JiRackTernaryPyTorch_405b.py

# ==============================================================================
# COPYRIGHT (C) 2025 KONSTANTIN VLADIMIROVICH GRABKO. ALL RIGHTS RESERVED.
# PATENT PENDING | CMS MANHATTAN JIRACK TECHNOLOGY
#
# This software is licensed under the Commercial License Agreement V.1.2.
# Any use, modification, or distribution of this code requires compliance with 
# the terms found in the LICENSE.md file in the root directory.
#
# NO PATENTING RIGHTS: Users are strictly prohibited from filing patent claims
# based on the BRE or SWA architectures disclosed herein.
# Contact: grabko@cmsmanhattan.com | +1 (516) 777-0945
# ==============================================================================
# COPYRIGHT (C) 2025 KONSTANTIN VLADIMIROVICH GRABKO. ALL RIGHTS RESERVED.
# PATENT PENDING | CMS MANHATTAN JIRACK TECHNOLOGY | MOE-VERSION 405B
# ==============================================================================

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional, Tuple
from transformers import PreTrainedModel, PretrainedConfig
from transformers.modeling_outputs import CausalLMOutputWithPast

class JiRackMoEConfig(PretrainedConfig):
    model_type = "jirack_moe_transformer"
    def __init__(
        self,
        vocab_size=128256,
        hidden_size=12288,
        num_hidden_layers=112,    # MoE allows reduced depth while increasing experts
        num_attention_heads=96,
        num_experts=16,           # Total of 16 experts per layer
        num_experts_per_tok=2,    # Activate only the top 2 experts per token
        intermediate_size=12288,  # Size of each expert
        max_position_embeddings=32768,
        rms_norm_eps=1e-5,
        author="Author: Konstantin Vladimirovich Grabko (CMS Manhattan) 2025",
        **kwargs
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.num_experts = num_experts
        self.num_experts_per_tok = num_experts_per_tok
        self.intermediate_size = intermediate_size
        self.max_position_embeddings = max_position_embeddings
        self.rms_norm_eps = rms_norm_eps
        self.author = author

# --- Ternary Linear Logic (Bit-Response Engine) ---

class JiRackBitLinear(nn.Linear):
    def __init__(self, in_features, out_features, bias=False):
        super().__init__(in_features, out_features, bias)
        nn.init.normal_(self.weight, std=0.02)
    def forward(self, x):
        w = self.weight
        gamma = w.abs().mean() + 1e-9
        w_quant = torch.clamp(torch.round(w / gamma), -1, 1)
        w_final = w + (w_quant * gamma - w).detach()
        x_norm = x - x.mean(dim=-1, keepdim=True)
        x_quant = x_norm + (torch.clamp(x_norm, -1.2, 1.2) - x_norm).detach()
        return F.linear(x_quant, w_final, self.bias)

# --- Expert Block ---

class JiRackExpert(nn.Module):
    def __init__(self, config: JiRackMoEConfig):
        super().__init__()
        self.w1 = JiRackBitLinear(config.hidden_size, config.intermediate_size)
        self.w2 = JiRackBitLinear(config.intermediate_size, config.hidden_size)
        self.w3 = JiRackBitLinear(config.hidden_size, config.intermediate_size)
    def forward(self, x):
        return self.w2(F.silu(self.w1(x)) * self.w3(x))

# --- Router (Gate) for Expert Selection ---

class JiRackMoEGate(nn.Module):
    def __init__(self, config: JiRackMoEConfig):
        super().__init__()
        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.top_k = config.num_experts_per_tok
    def forward(self, x):
        logits = self.gate(x)
        weights, indices = torch.topk(logits, self.top_k, dim=-1)
        weights = F.softmax(weights.float(), dim=-1).type_as(x)
        return weights, indices

# --- Signature and Phase CMS Layer ---

class SignatureLayer(nn.Module):
    def __init__(self, dim, author_name):
        super().__init__()
        self.gate = nn.Parameter(torch.ones(dim))
        seed = sum(ord(c) for c in author_name)
        torch.manual_seed(seed)
        self.signage_cms = nn.Parameter(torch.randn(dim, dim) * 0.001)
    def forward(self, x):
        sig = torch.tanh(F.linear(x, self.signage_cms))
        return x * torch.sigmoid(self.gate) + sig

# --- Main MoE Transformer Block ---

class JiRackMoEBlock(nn.Module):
    def __init__(self, config: JiRackMoEConfig):
        super().__init__()
        self.ln1 = nn.LayerNorm(config.hidden_size)
        self.attn = nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
        self.ln2 = nn.LayerNorm(config.hidden_size)
        
        # MoE Integration
        self.gate = JiRackMoEGate(config)
        self.experts = nn.ModuleList([JiRackExpert(config) for _ in range(config.num_experts)])
        
        self.signature = SignatureLayer(config.hidden_size, config.author)

    def forward(self, x):
        # Attention
        residual = x
        x = self.ln1(x)
        x, _ = self.attn(x, x, x, need_weights=False)
        x = residual + x
        
        # MoE Routing
        residual = x
        x = self.ln2(x)
        weights, indices = self.gate(x)
        
        final_output = torch.zeros_like(x)
        # A loop is used here for simplicity; scatter/gather is used in production
        for i, expert in enumerate(self.experts):
            mask = (indices == i).any(dim=-1)
            if mask.any():
                expert_output = expert(x[mask])
                # Applying expert contribution weights
                w = weights[indices == i].unsqueeze(-1)
                final_output[mask] += expert_output * w
        
        x = residual + self.signature(final_output)
        return x

# --- Main 405B MoE Model ---

class JiRackTernaryMoE405B(PreTrainedModel):
    config_class = JiRackMoEConfig
    def __init__(self, config: JiRackMoEConfig):
        super().__init__(config)
        self.token_emb = nn.Embedding(config.vocab_size, config.hidden_size)
        self.blocks = nn.ModuleList([JiRackMoEBlock(config) for _ in range(config.num_hidden_layers)])
        self.ln_f = nn.LayerNorm(config.hidden_size)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        
        self.post_init()

    def forward(self, input_ids, labels=None):
        x = self.token_emb(input_ids)
        for block in self.blocks:
            if self.training:
                x = torch.utils.checkpoint.checkpoint(block, x, use_reentrant=False)
            else:
                x = block(x)
        
        logits = self.lm_head(self.ln_f(x))
        loss = None
        if labels is not None:
            loss = F.cross_entropy(logits.view(-1, self.config.vocab_size), labels.view(-1))
        
        return CausalLMOutputWithPast(loss=loss, logits=logits)