Create JiRackTernaryPyTorch_1b_HF_SPEC.py

JiRackTernaryPyTorch_1b_HF_SPEC.py

@@ -0,0 +1,196 @@
+# ==============================================================================
+# COPYRIGHT (C) 2026 KONSTANTIN VLADIMIROVICH GRABKO. ALL RIGHTS RESERVED.
+# PATENT PENDING | CMS MANHATTAN JIRACK TECHNOLOGY
+# ==============================================================================
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PreTrainedModel, PretrainedConfig # added for HF
+
+# Contact for JiRack Signature Layer
+VOCAB_SIZE = 128256
+HIDDEN_SIZE = 2048
+NUM_LAYERS = 16
+NUM_HEADS = 32
+NUM_KV_HEADS = 8  
+INTERMEDIATE_SIZE = 8192
+MAX_SEQ_LEN = 4096
+RMS_EPS = 1e-6
+
+TERNARY_MIN = -1
+TERNARY_MAX = 1
+INT8_MIN = -128
+INT8_MAX = 127
+INT8_SCALE_TARGET = 127.0
+STABILITY_EPS = 1e-9
+
+class JiRackTernaryConfig(PretrainedConfig):
+    """Configuration to registration Hugging Face ecosystem """
+    model_type = "jirack_ternary"
+    def __init__(
+        self,
+        vocab_size=VOCAB_SIZE,
+        hidden_size=HIDDEN_SIZE,
+        num_hidden_layers=NUM_LAYERS,
+        num_attention_heads=NUM_HEADS,
+        num_key_value_heads=NUM_KV_HEADS,
+        intermediate_size=INTERMEDIATE_SIZE,
+        max_position_embeddings=MAX_SEQ_LEN,
+        rms_norm_eps=RMS_EPS,
+        **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_key_value_heads = num_key_value_heads
+        self.intermediate_size = intermediate_size
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+
+# Old class exist TernaryConfig to be compateble old ML scripts
+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):
+        w = self.weight
+        gamma = w.abs().mean().clamp(min=STABILITY_EPS)
+        w_quant = torch.clamp(torch.round(w / gamma), TERNARY_MIN, TERNARY_MAX)
+        w_final = w + (w_quant * gamma - w).detach()
+
+        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(INT8_MIN, INT8_MAX) / 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
+
+def precompute_freqs_cis(dim, seq_len):
+    base = 10000.0
+    freqs = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+    t = torch.arange(seq_len).float()
+    freqs = torch.outer(t, freqs)
+    return torch.polar(torch.ones_like(freqs), freqs)
+
+def apply_rotary_emb(xq, xk, freqs_cis):
+    xq_f = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_f = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+    freqs_cis = freqs_cis.to(xq_f.device)[None, None, :xq_f.shape[2], :]
+    xq_out = torch.view_as_real(xq_f * freqs_cis).flatten(3)
+    xk_out = torch.view_as_real(xk_f * freqs_cis).flatten(3)
+    return xq_out.type_as(xq), xk_out.type_as(xk)
+
+def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
+    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_cis):
+        B, T, D = x.shape
+        h = self.norm1(x)
+
+        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_cis)
+
+        k = repeat_kv(k, self.n_rep)
+        v = repeat_kv(v, self.n_rep)
+
+        attn_out = F.scaled_dot_product_attention(q, k, v, is_causal=True)
+
+        attn_out = attn_out.transpose(1, 2).reshape(B, T, D)
+        x = x + self.out_proj(attn_out)
+
+        m = self.norm2(x)
+        x = x + self.ffn_w2(F.silu(self.ffn_w1(m)) * self.ffn_w3(m))
+        return x
+
+class TernaryTransformer1B(PreTrainedModel):
+    config_class = JiRackTernaryConfig
+
+    def __init__(self, config):
+        # if old ojbect came TernaryConfig, then convert it to JiRackTernaryConfig
+        if not isinstance(config, PretrainedConfig):
+            config = JiRackTernaryConfig()
+            
+        super().__init__(config)
+        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)
+
+        # Reg buffer for RoPE
+        self.register_buffer("freqs_cis", precompute_freqs_cis(config.hidden_size // config.num_attention_heads, MAX_SEQ_LEN))
+        
+        # init for HF
+        self.post_init()
+
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = self.token_emb(input_ids)
+        for block in self.blocks:
+            x = block(x, self.freqs_cis)
+        
+        logits = self.lm_head(self.ln_f(x))
+        
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss = F.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+        return (loss, logits) if loss is not None else (logits, None)
+
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        # minimal model.generate()
+        return {"input_ids": input_ids}