        k = self.k_proj(x).view(B, T, NUM_HEADS, HEAD_DIM).transpose(1, 2)
        v = self.v_proj(x).view(B, T, NUM_HEADS, HEAD_DIM).transpose(1, 2)

        # 1. KV-кеш и определение смещения
        pos_offset = 0
        seqlen_k_new = k.size(2)
        if past_kv is not None:
            past_k, past_v = past_kv
            k = torch.cat([past_k, k], dim=2)
            v = torch.cat([past_v, v], dim=2)
            pos_offset = past_k.size(2)
        
        seqlen_k = k.size(2) # Общая длина K
        new_kv = (k, v)

        # 2. Расчет внимания
        attn = torch.matmul(q, k.transpose(-2, -1)) * self.scale

        # 3. КРИТИЧЕСКОЕ ИСПРАВЛЕНИЕ МАСКИРОВАНИЯ (Causal Mask)
        if T == seqlen_k_new and seqlen_k > 0:
            # Создаем маску T x seqlen_k
            mask = torch.full((T, seqlen_k), 
                              float("-inf"), 
                              device=x.device, 
                              dtype=attn.dtype)
            
            # Разрешаем всем новым токенам видеть все старые токены (past_kv)
            mask[:, :pos_offset] = 0.0

            # Применяем треугольную маску для текущих T токенов
            current_causal_mask = torch.tril(torch.ones(T, T, device=x.device, dtype=torch.bool))
            mask[:, pos_offset : pos_offset + T].masked_fill_(~current_causal_mask, float('-inf'))
            
            # Применяем маску к весам внимания
            attn = attn + mask[None, None, :, :]

        # 4. Выход
        attn = F.softmax(attn, dim=-1)
        out = torch.matmul(attn, v)
        out = out.transpose(1, 2).contiguous().view(B, T, D)
        out = self.out_proj(out)

        return out, new_kv


# -------------------------------
# FeedForward (GELU, GPT-style)
# -------------------------------
class FeedForward(nn.Module):
    def __init__(self):
        super().__init__()
        self.c_fc = nn.Linear(MODEL_DIM, FFN_HIDDEN_DIM, bias=False)
        self.c_proj = nn.Linear(FFN_HIDDEN_DIM, MODEL_DIM, bias=False)

    def forward(self, x):
        return self.c_proj(F.gelu(self.c_fc(x), approximate='tanh'))


# -------------------------------
# Transformer Block (Post-Norm, GPT-style)
# -------------------------------
class TransformerBlock(nn.Module):
    def __init__(self):
        super().__init__()
        self.attn = MultiHeadAttention()
        self.ffn = FeedForward()
        # ИСПРАВЛЕНИЕ: Добавлен стандартный eps
        self.norm1 = nn.LayerNorm(MODEL_DIM, eps=1e-5) 
        self.norm2 = nn.LayerNorm(MODEL_DIM, eps=1e-5)

    def forward(self, x, past_kv: Optional[Tuple[torch.Tensor, torch.Tensor]] = None):
        # Post-Normalization (GPT Style): Input -> Attn(Norm(Input)) + Input -> FFN(Norm(Result)) + Result
        attn_out, new_kv = self.attn(self.norm1(x), past_kv)
        x = x + attn_out
        x = x + self.ffn(self.norm2(x))
        return x, new_kv


# -------------------------------
# Главная модель GPTPyTorch (18 слоев)
# -------------------------------
class GPTPyTorch(nn.Module):
    def __init__(self):
        super().__init__()
        self.token_emb = nn.Embedding(VOCAB_SIZE, MODEL_DIM)
        self.pos_emb = LearnedPositionalEmbedding(MAX_SEQ_LEN, MODEL_DIM)
        self.blocks = nn.ModuleList([TransformerBlock() for _ in range(NUM_LAYERS)])
        self.ln_f = nn.LayerNorm(MODEL_DIM, eps=1e-5)
        self.lm_head = nn.Linear(MODEL_DIM, VOCAB_SIZE, bias=False)
        
        signature = "Konstantin V Gbabko .  original author © 2025"
        bytes_tensor = torch.tensor([ord(c) for c in signature], dtype=torch.uint8)
        self.register_buffer("konstantin_gbabko_proof_of_authorship", bytes_tensor)
        self.register_buffer("konstantin_gbabko_birth_date", torch.tensor([20251126], dtype=torch.int64))

        self.lm_head.weight = self.token_emb.weight
        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            # ИСПРАВЛЕНИЕ: Инициализация, масштабированная по глубине сети (TFixup/ReZero style)
            std = 0.02 / math.sqrt(2 * NUM_LAYERS) if isinstance(module, nn.Linear) and module.out_features == MODEL_DIM else 0.02
            torch.nn.init.normal_(module.weight, mean=0.0, std=std)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
        elif isinstance(module, nn.LayerNorm):
            nn.init.zeros_(module.bias)
            nn.init.ones_(module.weight)

    def forward(self, input_ids, past_kv: Optional[List[Tuple[torch.Tensor, torch.Tensor]]] = None):
        B, T = input_ids.shape
        x = self.token_emb(input_ids)

        pos_offset = 0
        if past_kv is not None and past_kv[0] is not None:
            pos_offset = past_kv[0][0].size(2)
            
        x = self.pos_emb(x, pos_offset=pos_offset)

        # ИСПРАВЛЕНИЕ: Инициализируем новый кеш
        new_kv_cache = [] if past_kv is not None or T > 1 else None
        current_past = past_kv

        for i, block in enumerate(self.blocks):
            layer_past = current_past[i] if (current_past and i < len(current_past)) else None
            x, layer_kv = block(x, layer_past)
            
            if new_kv_cache is not None:
                new_kv_cache.append(layer_kv)

        x = self.ln_f(x)
        logits = self.lm_head(x)
        return logits, new_kv_cache

    @torch.no_grad()
    def generate(
        self,
        input_ids: torch.Tensor,
        max_new_tokens: int = 100,
        temperature: float = 0.8,
        top_p: float = 0.95,
        repetition_penalty: float = 1.0,
        do_sample: bool = True,
        eos_token_id: int = 50256
    ) -> torch.Tensor:
        # ИСПРАВЛЕНИЕ: Инициализируем KV cache как список None кортежей
        kv_cache = [None] * NUM_LAYERS
        current_ids = input_ids.clone()

        for step in range(max_new_tokens):
            if step == 0:
                input_for_model = current_ids
            else:
                input_for_model = current_ids[:, -1].unsqueeze(-1)

            logits, kv_cache = self(input_for_model, kv_cache)
            next_token_logits = logits[:, -1, :]

            if temperature > 0:
                next_token_logits = next_token_logits / temperature

            # Repetition Penalty (логика сохранена)
            if repetition_penalty != 1.0:
                for i in range(current_ids.shape[0]):
                    unique_tokens = torch.unique(current_ids[i]).tolist()
                    for token_id in unique_tokens:
                        score = next_token_logits[i, token_id]
                        if score < 0:
                            next_token_logits[i, token_id] = score * repetition_penalty
                        else:
                            next_token_logits[i, token_id] = score / repetition_penalty

            # Top-P сэмплирование (логика сохранена)
            if do_sample and top_p < 1.0:
                sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
                cumulative_probs = torch.softmax(sorted_logits, dim=-1).cumsum(dim=-1)
                sorted_indices_to_remove = cumulative_probs > top_p
                sorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()
                sorted_indices_to_remove[:, 0] = False
                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
                next_token_logits = next_token_logits.masked_fill(indices_to_remove, float('-inf'))

            # Сэмплирование
            if do_sample and temperature > 0:
                probs = torch.softmax(next_token_logits, dim=-1)
                if torch.isnan(probs).any() or torch.isinf(probs).any():
                    next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
                else:
                    next_token = torch.multinomial(probs, num_samples=1)
            else:
                next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)

            if next_token.item() == eos_token_id:
                break

            current_ids = torch.cat([current_ids, next_token], dim=1)

        return current_ids


if __name__ == "__main__":
    os.makedirs("models", exist_ok=True)

    model = GPTPyTorch().to(device)
    model.eval()

    total_params = sum(p.numel() for p in model.parameters())
    print(f"Device: {device}")
    print(f"Total parameters: {total_params / 1e6:.2f}M") # ~167.33M

    # 1. Проверка первого прохода (T=50)
    input_ids_T50 = torch.randint(0, VOCAB_SIZE, (1, 50), device=device)
    with torch.no_grad():
        logits_50, kv_cache_50 = model(input_ids_T50)
        expected_k_shape = (1, NUM_HEADS, 50, HEAD_DIM)
        assert kv_cache_50[0][0].shape == expected_k_shape
        print(f"Initial logits shape: {logits_50.shape}")
        print(f"Initial KV-cache seqlen: {kv_cache_50[0][0].size(2)}")

    # 2. Проверка инкрементального прохода (T=1)
    input_ids_T1 = torch.randint(0, VOCAB_SIZE, (1, 1), device=device)
    with torch.no_grad():
        logits_51, kv_cache_51 = model(input_ids_T1, past_kv=kv_cache_50)

        # Проверка длины кеша: 50 + 1 = 51
        assert kv_cache_51[0][0].size(2) == 51
        print(f"Incremental logits shape: {logits_51.shape}")
        print(f"Incremental KV-cache seqlen: {kv_cache_51[0][0].size(2)}")

    # 3. Проверка функции generate
    generated = model.generate(input_ids_T50, max_new_tokens=5, temperature=0.8, top_p=0.9)
    print(f"Generated sequence length (50 + 5): {generated.shape[1]}")

    save_path = "models/JiRack_GPT_L18_PostNorm_fixed.pt"
    torch.save(model.state_dict(), save_path)
    print(f"Model successfully saved to {save_path}")