MURZIK-15B init weights

config.json

@@ -0,0 +1,31 @@
+{
+  "architectures": [
+    "MurzikForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "dtype": "bfloat16",
+  "eos_token_id": 2,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 20480,
+  "max_position_embeddings": 8192,
+  "model_type": "murzik",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.6.0",
+  "use_cache": true,
+  "use_qk_norm": true,
+  "vocab_size": 128256,
+  "auto_map": {
+    "AutoConfig": "murzik.configuration_murzik.MurzikConfig",
+    "AutoModelForCausalLM": "murzik.modeling_murzik.MurzikForCausalLM"
+  }
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:682eabba623085c2c33a5b59251d2033c476dcd432261bfac2a363e8752e71a9
+size 25473255848

murzik/__init__.py

@@ -0,0 +1,21 @@
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+
+from .configuration_murzik import MurzikConfig
+from .configuration_murzik_moe import MurzikMoeConfig
+from .modeling_murzik import MurzikForCausalLM
+from .modeling_murzik_moe import MurzikMoeForCausalLM
+from .tokenization_murzik import MurzikTokenizer
+
+AutoConfig.register("murzik", MurzikConfig)
+AutoConfig.register("murzik_moe", MurzikMoeConfig)
+AutoModelForCausalLM.register(MurzikConfig, MurzikForCausalLM)
+AutoModelForCausalLM.register(MurzikMoeConfig, MurzikMoeForCausalLM)
+AutoTokenizer.register(MurzikConfig, slow_tokenizer_class=MurzikTokenizer)
+
+__all__ = [
+    "MurzikConfig",
+    "MurzikMoeConfig",
+    "MurzikForCausalLM",
+    "MurzikMoeForCausalLM",
+    "MurzikTokenizer",
+]

murzik/configuration_murzik.py

@@ -0,0 +1,53 @@
+"""Murzik dense config (pilot 1B/15B)."""
+
+from transformers import PretrainedConfig
+
+
+class MurzikConfig(PretrainedConfig):
+    model_type = "murzik"
+
+    def __init__(
+        self,
+        vocab_size: int = 128256,
+        hidden_size: int = 1536,
+        intermediate_size: int = 6144,
+        num_hidden_layers: int = 24,
+        num_attention_heads: int = 16,
+        num_key_value_heads: int = 4,
+        head_dim: int | None = None,
+        hidden_act: str = "silu",
+        max_position_embeddings: int = 32768,
+        initializer_range: float = 0.02,
+        rms_norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = True,
+        rope_theta: float = 1_000_000.0,
+        attention_dropout: float = 0.0,
+        use_qk_norm: bool = True,
+        pad_token_id: int | None = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_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.head_dim = head_dim or hidden_size // num_attention_heads
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.use_qk_norm = use_qk_norm
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

murzik/configuration_murzik_moe.py

@@ -0,0 +1,69 @@
+"""MurzikMoE config (32B/64B/100B)."""
+
+from transformers import PretrainedConfig
+
+
+class MurzikMoeConfig(PretrainedConfig):
+    model_type = "murzik_moe"
+
+    def __init__(
+        self,
+        vocab_size: int = 128256,
+        hidden_size: int = 2560,
+        intermediate_size: int = 9728,
+        num_hidden_layers: int = 40,
+        num_attention_heads: int = 32,
+        num_key_value_heads: int = 8,
+        head_dim: int | None = None,
+        hidden_act: str = "silu",
+        max_position_embeddings: int = 32768,
+        initializer_range: float = 0.006,
+        rms_norm_eps: float = 1e-6,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = True,
+        rope_theta: float = 1_000_000.0,
+        attention_dropout: float = 0.0,
+        use_qk_norm: bool = True,
+        decoder_sparse_step: int = 1,
+        moe_intermediate_size: int = 2432,
+        num_experts: int = 96,
+        num_experts_per_tok: int = 6,
+        num_shared_experts: int = 2,
+        first_k_dense_replace: int = 2,
+        router_aux_loss_coef: float = 0.001,
+        expert_bias_update_speed: float = 0.001,
+        pad_token_id: int | None = None,
+        bos_token_id: int = 1,
+        eos_token_id: int = 2,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_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.head_dim = head_dim or hidden_size // num_attention_heads
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+        self.use_qk_norm = use_qk_norm
+        self.decoder_sparse_step = decoder_sparse_step
+        self.moe_intermediate_size = moe_intermediate_size
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_shared_experts = num_shared_experts
+        self.first_k_dense_replace = first_k_dense_replace
+        self.router_aux_loss_coef = router_aux_loss_coef
+        self.expert_bias_update_speed = expert_bias_update_speed
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

murzik/modeling_murzik.py

@@ -0,0 +1,280 @@
+"""Murzik dense decoder (pilot). GQA + RoPE + SwiGLU + RMSNorm."""
+
+from __future__ import annotations
+
+import math
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from transformers import PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.utils import logging
+
+from .configuration_murzik import MurzikConfig
+
+logger = logging.get_logger(__name__)
+
+
+class MurzikRMSNorm(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+def rotate_half(x: torch.Tensor) -> torch.Tensor:
+    x1, x2 = x.chunk(2, dim=-1)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin):
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class MurzikRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, max_position_embeddings: int, base: float, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.int64).float() / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.max_seq_len_cached = max_position_embeddings
+        t = torch.arange(max_position_embeddings, device=device, dtype=torch.int64).type_as(self.inv_freq)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
+        self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
+
+    def forward(self, x: torch.Tensor, seq_len: int):
+        return (
+            self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+            self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
+        )
+
+
+class MurzikMLP(nn.Module):
+    def __init__(self, config: MurzikConfig):
+        super().__init__()
+        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class MurzikAttention(nn.Module):
+    def __init__(self, config: MurzikConfig, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = config.num_key_value_heads
+        self.head_dim = config.head_dim
+        self.num_kv_groups = self.num_heads // self.num_kv_heads
+
+        self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=False)
+        self.q_norm = MurzikRMSNorm(self.head_dim, eps=config.rms_norm_eps) if config.use_qk_norm else None
+        self.k_norm = MurzikRMSNorm(self.head_dim, eps=config.rms_norm_eps) if config.use_qk_norm else None
+        self.dropout = nn.Dropout(config.attention_dropout)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        past_key_value: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+        use_cache: bool = False,
+    ):
+        bsz, q_len, _ = hidden_states.size()
+        q = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(hidden_states).view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+
+        if self.q_norm is not None:
+            q = self.q_norm(q)
+        if self.k_norm is not None:
+            k = self.k_norm(k)
+
+        cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        if past_key_value is not None:
+            k = torch.cat([past_key_value[0], k], dim=2)
+            v = torch.cat([past_key_value[1], v], dim=2)
+        past = (k, v) if use_cache else None
+
+        k = k.repeat_interleave(self.num_kv_groups, dim=1)
+        v = v.repeat_interleave(self.num_kv_groups, dim=1)
+
+        attn_weights = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(self.head_dim)
+        if attention_mask is not None:
+            attn_weights = attn_weights + attention_mask
+        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
+        attn_weights = self.dropout(attn_weights)
+        attn_output = torch.matmul(attn_weights, v)
+        attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, q_len, -1)
+        return self.o_proj(attn_output), past
+
+
+class MurzikDecoderLayer(nn.Module):
+    def __init__(self, config: MurzikConfig, layer_idx: int):
+        super().__init__()
+        self.self_attn = MurzikAttention(config, layer_idx)
+        self.mlp = MurzikMLP(config)
+        self.input_layernorm = MurzikRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = MurzikRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(self, hidden_states, attention_mask, position_embeddings, past_key_value=None, use_cache=False):
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states, present = self.self_attn(
+            hidden_states, attention_mask, position_embeddings, past_key_value, use_cache
+        )
+        hidden_states = residual + hidden_states
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = residual + self.mlp(hidden_states)
+        return hidden_states, present
+
+
+class MurzikPreTrainedModel(PreTrainedModel):
+    config_class = MurzikConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MurzikDecoderLayer"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+
+
+class MurzikModel(MurzikPreTrainedModel):
+    def __init__(self, config: MurzikConfig):
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id)
+        self.layers = nn.ModuleList([MurzikDecoderLayer(config, i) for i in range(config.num_hidden_layers)])
+        self.norm = MurzikRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = MurzikRotaryEmbedding(
+            config.head_dim, config.max_position_embeddings, config.rope_theta
+        )
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[list] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ):
+        bsz, seq_len = input_ids.shape
+        hidden_states = self.embed_tokens(input_ids)
+        cos, sin = self.rotary_emb(hidden_states, seq_len)
+        position_embeddings = (cos, sin)
+
+        if attention_mask is None:
+            attention_mask = torch.triu(
+                torch.full((seq_len, seq_len), float("-inf"), device=input_ids.device),
+                diagonal=1,
+            ).unsqueeze(0).unsqueeze(0)
+        else:
+            attention_mask = attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
+            attention_mask = (1.0 - attention_mask) * torch.finfo(hidden_states.dtype).min
+
+        presents = [] if use_cache else None
+        for idx, layer in enumerate(self.layers):
+            past = past_key_values[idx] if past_key_values is not None else None
+            if self.gradient_checkpointing and self.training:
+                hidden_states, present = self._checkpoint_layer(
+                    layer, hidden_states, attention_mask, position_embeddings, past, use_cache
+                )
+            else:
+                hidden_states, present = layer(
+                    hidden_states, attention_mask, position_embeddings, past, use_cache
+                )
+            if use_cache:
+                presents.append(present)
+
+        hidden_states = self.norm(hidden_states)
+        return hidden_states, presents
+
+    def _checkpoint_layer(self, layer, hidden_states, attention_mask, position_embeddings, past, use_cache):
+        def custom_forward(hs):
+            out, pr = layer(hs, attention_mask, position_embeddings, past, use_cache)
+            return out, pr
+
+        return torch.utils.checkpoint.checkpoint(custom_forward, hidden_states, use_reentrant=False)
+
+
+class MurzikForCausalLM(MurzikPreTrainedModel):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+
+    def __init__(self, config: MurzikConfig):
+        super().__init__(config)
+        self.model = MurzikModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[list] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> CausalLMOutputWithPast:
+        hidden_states, past_key_values = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            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),
+                ignore_index=-100,
+            )
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=past_key_values,
+        )

murzik/modeling_murzik_moe.py

@@ -0,0 +1,234 @@
+"""MurzikMoE — sparse MoE FFN on top of Murzik decoder blocks."""
+
+from __future__ import annotations
+
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from .configuration_murzik_moe import MurzikMoeConfig
+from .modeling_murzik import (
+    MurzikAttention,
+    MurzikMLP,
+    MurzikPreTrainedModel,
+    MurzikRMSNorm,
+    MurzikRotaryEmbedding,
+)
+
+
+class MurzikMoeMLP(nn.Module):
+    """Single expert SwiGLU block."""
+
+    def __init__(self, config: MurzikMoeConfig, intermediate_size: int):
+        super().__init__()
+        self.gate_proj = nn.Linear(config.hidden_size, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(config.hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, config.hidden_size, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class MurzikSparseMoeBlock(nn.Module):
+    def __init__(self, config: MurzikMoeConfig):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.top_k = config.num_experts_per_tok
+        self.hidden_size = config.hidden_size
+
+        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
+        self.experts = nn.ModuleList(
+            [MurzikMoeMLP(config, config.moe_intermediate_size) for _ in range(config.num_experts)]
+        )
+        self.shared_experts = nn.ModuleList(
+            [MurzikMoeMLP(config, config.moe_intermediate_size) for _ in range(config.num_shared_experts)]
+        )
+        self.register_buffer("expert_bias", torch.zeros(config.num_experts), persistent=True)
+        self.router_aux_loss_coef = config.router_aux_loss_coef
+
+    def forward(self, hidden_states: torch.Tensor):
+        batch_size, seq_len, hidden_dim = hidden_states.shape
+        flat = hidden_states.view(-1, hidden_dim)
+        router_logits = self.gate(flat)
+        routing_weights = F.softmax(router_logits + self.expert_bias, dim=-1, dtype=torch.float32)
+        routing_weights, selected = torch.topk(routing_weights, self.top_k, dim=-1)
+        routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights = routing_weights.to(flat.dtype)
+
+        out = torch.zeros_like(flat)
+        for expert_idx, expert in enumerate(self.experts):
+            mask = (selected == expert_idx).any(dim=-1)
+            if not mask.any():
+                continue
+            idx = mask.nonzero(as_tuple=True)[0]
+            expert_input = flat[idx]
+            expert_out = expert(expert_input)
+            weight = (selected[idx] == expert_idx).float() * routing_weights[idx]
+            weight = weight.sum(dim=-1, keepdim=True)
+            out[idx] += expert_out * weight
+
+        for shared in self.shared_experts:
+            out += shared(flat)
+
+        aux_loss = self._aux_loss(router_logits, selected)
+        return out.view(batch_size, seq_len, hidden_dim), aux_loss
+
+    def _aux_loss(self, router_logits: torch.Tensor, selected: torch.Tensor) -> torch.Tensor:
+        probs = F.softmax(router_logits, dim=-1, dtype=torch.float32)
+        one_hot = F.one_hot(selected, num_classes=self.num_experts).float().sum(dim=1)
+        load = one_hot.mean(dim=0)
+        balance = probs.mean(dim=0)
+        aux = self.num_experts * (load * balance).sum()
+        return aux * self.router_aux_loss_coef
+
+
+class MurzikMoeDecoderLayer(nn.Module):
+    def __init__(self, config: MurzikMoeConfig, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.self_attn = MurzikAttention(config, layer_idx)
+        self.input_layernorm = MurzikRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = MurzikRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        use_moe = layer_idx >= config.first_k_dense_replace
+        if use_moe:
+            self.mlp = MurzikSparseMoeBlock(config)
+            self.is_moe = True
+        else:
+            self.mlp = MurzikMLP(config)
+            self.is_moe = False
+
+    def forward(self, hidden_states, attention_mask, position_embeddings, past_key_value=None, use_cache=False):
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states, present = self.self_attn(
+            hidden_states, attention_mask, position_embeddings, past_key_value, use_cache
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        aux_loss = None
+        if self.is_moe:
+            hidden_states, aux_loss = self.mlp(hidden_states)
+        else:
+            hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states, present, aux_loss
+
+
+class MurzikMoePreTrainedModel(MurzikPreTrainedModel):
+    config_class = MurzikMoeConfig
+    _no_split_modules = ["MurzikMoeDecoderLayer"]
+
+
+class MurzikMoeModel(MurzikMoePreTrainedModel):
+    def __init__(self, config: MurzikMoeConfig):
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id)
+        self.layers = nn.ModuleList(
+            [MurzikMoeDecoderLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+        self.norm = MurzikRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = MurzikRotaryEmbedding(
+            config.head_dim, config.max_position_embeddings, config.rope_theta
+        )
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[list] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ):
+        bsz, seq_len = input_ids.shape
+        hidden_states = self.embed_tokens(input_ids)
+        cos, sin = self.rotary_emb(hidden_states, seq_len)
+        position_embeddings = (cos, sin)
+
+        if attention_mask is None:
+            attention_mask = torch.triu(
+                torch.full((seq_len, seq_len), float("-inf"), device=input_ids.device),
+                diagonal=1,
+            ).unsqueeze(0).unsqueeze(0)
+        else:
+            attention_mask = attention_mask[:, None, None, :].to(dtype=hidden_states.dtype)
+            attention_mask = (1.0 - attention_mask) * torch.finfo(hidden_states.dtype).min
+
+        presents = [] if use_cache else None
+        aux_loss = torch.tensor(0.0, device=input_ids.device)
+        for idx, layer in enumerate(self.layers):
+            past = past_key_values[idx] if past_key_values is not None else None
+            hidden_states, present, layer_aux = layer(
+                hidden_states, attention_mask, position_embeddings, past, use_cache
+            )
+            if layer_aux is not None:
+                aux_loss = aux_loss + layer_aux
+            if use_cache:
+                presents.append(present)
+
+        hidden_states = self.norm(hidden_states)
+        return hidden_states, presents, aux_loss
+
+
+class MurzikMoeForCausalLM(MurzikMoePreTrainedModel):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+
+    def __init__(self, config: MurzikMoeConfig):
+        super().__init__(config)
+        self.model = MurzikMoeModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[list] = None,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> CausalLMOutputWithPast:
+        hidden_states, past_key_values, aux_loss = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+        )
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            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),
+                ignore_index=-100,
+            )
+            if aux_loss is not None:
+                loss = loss + aux_loss
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=past_key_values,
+        )

murzik/tokenization_murzik.py

@@ -0,0 +1,90 @@
+"""Murzik tokenizer — SentencePiece wrapper for Hugging Face."""
+
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+from transformers import PreTrainedTokenizer
+
+# Special tokens (must match SFT template)
+SPECIAL_TOKENS = {
+    "pad_token": "<|pad|>",
+    "bos_token": "<|murzik|>",
+    "eos_token": "<|end|>",
+    "unk_token": "<|unk|>",
+    "additional_special_tokens": [
+        "<|user|>",
+        "<|assistant|>",
+        "<|system|>",
+    ],
+}
+
+
+class MurzikTokenizer(PreTrainedTokenizer):
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file: str,
+        bos_token: str = SPECIAL_TOKENS["bos_token"],
+        eos_token: str = SPECIAL_TOKENS["eos_token"],
+        pad_token: str = SPECIAL_TOKENS["pad_token"],
+        unk_token: str = SPECIAL_TOKENS["unk_token"],
+        **kwargs,
+    ):
+        self.vocab_file = vocab_file
+        self.sp_model = spm.SentencePieceProcessor()
+        if vocab_file and Path(vocab_file).exists():
+            self.sp_model.Load(vocab_file)
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            unk_token=unk_token,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return self.sp_model.get_piece_size()
+
+    def get_vocab(self):
+        return {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+
+    def _tokenize(self, text: str) -> list[str]:
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token: str) -> int:
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index: int) -> str:
+        return self.sp_model.id_to_piece(index)
+
+    def convert_tokens_to_string(self, tokens: list[str]) -> str:
+        return self.sp_model.decode(tokens)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        if token_ids_1 is None:
+            return token_ids_0
+        return token_ids_0 + token_ids_1
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+        if token_ids_1 is not None:
+            return ([0] * len(token_ids_0)) + ([1] + [0] * (len(token_ids_1) - 1))
+        return [0] * len(token_ids_0)
+
+    def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
+        if token_ids_1 is None:
+            return len(token_ids_0) * [0]
+        return [0] * (len(token_ids_0) + len(token_ids_1))
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple[str]:
+        out = Path(save_directory) / f"{filename_prefix or ''}murzik.model"
+        if self.vocab_file:
+            import shutil
+            shutil.copy(self.vocab_file, out)
+        return (str(out),)

tokenizer.json

@@ -0,0 +1,116 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|pad|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|murzik|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|end|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "<|unk|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 4,
+      "content": "<|user|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 5,
+      "content": "<|assistant|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 6,
+      "content": "<|system|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "TemplateProcessing",
+    "single": [
+      {
+        "Sequence": {
+          "id": "A",
+          "type_id": 0
+        }
+      }
+    ],
+    "pair": [
+      {
+        "Sequence": {
+          "id": "A",
+          "type_id": 0
+        }
+      },
+      {
+        "Sequence": {
+          "id": "B",
+          "type_id": 1
+        }
+      }
+    ],
+    "special_tokens": {}
+  },
+  "decoder": null,
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {},
+    "merges": []
+  }
+}

tokenizer_config.json

@@ -0,0 +1,9 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<|murzik|>",
+  "eos_token": "<|end|>",
+  "model_max_length": 8192,
+  "pad_token": "<|pad|>",
+  "tokenizer_class": "TokenizersBackend",
+  "unk_token": "<|unk|>"
+}