# ============================================================================= # COPYRIGHT © 2025-2026 Konstantin Vladimirovich Grabko. ALL RIGHTS RESERVED. # CMS Manhattan JiRack Technology — PATENT PENDING # # This code is proprietary. # Personal and non-commercial research use is allowed. # Any commercial use, derivative works for profit, or distribution # requires a paid license and 5% royalty. # # Unauthorized commercial use is strictly prohibited. # Contact: grabko@cmsmanhattan.com # ============================================================================= # # Model updated for new last tokenizer version # # Replace toknizer in current model: Just use the class and call resize function in your train script # New model: Use our conversion script to rapidly initialize new models by copying existing embeddings and LM_head weights. This enables fast model bootstrapping, or alternatively, provides the foundation to train a new model entirely from scratch. # # ============================================================================= import torch import torch.nn as nn import torch.nn.functional as F # --- JIRACK 1B ARCHITECTURE CONSTANTS --- VOCAB_SIZE = 128259 HIDDEN_SIZE = 2048 NUM_LAYERS = 16 NUM_HEADS = 32 NUM_KV_HEADS = 8 INTERMEDIATE_SIZE = 8192 MAX_SEQ_LEN = 4096 RMS_EPS = 1e-6 # --- QUANTIZATION PARAMETERS --- STABILITY_EPS = 1e-9 INT8_SCALE_TARGET = 127.0 class TernaryConfig: def __init__(self): self.vocab_size = VOCAB_SIZE self.hidden_size = HIDDEN_SIZE self.num_hidden_layers = NUM_LAYERS self.num_attention_heads = NUM_HEADS self.num_key_value_heads = NUM_KV_HEADS self.intermediate_size = INTERMEDIATE_SIZE self.max_position_embeddings = MAX_SEQ_LEN self.rms_norm_eps = RMS_EPS class BitLinear(nn.Linear): def __init__(self, in_features, out_features, bias=False): super().__init__(in_features, out_features, bias) def forward(self, x): # Weight Quantization w = self.weight gamma = w.abs().mean().clamp(min=STABILITY_EPS) w_quant = torch.clamp(torch.round(w / gamma), -1, 1) w_final = w + (w_quant * gamma - w).detach() # Activation Quantization (Absmax) x_norm = x - x.mean(dim=-1, keepdim=True) x_max = x_norm.abs().max(dim=-1, keepdim=True).values.clamp(min=STABILITY_EPS) scale = INT8_SCALE_TARGET / x_max x_quant = (x_norm * scale).round().clamp(-128, 127) / scale x_final = x + (x_quant - x).detach() return F.linear(x_final, w_final, self.bias) class RMSNorm(nn.Module): def __init__(self, dim, eps=RMS_EPS): super().__init__() self.eps = eps self.weight = nn.Parameter(torch.ones(dim)) def forward(self, x): return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight # --- ROPE WITHOUT COMPLEX NUMBERS --- def precompute_freqs_cis(dim, seq_len, theta=500000.0): freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)) t = torch.arange(seq_len).float() freqs = torch.outer(t, freqs) return torch.cos(freqs), torch.sin(freqs) def apply_rotary_emb(xq, xk, freqs_cos, freqs_sin): def rotate_half(x): # Split 64 into two 32s x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) T = xq.shape[2] # FIX: Repeat frequencies (32 -> 64) to match head_dim f_cos = freqs_cos[:T].to(xq.device).view(1, 1, T, -1).repeat(1, 1, 1, 2) f_sin = freqs_sin[:T].to(xq.device).view(1, 1, T, -1).repeat(1, 1, 1, 2) xq_out = (xq * f_cos) + (rotate_half(xq) * f_sin) xk_out = (xk * f_cos) + (rotate_half(xk) * f_sin) return xq_out, xk_out def repeat_kv(x, n_rep): if n_rep == 1: return x bs, n_kv_heads, seqlen, head_dim = x.shape return x[:, :, None, :, :].expand(bs, n_kv_heads, n_rep, seqlen, head_dim).reshape(bs, n_kv_heads * n_rep, seqlen, head_dim) class TransformerBlock(nn.Module): def __init__(self, config): super().__init__() self.n_heads = config.num_attention_heads self.n_kv_heads = config.num_key_value_heads self.n_rep = self.n_heads // self.n_kv_heads self.head_dim = config.hidden_size // self.n_heads self.q_proj = BitLinear(config.hidden_size, config.hidden_size) self.k_proj = BitLinear(config.hidden_size, self.n_kv_heads * self.head_dim) self.v_proj = BitLinear(config.hidden_size, self.n_kv_heads * self.head_dim) self.out_proj = BitLinear(config.hidden_size, config.hidden_size) self.ffn_w1 = BitLinear(config.hidden_size, config.intermediate_size) self.ffn_w3 = BitLinear(config.hidden_size, config.intermediate_size) self.ffn_w2 = BitLinear(config.intermediate_size, config.hidden_size) self.norm1 = RMSNorm(config.hidden_size) self.norm2 = RMSNorm(config.hidden_size) def forward(self, x, freqs_cos, freqs_sin): h = self.norm1(x) B, T, D = x.shape q = self.q_proj(h).view(B, T, self.n_heads, self.head_dim).transpose(1, 2) k = self.k_proj(h).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2) v = self.v_proj(h).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2) q, k = apply_rotary_emb(q, k, freqs_cos, freqs_sin) k, v = repeat_kv(k, self.n_rep), repeat_kv(v, self.n_rep) attn_out = F.scaled_dot_product_attention(q, k, v, is_causal=True) x = x + self.out_proj(attn_out.transpose(1, 2).reshape(B, T, D)) m = self.norm2(x) x = x + self.ffn_w2(F.silu(self.ffn_w1(m)) * self.ffn_w3(m)) return x class TernaryTransformer1B(nn.Module): def __init__(self, config): super().__init__() self.token_emb = nn.Embedding(config.vocab_size, config.hidden_size) self.blocks = nn.ModuleList([TransformerBlock(config) for _ in range(config.num_hidden_layers)]) self.ln_f = RMSNorm(config.hidden_size) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # RoPE frequencies (64 head_dim -> 32 pairs) cos, sin = precompute_freqs_cis(config.hidden_size // config.num_attention_heads, MAX_SEQ_LEN) self.register_buffer("freqs_cos", cos) self.register_buffer("freqs_sin", sin) def forward(self, input_ids): x = self.token_emb(input_ids) for block in self.blocks: x = block(x, self.freqs_cos, self.freqs_sin) return self.lm_head(self.ln_f(x)), None