JiRackDeltaNet_27b.py
| 1 | # ============================================================================= |
| 2 | # COPYRIGHT © 2026 Konstantin Vladimirovich Grabko. ALL RIGHTS RESERVED. |
| 3 | # JiRack Ultra Ternary Transformer |
| 4 | # |
| 5 | # CMS Manhattan JiRack Technology — PATENT PENDING |
| 6 | # |
| 7 | # This code is proprietary. |
| 8 | # Personal and non-commercial research use is allowed. |
| 9 | # Any commercial use, derivative works for profit, or distribution |
| 10 | # requires a paid license and 5% royalty. |
| 11 | # |
| 12 | # Unauthorized commercial use is strictly prohibited. |
| 13 | # Contact: grabko@cmsmanhattan.com |
| 14 | # ============================================================================= |
| 15 | """ |
| 16 | The custom ("hand-rolled") "JiRack Qwen3.8-27B" architecture follows the style of the earlier DS32B project: a single file, a custom `TransformerBlock`, no `AutoConfig`/`AutoModelForCausalLM` inside the architecture class, and no `model_id` in the constructor. All dimensions are hardcoded as constants below—verified via live inspection of the actual checkpoint (`inspect_qwen38_keys.py`, 851 tensors, cross-checked). |
| 17 | |
| 18 | The mathematics for DeltaNet recurrence, convolution, gating, and partial-rotary mRoPE attention were **ported** directly from the actual `modeling_qwen3_5.py` (Apache 2.0, HuggingFace/Qwen) rather than reinvented—to avoid the risk of subtle numerical errors in the delicate recurrence formulas. While the logic was ported, the code was rewritten to match our naming conventions and style, and stripped down: no vision tower, and no `transformers`-style KV-cache classes (full sequence forward passes aren't needed for training; for generation, we use full recomputation at each step without an incremental cache—a deliberate simplification, though the cache could be added separately if desired). |
| 19 | |
| 20 | Submodule names (`linear_attn` / `self_attn`, `in_proj_qkv`, `out_proj`, `mlp.gate_proj`, etc.) intentionally **match** those in the actual `Qwen3_5ForCausalLM` `state_dict`, allowing weights from the HF checkpoint to be loaded directly via `load_state_dict()` without manual key mapping. The module supports the following: |
| 21 | - building the model (JiRackQwen38ForCausalLM) from scratch with random weights; |
| 22 | - accepting a state_dict in the HF Qwen3_5ForCausalLM format (obtained via |
| 23 | `AutoModelForCausalLM.from_pretrained(...).state_dict()` in the converter) |
| 24 | and loading it directly; |
| 25 | - patching the necessary `nn.Linear` layers to `BitLinear` (`TERNARIZE_PATTERNS`); |
| 26 | - saving and loading its own JiRack checkpoints (`save_checkpoint`/`load_checkpoint`). |
| 27 | |
| 28 | The converter (`convert_qwen38_to_jirack.py`) remains aware of the `model_id` |
| 29 | and the HF Hub—but this knowledge resides ONLY within the converter, not the architecture itself. |
| 30 | """ |
| 31 | |
| 32 | import os |
| 33 | import re |
| 34 | import torch |
| 35 | import torch.nn as nn |
| 36 | import torch.nn.functional as F |
| 37 | |
| 38 | |
| 39 | # ===================================================================== |
| 40 | # 1. CONFIRMED АРХИТЕКТУРА -- жёстко зашитые константы (не AutoConfig) |
| 41 | # ===================================================================== |
| 42 | |
| 43 | HIDDEN_SIZE = 5120 |
| 44 | VOCAB_SIZE = 248320 |
| 45 | INTERMEDIATE_SIZE = 17408 |
| 46 | NUM_LAYERS = 64 |
| 47 | RMS_NORM_EPS = 1e-6 |
| 48 | |
| 49 | # Чередование слоёв: full_attention_interval = 4 -> |
| 50 | # 3x linear_attention, 1x full_attention, повторяется на все 64 слоя. |
| 51 | LAYER_TYPES = ["linear_attention"] * 3 + ["full_attention"] |
| 52 | LAYER_TYPES = (LAYER_TYPES * (NUM_LAYERS // len(LAYER_TYPES) + 1))[:NUM_LAYERS] |
| 53 | |
| 54 | # --- Gated DeltaNet (linear_attention) --- |
| 55 | LINEAR_NUM_KEY_HEADS = 16 |
| 56 | LINEAR_KEY_HEAD_DIM = 128 |
| 57 | LINEAR_NUM_VALUE_HEADS = 48 |
| 58 | LINEAR_VALUE_HEAD_DIM = 128 |
| 59 | LINEAR_CONV_KERNEL_DIM = 4 |
| 60 | DELTA_KEY_DIM = LINEAR_NUM_KEY_HEADS * LINEAR_KEY_HEAD_DIM # 2048 |
| 61 | DELTA_VALUE_DIM = LINEAR_NUM_VALUE_HEADS * LINEAR_VALUE_HEAD_DIM # 6144 |
| 62 | DELTA_CONV_DIM = DELTA_KEY_DIM * 2 + DELTA_VALUE_DIM # 10240 |
| 63 | DELTA_CHUNK_SIZE = 64 |
| 64 | |
| 65 | # --- Gated Attention (full_attention) --- |
| 66 | NUM_ATTENTION_HEADS = 24 |
| 67 | NUM_KEY_VALUE_HEADS = 4 |
| 68 | HEAD_DIM = 256 |
| 69 | PARTIAL_ROTARY_FACTOR = 0.25 |
| 70 | ROPE_THETA = 1e7 |
| 71 | MROPE_SECTION = [11, 11, 10] |
| 72 | ATTENTION_BIAS = False |
| 73 | |
| 74 | ARCH_SUMMARY = f""" |
| 75 | Qwen3.8-27B (qwen3_5), hand-rolled JiRack TransformerBlock, text-only. |
| 76 | hidden_size={HIDDEN_SIZE}, vocab_size={VOCAB_SIZE}, |
| 77 | intermediate_size={INTERMEDIATE_SIZE}, num_hidden_layers={NUM_LAYERS} |
| 78 | Чередование (interval=4): 3x linear_attention (DeltaNet), 1x full_attention |
| 79 | """ |
| 80 | |
| 81 | KEPT_FULL_PRECISION = """ |
| 82 | The following always remain at full precision (not GEMM, or otherwise too sensitive): |
| 83 | - dt_bias, A_log (raw Parameters, DeltaNet decay) |
| 84 | - linear_attn.in_proj_b/in_proj_a (default behavior; see GATE_PROJ_PATTERNS) |
| 85 | - conv1d (depthwise grouped conv; not GEMM) |
| 86 | - all norms: RMSNormGated, q_norm/k_norm, model.norm |
| 87 | - embed_tokens, lm_head |
| 88 | - mtp.* (entire MTP/NextN head) (speculative decoding draft layer; |
| 89 | ternarization destroys acceptance rate; |
| 90 | the GGUF community keeps it at Q8_0) |
| 91 | """ |
| 92 | |
| 93 | |
| 94 | # ===================================================================== |
| 95 | # 2. BitLinear -- наш тернарный слой |
| 96 | # ===================================================================== |
| 97 | |
| 98 | class BitLinear(nn.Linear): |
| 99 | """b1.58-style per-tensor absmean ternary weights, per-token int8 |
| 100 | activations, непрерывный lambda-warmup (STE).""" |
| 101 | |
| 102 | def __init__(self, in_features, out_features, bias=False): |
| 103 | super().__init__(in_features, out_features, bias=bias) |
| 104 | self.eps = 1e-5 |
| 105 | self.register_buffer("lambda_", torch.zeros(()), persistent=True) |
| 106 | |
| 107 | @classmethod |
| 108 | def from_linear(cls, lin: nn.Linear) -> "BitLinear": |
| 109 | bl = cls(lin.in_features, lin.out_features, bias=lin.bias is not None) |
| 110 | bl = bl.to(dtype=lin.weight.dtype, device=lin.weight.device) |
| 111 | with torch.no_grad(): |
| 112 | bl.weight.copy_(lin.weight) |
| 113 | if lin.bias is not None: |
| 114 | bl.bias.copy_(lin.bias) |
| 115 | return bl |
| 116 | |
| 117 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 118 | if not self.training and float(self.lambda_) < 1e-6: |
| 119 | return F.linear(x, self.weight, self.bias) |
| 120 | |
| 121 | lam = self.lambda_.to(x.dtype) |
| 122 | w = self.weight |
| 123 | gamma = w.float().abs().mean().clamp(min=self.eps).to(w.dtype) |
| 124 | w_quant = torch.clamp(torch.round(w / gamma), -1.0, 1.0) * gamma |
| 125 | w_effective = w + lam * (w_quant - w).detach() |
| 126 | |
| 127 | x_scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp(min=self.eps) |
| 128 | x_quant = torch.clamp(torch.round(x * x_scale), -128.0, 127.0) / x_scale |
| 129 | x_effective = x + lam * (x_quant - x).detach() |
| 130 | |
| 131 | return F.linear(x_effective, w_effective, self.bias) |
| 132 | |
| 133 | def set_lambda(self, v: float): |
| 134 | self.lambda_.fill_(v) |
| 135 | |
| 136 | |
| 137 | # ===================================================================== |
| 138 | # 3. Карта: что тернаризуем, что оставляем full precision |
| 139 | # ===================================================================== |
| 140 | |
| 141 | TERNARIZE_PATTERNS = [ |
| 142 | r"linear_attn\.in_proj_qkv$", |
| 143 | r"linear_attn\.in_proj_z$", |
| 144 | r"linear_attn\.out_proj$", |
| 145 | r"self_attn\.q_proj$", |
| 146 | r"self_attn\.k_proj$", |
| 147 | r"self_attn\.v_proj$", |
| 148 | r"self_attn\.o_proj$", |
| 149 | r"mlp\.gate_proj$", |
| 150 | r"mlp\.up_proj$", |
| 151 | r"mlp\.down_proj$", |
| 152 | ] |
| 153 | |
| 154 | GATE_PROJ_PATTERNS = [ |
| 155 | r"linear_attn\.in_proj_b$", |
| 156 | r"linear_attn\.in_proj_a$", |
| 157 | ] |
| 158 | |
| 159 | |
| 160 | def apply_bitlinear_patch(model: nn.Module, include_gate_proj: bool = False): |
| 161 | """Точечная замена nn.Linear -> BitLinear по TERNARIZE_PATTERNS.""" |
| 162 | patterns = list(TERNARIZE_PATTERNS) |
| 163 | if include_gate_proj: |
| 164 | patterns += GATE_PROJ_PATTERNS |
| 165 | compiled = [re.compile(p) for p in patterns] |
| 166 | |
| 167 | patched = [] |
| 168 | for name, module in list(model.named_modules()): |
| 169 | if not isinstance(module, nn.Linear): |
| 170 | continue |
| 171 | # MTP/NextN-голова НЕ тернаризуется -- всегда full precision. |
| 172 | # Суффиксные паттерны (self_attn.q_proj$ и т.п.) иначе зацепили бы |
| 173 | # mtp.layers.0.self_attn.q_proj. |
| 174 | if name.startswith("mtp"): |
| 175 | continue |
| 176 | if any(p.search(name) for p in compiled): |
| 177 | parent_name, _, child_name = name.rpartition(".") |
| 178 | parent = model.get_submodule(parent_name) if parent_name else model |
| 179 | setattr(parent, child_name, BitLinear.from_linear(module)) |
| 180 | patched.append(name) |
| 181 | return patched |
| 182 | |
| 183 | |
| 184 | # ===================================================================== |
| 185 | # 4. Общие блоки (порт математики из реального modeling_qwen3_5.py) |
| 186 | # ===================================================================== |
| 187 | |
| 188 | class RMSNorm(nn.Module): |
| 189 | """(x * w).to(dtype), w хранится как (weight) с offset +1 -- как в |
| 190 | оригинале Qwen3_5RMSNorm (инициализация нулями, эффективный вес 1+w).""" |
| 191 | |
| 192 | def __init__(self, dim: int, eps: float = RMS_NORM_EPS): |
| 193 | super().__init__() |
| 194 | self.eps = eps |
| 195 | self.weight = nn.Parameter(torch.zeros(dim)) |
| 196 | |
| 197 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 198 | dtype = x.dtype |
| 199 | x = x.float() |
| 200 | x = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) |
| 201 | x = x * (1.0 + self.weight.float()) |
| 202 | return x.to(dtype) |
| 203 | |
| 204 | |
| 205 | class RMSNormGated(nn.Module): |
| 206 | """RMSNorm с гейтингом через SiLU(gate), применяется ПЕРЕД гейтингом |
| 207 | (норма считается по hidden_states, гейт применяется после).""" |
| 208 | |
| 209 | def __init__(self, dim: int, eps: float = RMS_NORM_EPS): |
| 210 | super().__init__() |
| 211 | self.weight = nn.Parameter(torch.ones(dim)) |
| 212 | self.eps = eps |
| 213 | |
| 214 | def forward(self, x: torch.Tensor, gate: torch.Tensor) -> torch.Tensor: |
| 215 | dtype = x.dtype |
| 216 | x = x.float() |
| 217 | var = x.pow(2).mean(-1, keepdim=True) |
| 218 | x = x * torch.rsqrt(var + self.eps) |
| 219 | x = self.weight * x.to(dtype) |
| 220 | x = x * F.silu(gate.float()) |
| 221 | return x.to(dtype) |
| 222 | |
| 223 | |
| 224 | def rotate_half(x: torch.Tensor) -> torch.Tensor: |
| 225 | x1 = x[..., : x.shape[-1] // 2] |
| 226 | x2 = x[..., x.shape[-1] // 2:] |
| 227 | return torch.cat((-x2, x1), dim=-1) |
| 228 | |
| 229 | |
| 230 | def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1): |
| 231 | """Partial rotary: только первые cos.shape[-1] каналов вращаются, |
| 232 | остальное (q_pass/k_pass) проходит без изменений.""" |
| 233 | cos = cos.unsqueeze(unsqueeze_dim) |
| 234 | sin = sin.unsqueeze(unsqueeze_dim) |
| 235 | rotary_dim = cos.shape[-1] |
| 236 | q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] |
| 237 | k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] |
| 238 | q_embed = torch.cat([(q_rot * cos) + (rotate_half(q_rot) * sin), q_pass], dim=-1) |
| 239 | k_embed = torch.cat([(k_rot * cos) + (rotate_half(k_rot) * sin), k_pass], dim=-1) |
| 240 | return q_embed, k_embed |
| 241 | |
| 242 | |
| 243 | class TextRotaryEmbedding(nn.Module): |
| 244 | """mRoPE с partial_rotary_factor. Для чистого текста (без vision) |
| 245 | 3 позиционные "оси" (T,H,W) совпадают -> interleaved mRoPE |
| 246 | вырождается в обычный RoPE, но мы всё равно считаем честно тем же |
| 247 | путём, что и оригинал, для полной числовой идентичности.""" |
| 248 | |
| 249 | def __init__(self, head_dim=HEAD_DIM, partial_rotary_factor=PARTIAL_ROTARY_FACTOR, |
| 250 | theta=ROPE_THETA, mrope_section=MROPE_SECTION): |
| 251 | super().__init__() |
| 252 | self.rotary_dim = int(head_dim * partial_rotary_factor) |
| 253 | inv_freq = 1.0 / (theta ** (torch.arange(0, self.rotary_dim, 2, dtype=torch.float32) / self.rotary_dim)) |
| 254 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 255 | self.mrope_section = mrope_section |
| 256 | |
| 257 | def _apply_interleaved_mrope(self, freqs: torch.Tensor) -> torch.Tensor: |
| 258 | # freqs: (3, bs, seq, rotary_dim // 2) |
| 259 | freqs_t = freqs[0].clone() |
| 260 | for dim, offset in enumerate((1, 2), start=1): |
| 261 | length = self.mrope_section[dim] * 3 |
| 262 | idx = slice(offset, length, 3) |
| 263 | freqs_t[..., idx] = freqs[dim, ..., idx] |
| 264 | return freqs_t |
| 265 | |
| 266 | @torch.no_grad() |
| 267 | def forward(self, x: torch.Tensor, position_ids: torch.LongTensor): |
| 268 | # position_ids: (bs, seq) -> реплицируем в 3 идентичные оси (text-only) |
| 269 | if position_ids.ndim == 2: |
| 270 | position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1) |
| 271 | inv_freq_exp = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1) |
| 272 | pos_exp = position_ids[:, :, None, :].float() |
| 273 | freqs = (inv_freq_exp @ pos_exp).transpose(2, 3) # (3, bs, seq, rotary_dim//2) |
| 274 | freqs = self._apply_interleaved_mrope(freqs) # (bs, seq, rotary_dim//2) |
| 275 | emb = torch.cat((freqs, freqs), dim=-1) # (bs, seq, rotary_dim) |
| 276 | return emb.cos().to(x.dtype), emb.sin().to(x.dtype) |
| 277 | |
| 278 | |
| 279 | def l2norm(x: torch.Tensor, dim: int = -1, eps: float = 1e-6): |
| 280 | return x * torch.rsqrt((x * x).sum(dim=dim, keepdim=True) + eps) |
| 281 | |
| 282 | |
| 283 | def causal_conv1d(hidden_states: torch.Tensor, weight: torch.Tensor, activation: str = "silu"): |
| 284 | """hidden_states: (batch, channels, seq_len). Depthwise causal conv, |
| 285 | left-padded by kernel_size - 1, then activation.""" |
| 286 | channels, seq_len = hidden_states.shape[1], hidden_states.shape[2] |
| 287 | padding = weight.shape[-1] - 1 |
| 288 | out = F.conv1d(hidden_states, weight=weight.unsqueeze(1), bias=None, padding=padding, groups=channels) |
| 289 | out = out[:, :, :seq_len] |
| 290 | return F.silu(out) if activation == "silu" else out |
| 291 | |
| 292 | |
| 293 | def chunked_gated_delta_rule(query, key, value, g, beta, chunk_size=DELTA_CHUNK_SIZE): |
| 294 | """Port of `torch_chunk_gated_delta_rule` from `modeling_qwen3_5.py` — excluding |
| 295 | `initial_state`/`output_final_state` (training/full forward pass without |
| 296 | incremental cache). query/key/value: (b, h, t, d), g/beta: (b, h, t).""" |
| 297 | initial_dtype = query.dtype |
| 298 | query = l2norm(query, dim=-1, eps=1e-6) |
| 299 | key = l2norm(key, dim=-1, eps=1e-6) |
| 300 | query, key, value, beta, g = [x.contiguous().to(torch.float32) for x in (query, key, value, beta, g)] |
| 301 | |
| 302 | b, h, t, k_dim = key.shape |
| 303 | v_dim = value.shape[-1] |
| 304 | pad = (chunk_size - t % chunk_size) % chunk_size |
| 305 | query = F.pad(query, (0, 0, 0, pad)) |
| 306 | key = F.pad(key, (0, 0, 0, pad)) |
| 307 | value = F.pad(value, (0, 0, 0, pad)) |
| 308 | beta = F.pad(beta, (0, pad)) |
| 309 | g = F.pad(g, (0, pad)) |
| 310 | total_t = t + pad |
| 311 | scale = 1 / (query.shape[-1] ** 0.5) |
| 312 | query = query * scale |
| 313 | |
| 314 | v_beta = value * beta.unsqueeze(-1) |
| 315 | k_beta = key * beta.unsqueeze(-1) |
| 316 | query, key, value, k_beta, v_beta = [ |
| 317 | x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1]) for x in (query, key, value, k_beta, v_beta) |
| 318 | ] |
| 319 | g = g.reshape(g.shape[0], g.shape[1], -1, chunk_size) |
| 320 | mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=0) |
| 321 | |
| 322 | g = g.cumsum(dim=-1) |
| 323 | decay_mask = ((g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().float()).tril() |
| 324 | attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0) |
| 325 | for i in range(1, chunk_size): |
| 326 | row = attn[..., i, :i].clone() |
| 327 | sub = attn[..., :i, :i].clone() |
| 328 | attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2) |
| 329 | attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device) |
| 330 | value = attn @ v_beta |
| 331 | k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1)) |
| 332 | |
| 333 | last_state = torch.zeros(b, h, k_dim, v_dim, dtype=value.dtype, device=value.device) |
| 334 | core_out = torch.zeros_like(value) |
| 335 | |
| 336 | for i in range(total_t // chunk_size): |
| 337 | q_i, k_i, v_i = query[:, :, i], key[:, :, i], value[:, :, i] |
| 338 | attn_i = q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i] |
| 339 | v_prime = k_cumdecay[:, :, i] @ last_state |
| 340 | v_new = v_i - v_prime |
| 341 | attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_state |
| 342 | core_out[:, :, i] = attn_inter + attn_i @ v_new |
| 343 | last_state = ( |
| 344 | last_state * g[:, :, i, -1, None, None].exp() |
| 345 | + (k_i * (g[:, :, i, -1, None] - g[:, :, i]).exp()[..., None]).transpose(-1, -2) @ v_new |
| 346 | ) |
| 347 | |
| 348 | core_out = core_out.reshape(core_out.shape[0], core_out.shape[1], -1, core_out.shape[-1]) |
| 349 | core_out = core_out[:, :, :t] |
| 350 | core_out = core_out.transpose(1, 2).contiguous().to(initial_dtype) |
| 351 | return core_out |
| 352 | |
| 353 | |
| 354 | # ===================================================================== |
| 355 | # 5. Наш собственный TransformerBlock (и подмодули linear_attn/self_attn) |
| 356 | # ===================================================================== |
| 357 | |
| 358 | class GatedDeltaNetBlock(nn.Module): |
| 359 | """Linear attention layer (Gated DeltaNet). Submodule names match the HF state_dict so that weights load directly.""" |
| 360 | |
| 361 | def __init__(self): |
| 362 | super().__init__() |
| 363 | self.conv1d = nn.Conv1d( |
| 364 | in_channels=DELTA_CONV_DIM, out_channels=DELTA_CONV_DIM, bias=False, |
| 365 | kernel_size=LINEAR_CONV_KERNEL_DIM, groups=DELTA_CONV_DIM, |
| 366 | padding=LINEAR_CONV_KERNEL_DIM - 1, |
| 367 | ) |
| 368 | self.dt_bias = nn.Parameter(torch.ones(LINEAR_NUM_VALUE_HEADS)) |
| 369 | A = torch.empty(LINEAR_NUM_VALUE_HEADS).uniform_(0.01, 16) |
| 370 | self.A_log = nn.Parameter(torch.log(A)) |
| 371 | self.norm = RMSNormGated(LINEAR_VALUE_HEAD_DIM, eps=RMS_NORM_EPS) |
| 372 | self.out_proj = nn.Linear(DELTA_VALUE_DIM, HIDDEN_SIZE, bias=False) |
| 373 | self.in_proj_qkv = nn.Linear(HIDDEN_SIZE, DELTA_KEY_DIM * 2 + DELTA_VALUE_DIM, bias=False) |
| 374 | self.in_proj_z = nn.Linear(HIDDEN_SIZE, DELTA_VALUE_DIM, bias=False) |
| 375 | self.in_proj_b = nn.Linear(HIDDEN_SIZE, LINEAR_NUM_VALUE_HEADS, bias=False) |
| 376 | self.in_proj_a = nn.Linear(HIDDEN_SIZE, LINEAR_NUM_VALUE_HEADS, bias=False) |
| 377 | |
| 378 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 379 | bsz, seq_len, _ = hidden_states.shape |
| 380 | |
| 381 | mixed_qkv = self.in_proj_qkv(hidden_states).transpose(1, 2) # (b, conv_dim, t) |
| 382 | z = self.in_proj_z(hidden_states).reshape(bsz, seq_len, -1, LINEAR_VALUE_HEAD_DIM) |
| 383 | b = self.in_proj_b(hidden_states) |
| 384 | a = self.in_proj_a(hidden_states) |
| 385 | |
| 386 | mixed_qkv = causal_conv1d(mixed_qkv, self.conv1d.weight.squeeze(1), activation="silu") |
| 387 | mixed_qkv = mixed_qkv.transpose(1, 2) # (b, t, conv_dim) |
| 388 | |
| 389 | query, key, value = torch.split(mixed_qkv, [DELTA_KEY_DIM, DELTA_KEY_DIM, DELTA_VALUE_DIM], dim=-1) |
| 390 | query = query.reshape(bsz, seq_len, -1, LINEAR_KEY_HEAD_DIM) |
| 391 | key = key.reshape(bsz, seq_len, -1, LINEAR_KEY_HEAD_DIM) |
| 392 | value = value.reshape(bsz, seq_len, -1, LINEAR_VALUE_HEAD_DIM) |
| 393 | |
| 394 | beta = b.sigmoid() |
| 395 | g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias) |
| 396 | |
| 397 | rep = LINEAR_NUM_VALUE_HEADS // LINEAR_NUM_KEY_HEADS |
| 398 | if rep > 1: |
| 399 | query = query.repeat_interleave(rep, dim=2) |
| 400 | key = key.repeat_interleave(rep, dim=2) |
| 401 | |
| 402 | # -> (b, h, t, d) for the chunked kernel |
| 403 | query_h = query.transpose(1, 2) |
| 404 | key_h = key.transpose(1, 2) |
| 405 | value_h = value.transpose(1, 2) |
| 406 | beta_h = beta.transpose(1, 2) |
| 407 | g_h = g.transpose(1, 2) |
| 408 | |
| 409 | core_attn_out = chunked_gated_delta_rule(query_h, key_h, value_h, g_h, beta_h) |
| 410 | |
| 411 | core_attn_out = core_attn_out.reshape(-1, LINEAR_VALUE_HEAD_DIM) |
| 412 | z_flat = z.reshape(-1, LINEAR_VALUE_HEAD_DIM) |
| 413 | core_attn_out = self.norm(core_attn_out, z_flat) |
| 414 | core_attn_out = core_attn_out.reshape(bsz, seq_len, -1) |
| 415 | |
| 416 | return self.out_proj(core_attn_out) |
| 417 | |
| 418 | |
| 419 | class AttentionBlock(nn.Module): |
| 420 | """full_attention layer (Gated Attention + partial rotary mRoPE).""" |
| 421 | |
| 422 | def __init__(self): |
| 423 | super().__init__() |
| 424 | self.num_key_value_groups = NUM_ATTENTION_HEADS // NUM_KEY_VALUE_HEADS |
| 425 | self.scaling = HEAD_DIM ** -0.5 |
| 426 | self.q_proj = nn.Linear(HIDDEN_SIZE, NUM_ATTENTION_HEADS * HEAD_DIM * 2, bias=ATTENTION_BIAS) |
| 427 | self.k_proj = nn.Linear(HIDDEN_SIZE, NUM_KEY_VALUE_HEADS * HEAD_DIM, bias=ATTENTION_BIAS) |
| 428 | self.v_proj = nn.Linear(HIDDEN_SIZE, NUM_KEY_VALUE_HEADS * HEAD_DIM, bias=ATTENTION_BIAS) |
| 429 | self.o_proj = nn.Linear(NUM_ATTENTION_HEADS * HEAD_DIM, HIDDEN_SIZE, bias=ATTENTION_BIAS) |
| 430 | self.q_norm = RMSNorm(HEAD_DIM, eps=RMS_NORM_EPS) |
| 431 | self.k_norm = RMSNorm(HEAD_DIM, eps=RMS_NORM_EPS) |
| 432 | |
| 433 | def forward(self, hidden_states, cos, sin, causal_mask): |
| 434 | bsz, seq_len, _ = hidden_states.shape |
| 435 | hidden_shape = (bsz, seq_len, -1, HEAD_DIM) |
| 436 | |
| 437 | query_states, gate = torch.chunk( |
| 438 | self.q_proj(hidden_states).view(bsz, seq_len, -1, HEAD_DIM * 2), 2, dim=-1 |
| 439 | ) |
| 440 | gate = gate.reshape(bsz, seq_len, -1) |
| 441 | |
| 442 | query_states = self.q_norm(query_states.view(hidden_shape)).transpose(1, 2) |
| 443 | key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2) |
| 444 | value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) |
| 445 | |
| 446 | query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) |
| 447 | |
| 448 | rep = self.num_key_value_groups |
| 449 | key_states = key_states.repeat_interleave(rep, dim=1) |
| 450 | value_states = value_states.repeat_interleave(rep, dim=1) |
| 451 | |
| 452 | attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scaling |
| 453 | attn_weights = attn_weights + causal_mask |
| 454 | attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) |
| 455 | attn_output = torch.matmul(attn_weights, value_states) |
| 456 | attn_output = attn_output.transpose(1, 2).contiguous().reshape(bsz, seq_len, -1) |
| 457 | attn_output = attn_output * torch.sigmoid(gate) |
| 458 | |
| 459 | return self.o_proj(attn_output) |
| 460 | |
| 461 | |
| 462 | class MLPBlock(nn.Module): |
| 463 | def __init__(self): |
| 464 | super().__init__() |
| 465 | self.gate_proj = nn.Linear(HIDDEN_SIZE, INTERMEDIATE_SIZE, bias=False) |
| 466 | self.up_proj = nn.Linear(HIDDEN_SIZE, INTERMEDIATE_SIZE, bias=False) |
| 467 | self.down_proj = nn.Linear(INTERMEDIATE_SIZE, HIDDEN_SIZE, bias=False) |
| 468 | |
| 469 | def forward(self, x): |
| 470 | return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x)) |
| 471 | |
| 472 | |
| 473 | class TransformerBlock(nn.Module): |
| 474 | """Наш decoder layer -- linear_attn ИЛИ self_attn, в зависимости от |
| 475 | layer_type, + общий MLP. Имена атрибутов совпадают с HF state_dict.""" |
| 476 | |
| 477 | def __init__(self, layer_idx: int, layer_type: str = None): |
| 478 | super().__init__() |
| 479 | # layer_type задан явно -> используем его (нужно для MTP-слоя, |
| 480 | # который всегда full_attention и лежит ВНЕ LAYER_TYPES). |
| 481 | self.block_type = layer_type if layer_type is not None else LAYER_TYPES[layer_idx] |
| 482 | if self.block_type == "linear_attention": |
| 483 | self.linear_attn = GatedDeltaNetBlock() |
| 484 | else: |
| 485 | self.self_attn = AttentionBlock() |
| 486 | self.mlp = MLPBlock() |
| 487 | self.input_layernorm = RMSNorm(HIDDEN_SIZE, eps=RMS_NORM_EPS) |
| 488 | self.post_attention_layernorm = RMSNorm(HIDDEN_SIZE, eps=RMS_NORM_EPS) |
| 489 | |
| 490 | def forward(self, hidden_states, cos, sin, causal_mask): |
| 491 | residual = hidden_states |
| 492 | hidden_states = self.input_layernorm(hidden_states) |
| 493 | if self.block_type == "linear_attention": |
| 494 | hidden_states = self.linear_attn(hidden_states) |
| 495 | else: |
| 496 | hidden_states = self.self_attn(hidden_states, cos, sin, causal_mask) |
| 497 | hidden_states = residual + hidden_states |
| 498 | |
| 499 | residual = hidden_states |
| 500 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 501 | hidden_states = self.mlp(hidden_states) |
| 502 | hidden_states = residual + hidden_states |
| 503 | return hidden_states |
| 504 | |
| 505 | |
| 506 | # ===================================================================== |
| 507 | # 5b. MTP / NextN-голова (draft-слой спекулятивного декода) |
| 508 | # ===================================================================== |
| 509 | |
| 510 | class MTPBlock(nn.Module): |
| 511 | """Multi-Token Prediction голова -- ровно как в реальном чекпоинте |
| 512 | Qwen3.5/3.8 (mtp_num_hidden_layers=1, mtp_use_dedicated_embeddings=false). |
| 513 | |
| 514 | Имена подмодулей СОВПАДАЮТ с ключами HF state_dict (top-level `mtp.`): |
| 515 | mtp.pre_fc_norm_embedding.weight |
| 516 | mtp.pre_fc_norm_hidden.weight |
| 517 | mtp.fc.weight [HIDDEN, 2*HIDDEN] |
| 518 | mtp.layers.0.* (обычный full_attention decoder) |
| 519 | mtp.norm.weight (final norm перед общим lm_head) |
| 520 | |
| 521 | embed_tokens и lm_head ОБЩИЕ со стволом (dedicated embeddings = false), |
| 522 | здесь их нет. В GGUF это blk.{NUM_LAYERS}.nextn.eh_proj / |
| 523 | shared_head_norm + тензоры draft-слоя. |
| 524 | |
| 525 | Механика (DeepSeek-style, порядок как в vLLM Qwen3NextMTP): |
| 526 | h = fc(cat(pre_fc_norm_embedding(emb_next), pre_fc_norm_hidden(h_prev))) |
| 527 | h = layers[0](h) # full attention |
| 528 | logits2 = lm_head(norm(h)) |
| 529 | """ |
| 530 | |
| 531 | def __init__(self): |
| 532 | super().__init__() |
| 533 | self.pre_fc_norm_embedding = RMSNorm(HIDDEN_SIZE, eps=RMS_NORM_EPS) |
| 534 | self.pre_fc_norm_hidden = RMSNorm(HIDDEN_SIZE, eps=RMS_NORM_EPS) |
| 535 | self.fc = nn.Linear(2 * HIDDEN_SIZE, HIDDEN_SIZE, bias=False) |
| 536 | self.layers = nn.ModuleList([ |
| 537 | TransformerBlock(layer_idx=NUM_LAYERS, layer_type="full_attention"), |
| 538 | ]) |
| 539 | self.norm = RMSNorm(HIDDEN_SIZE, eps=RMS_NORM_EPS) |
| 540 | |
| 541 | def forward(self, inputs_embeds, previous_hidden, cos, sin, causal_mask): |
| 542 | """inputs_embeds -- embed_tokens(токены, сдвинутые на +1) |
| 543 | previous_hidden -- hidden ствола ДО финального model.norm |
| 544 | (см. контракт: MTP сам применяет pre_fc_norm_hidden).""" |
| 545 | h = self.fc(torch.cat( |
| 546 | (self.pre_fc_norm_embedding(inputs_embeds), |
| 547 | self.pre_fc_norm_hidden(previous_hidden)), |
| 548 | dim=-1, |
| 549 | )) |
| 550 | h = self.layers[0](h, cos, sin, causal_mask) |
| 551 | return self.norm(h) |
| 552 | |
| 553 | |
| 554 | # ===================================================================== |
| 555 | # 6. Полная модель |
| 556 | # ===================================================================== |
| 557 | |
| 558 | class JiRackQwen38Model(nn.Module): |
| 559 | def __init__(self): |
| 560 | super().__init__() |
| 561 | self.embed_tokens = nn.Embedding(VOCAB_SIZE, HIDDEN_SIZE) |
| 562 | self.layers = nn.ModuleList([TransformerBlock(i) for i in range(NUM_LAYERS)]) |
| 563 | self.norm = RMSNorm(HIDDEN_SIZE, eps=RMS_NORM_EPS) |
| 564 | self.rotary_emb = TextRotaryEmbedding() |
| 565 | |
| 566 | def forward(self, input_ids: torch.LongTensor, return_prenorm: bool = False): |
| 567 | bsz, seq_len = input_ids.shape |
| 568 | hidden_states = self.embed_tokens(input_ids) |
| 569 | |
| 570 | position_ids = torch.arange(seq_len, device=input_ids.device).unsqueeze(0).expand(bsz, -1) |
| 571 | cos, sin = self.rotary_emb(hidden_states, position_ids) |
| 572 | |
| 573 | causal_mask = torch.full((seq_len, seq_len), float("-inf"), device=input_ids.device, dtype=torch.float32) |
| 574 | causal_mask = torch.triu(causal_mask, diagonal=1)[None, None, :, :].to(hidden_states.dtype) |
| 575 | |
| 576 | for layer in self.layers: |
| 577 | hidden_states = layer(hidden_states, cos, sin, causal_mask) |
| 578 | |
| 579 | if return_prenorm: |
| 580 | # Для MTP: голова сама применяет pre_fc_norm_hidden, финальный |
| 581 | # model.norm сюда НЕ входит (иначе двойная нормализация). |
| 582 | return self.norm(hidden_states), hidden_states, (cos, sin, causal_mask) |
| 583 | return self.norm(hidden_states) |
| 584 | |
| 585 | |
| 586 | class JiRackQwen38ForCausalLM(nn.Module): |
| 587 | """Верхнеуровневая модель. Имена атрибутов (`model`, `lm_head`) |
| 588 | совпадают с HF Qwen3_5ForCausalLM -- state_dict грузится напрямую.""" |
| 589 | |
| 590 | def __init__(self): |
| 591 | super().__init__() |
| 592 | self.model = JiRackQwen38Model() |
| 593 | self.lm_head = nn.Linear(HIDDEN_SIZE, VOCAB_SIZE, bias=False) |
| 594 | # MTP/NextN-голова: top-level атрибут `mtp` -- ключи state_dict |
| 595 | # получаются ровно mtp.*, как в реальном HF-чекпоинте. |
| 596 | self.mtp = MTPBlock() |
| 597 | |
| 598 | def forward(self, input_ids: torch.LongTensor) -> torch.Tensor: |
| 599 | hidden_states = self.model(input_ids) |
| 600 | return self.lm_head(hidden_states) |
| 601 | |
| 602 | def mtp_forward(self, input_ids: torch.LongTensor): |
| 603 | """Возвращает (logits_t1, logits_t2): обычные логиты ствола и |
| 604 | draft-логиты MTP-головы (предсказание токена t+2 из позиции t). |
| 605 | Для инференса/валидации MTP; в GGUF-конвертации не участвует.""" |
| 606 | _, prenorm_hidden, (cos, sin, causal_mask) = self.model( |
| 607 | input_ids, return_prenorm=True) |
| 608 | logits_t1 = self.lm_head(self.model.norm(prenorm_hidden)) |
| 609 | |
| 610 | # Вход MTP -- эмбеддинги токенов, сдвинутых на +1 (следующий токен |
| 611 | # уже известен на шаге верификации). Последнюю позицию дублируем, |
| 612 | # чтобы сохранить форму (draft для неё не используется). |
| 613 | shifted = torch.cat([input_ids[:, 1:], input_ids[:, -1:]], dim=1) |
| 614 | emb_next = self.model.embed_tokens(shifted) |
| 615 | h2 = self.mtp(emb_next, prenorm_hidden, cos, sin, causal_mask) |
| 616 | logits_t2 = self.lm_head(h2) |
| 617 | return logits_t1, logits_t2 |
| 618 | |
| 619 | # ----------------------------------------------------------------- |
| 620 | # Загрузка реальных весов (state_dict уже скачан конвертером через |
| 621 | # HF Hub -- сюда просто передаётся словарь тензоров) |
| 622 | # ----------------------------------------------------------------- |
| 623 | def load_hf_state_dict(self, state_dict: dict, strict: bool = False): |
| 624 | """Грузит state_dict в формате оригинального Qwen3_5ForCausalLM. |
| 625 | mtp.* ключи теперь ГРУЗЯТСЯ в self.mtp (MTPBlock) -- фильтра |
| 626 | больше нет. strict=False оставлен из-за служебных lambda_-буферов |
| 627 | BitLinear, которых нет в реальном чекпоинте.""" |
| 628 | missing, unexpected = self.load_state_dict(state_dict, strict=strict) |
| 629 | real_missing = [k for k in missing if not k.endswith("lambda_")] |
| 630 | if real_missing: |
| 631 | print(f"⚠️ {len(real_missing)} missing keys (первые 10): {real_missing[:10]}") |
| 632 | if unexpected: |
| 633 | print(f"⚠️ {len(unexpected)} unexpected keys (первые 10): {unexpected[:10]}") |
| 634 | return missing, unexpected |
| 635 | |
| 636 | # ----------------------------------------------------------------- |
| 637 | # Подсадка РЕАЛЬНЫХ mtp.* весов из оригинального HF-чекпоинта. |
| 638 | # Нужна, потому что старые JiRack .pt сохранены ДО появления MTPBlock |
| 639 | # и mtp-весов не содержат (там была бы случайная инициализация). |
| 640 | # ----------------------------------------------------------------- |
| 641 | def load_mtp_from_hf_dir(self, hf_dir: str): |
| 642 | """Читает из папки оригинального HF-чекпоинта ТОЛЬКО mtp.* тензоры |
| 643 | (по index.json или перебором *.safetensors) и грузит их в self.mtp. |
| 644 | Возвращает число загруженных тензоров.""" |
| 645 | import glob |
| 646 | import json as _json |
| 647 | from safetensors import safe_open |
| 648 | |
| 649 | index_path = os.path.join(hf_dir, "model.safetensors.index.json") |
| 650 | mtp_map = {} # key -> shard file |
| 651 | if os.path.exists(index_path): |
| 652 | with open(index_path) as f: |
| 653 | weight_map = _json.load(f)["weight_map"] |
| 654 | for k, shard in weight_map.items(): |
| 655 | if k.startswith("mtp."): |
| 656 | mtp_map[k] = os.path.join(hf_dir, shard) |
| 657 | else: |
| 658 | for shard in glob.glob(os.path.join(hf_dir, "*.safetensors")): |
| 659 | with safe_open(shard, framework="pt", device="cpu") as f: |
| 660 | for k in f.keys(): |
| 661 | if k.startswith("mtp."): |
| 662 | mtp_map[k] = shard |
| 663 | |
| 664 | if not mtp_map: |
| 665 | raise FileNotFoundError( |
| 666 | f"В {hf_dir} не найдено ни одного mtp.* тензора. Нужен " |
| 667 | f"ОРИГИНАЛЬНЫЙ чекпоинт Qwen3.8-27B (не JiRack-экспорт!). " |
| 668 | f"Если оригинальные шарды удалены -- перекачай с HF Hub " |
| 669 | f"только шарды с mtp.* по model.safetensors.index.json.") |
| 670 | |
| 671 | own_dtype = next(self.mtp.parameters()).dtype |
| 672 | loaded = {} |
| 673 | by_shard = {} |
| 674 | for k, shard in mtp_map.items(): |
| 675 | by_shard.setdefault(shard, []).append(k) |
| 676 | for shard, keys in by_shard.items(): |
| 677 | with safe_open(shard, framework="pt", device="cpu") as f: |
| 678 | for k in keys: |
| 679 | loaded[k] = f.get_tensor(k).to(own_dtype) |
| 680 | |
| 681 | missing, unexpected = self.load_state_dict(loaded, strict=False) |
| 682 | n_ok = len(loaded) |
| 683 | print(f"✅ MTP: загружено {n_ok} тензоров из {len(by_shard)} шардов ({hf_dir})") |
| 684 | if unexpected: |
| 685 | print(f"⚠️ MTP: unexpected keys: {unexpected[:10]}") |
| 686 | return n_ok |
| 687 | |
| 688 | # ----------------------------------------------------------------- |
| 689 | # JiRack-чекпоинт (свой формат, с lambda/patched_linear_names) |
| 690 | # ----------------------------------------------------------------- |
| 691 | def save_checkpoint(self, path: str, lam: float, global_step: int = 0, patched_linear_names=None): |
| 692 | torch.save({ |
| 693 | "model": self.state_dict(), |
| 694 | "lambda": lam, |
| 695 | "global_step": global_step, |
| 696 | "patched_linear_names": patched_linear_names or [], |
| 697 | }, path) |
| 698 | |
| 699 | @classmethod |
| 700 | def load_checkpoint(cls, path: str, dtype=torch.bfloat16, include_gate_proj: bool = False): |
| 701 | ckpt = torch.load(path, map_location="cpu", weights_only=False) |
| 702 | model = cls() |
| 703 | patched_names = ckpt.get("patched_linear_names", []) |
| 704 | if patched_names: |
| 705 | for name in patched_names: |
| 706 | *parent_path, leaf = name.split(".") |
| 707 | parent = model |
| 708 | for p in parent_path: |
| 709 | parent = getattr(parent, p) |
| 710 | old = getattr(parent, leaf) |
| 711 | if not isinstance(old, BitLinear): |
| 712 | setattr(parent, leaf, BitLinear.from_linear(old)) |
| 713 | else: |
| 714 | apply_bitlinear_patch(model, include_gate_proj=include_gate_proj) |
| 715 | |
| 716 | has_mtp_in_ckpt = any(k.startswith("mtp.") for k in ckpt["model"]) |
| 717 | model.load_state_dict(ckpt["model"], strict=False) |
| 718 | model = model.to(dtype) |
| 719 | |
| 720 | if not has_mtp_in_ckpt: |
| 721 | print( |
| 722 | "⚠️ В этом .pt НЕТ mtp.* весов (чекпоинт сохранён до " |
| 723 | "появления MTPBlock). Голова сейчас со СЛУЧАЙНОЙ " |
| 724 | "инициализацией! Перед экспортом обязательно вызови " |
| 725 | "model.load_mtp_from_hf_dir('<папка оригинального HF-чекпоинта>')." |
| 726 | ) |
| 727 | |
| 728 | lam = ckpt.get("lambda", 0.0) |
| 729 | for m in model.modules(): |
| 730 | if isinstance(m, BitLinear): |
| 731 | m.set_lambda(lam) |
| 732 | |
| 733 | print(f"✅ Загружено. global_step={ckpt.get('global_step', 0)}, lambda={lam}") |
| 734 | return model |
| 735 | |
| 736 | |
| 737 | # ===================================================================== |
| 738 | # 7. Демонстрация (без скачивания весов -- только структура) |
| 739 | # ===================================================================== |
| 740 | |
| 741 | if __name__ == "__main__": |
| 742 | print(ARCH_SUMMARY) |
| 743 | print(KEPT_FULL_PRECISION) |
| 744 | |
| 745 | model = JiRackQwen38ForCausalLM() |
| 746 | n_params = sum(p.numel() for p in model.parameters()) |
| 747 | print(f"✅ Модель собрана (случайные веса). Параметров: {n_params / 1e9:.2f}B") |
| 748 | |
| 749 | patched = apply_bitlinear_patch(model) |
| 750 | print(f"✅ Заменено подмодулей на BitLinear: {len(patched)}") |
| 751 | print("Примеры первых 10:") |
| 752 | for n in patched[:10]: |
| 753 | print(f" {n}") |
| 754 | |
| 755 | ids = torch.randint(0, VOCAB_SIZE, (1, 8)) |
| 756 | with torch.no_grad(): |
| 757 | logits = model(ids) |
| 758 | print(f"✅ Пробный forward (случайные веса) -> logits shape: {tuple(logits.shape)}") |
| 759 | |