File size: 7,507 Bytes
431ffa3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 | # -*- coding: utf-8 -*-
from __future__ import annotations
import torch
import torch.nn as nn
from Model.layers.mla import MLA
from Model.layers.mamba3_layer import Mamba3Layer
from Model.layers.mixture_lora_ffn import MixtureLoRAFFN
from Model.layers.rmsnorm import RMSNorm
from Model.layers.swiglu import SwiGLU
def _build_ffn(cfg):
"""Recurrent-block FFN: step-aware Mixture-of-LoRAs when ``use_mol``, else the
plain shared SwiGLU. Prelude/coda (``StandardBlock``) always use plain SwiGLU
-- they are non-recurrent, so step-aware expert routing has no step to
condition on."""
if getattr(cfg, "use_mol", False):
return MixtureLoRAFFN(
cfg.d_model,
cfg.ffn_hidden,
cfg.mol_experts,
cfg.mol_rank,
cfg.recurrence_step_table,
step_aware=cfg.mol_step_aware,
router_temp=cfg.mol_router_temp,
top_k=cfg.mol_top_k,
init_std=cfg.init_std,
rmsnorm_eps=cfg.rmsnorm_eps,
)
return SwiGLU(cfg.d_model, cfg.ffn_hidden)
def _apply_ffn(ffn, x, step):
"""Forward a ``_build_ffn`` FFN, threading the recurrent step index only into
a MixtureLoRAFFN (plain SwiGLU has no ``step`` kwarg and ignores it)."""
if isinstance(ffn, MixtureLoRAFFN):
return ffn(x, step=step)
return ffn(x)
class StandardBlock(nn.Module):
def __init__(self, cfg, layer_idx: int | None = None):
super().__init__()
self.layer_idx = layer_idx
self.attn_norm = RMSNorm(cfg.d_model, eps=cfg.rmsnorm_eps)
self.attn = MLA(cfg)
self.ffn_norm = RMSNorm(cfg.d_model, eps=cfg.rmsnorm_eps)
self.ffn = SwiGLU(cfg.d_model, cfg.ffn_hidden)
def forward(
self,
x: torch.Tensor,
word_pos: torch.Tensor | None = None,
morph_depth: torch.Tensor | None = None,
attn_mask: torch.Tensor | None = None,
causal: bool = True,
cache=None,
pos_offset: int = 0,
) -> torch.Tensor:
x = x + self.attn(
self.attn_norm(x),
word_pos=word_pos,
morph_depth=morph_depth,
attn_mask=attn_mask,
causal=causal,
cache=cache,
pos_offset=pos_offset,
)
x = x + self.ffn(self.ffn_norm(x))
return x
class AttnSubLayer(nn.Module):
def __init__(self, cfg, layer_idx: int | None = None):
super().__init__()
self.layer_idx = layer_idx
self.attn_norm = RMSNorm(cfg.d_model, eps=cfg.rmsnorm_eps)
self.attn = MLA(cfg)
self.ffn_norm = RMSNorm(cfg.d_model, eps=cfg.rmsnorm_eps)
self.ffn = _build_ffn(cfg)
def forward(
self,
x: torch.Tensor,
word_pos: torch.Tensor | None = None,
morph_depth: torch.Tensor | None = None,
attn_mask: torch.Tensor | None = None,
causal: bool = True,
cache=None,
pos_offset: int = 0,
step: int | None = None,
) -> torch.Tensor:
x = x + self.attn(
self.attn_norm(x),
word_pos=word_pos,
morph_depth=morph_depth,
attn_mask=attn_mask,
causal=causal,
cache=cache,
pos_offset=pos_offset,
)
x = x + _apply_ffn(self.ffn, self.ffn_norm(x), step)
return x
class MambaSubLayer(nn.Module):
def __init__(self, cfg, layer_idx: int | None = None):
super().__init__()
self.layer_idx = layer_idx
self.mamba = Mamba3Layer(cfg, layer_idx=layer_idx)
self.ffn_norm = RMSNorm(cfg.d_model, eps=cfg.rmsnorm_eps)
self.ffn = _build_ffn(cfg)
def forward(
self,
x: torch.Tensor,
attn_mask: torch.Tensor | None = None,
cache=None,
step: int | None = None,
**kwargs,
) -> torch.Tensor:
x = self.mamba(x, attn_mask=attn_mask, cache=cache)
x = x + _apply_ffn(self.ffn, self.ffn_norm(x), step)
if attn_mask is not None:
x = x * attn_mask.to(device=x.device, dtype=x.dtype).unsqueeze(-1)
return x
class RecurrentBlock(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
self.layer_types = self._interleave(
cfg.mamba_per_block,
cfg.attn_per_block,
)
self.layers = nn.ModuleList(
self._make_layer(cfg, kind, idx)
for idx, kind in enumerate(self.layer_types)
)
@staticmethod
def _interleave(n_mamba: int, n_attn: int) -> list[str]:
if n_mamba < 0 or n_attn < 0:
raise ValueError("layer counts must be non-negative")
total = n_mamba + n_attn
if total <= 0:
raise ValueError("recurrent block cannot be empty")
if n_attn == 0:
return ["mamba"] * n_mamba
if n_mamba == 0:
return ["attn"] * n_attn
result: list[str] = []
m_used = 0
a_used = 0
for pos in range(total):
target_attn = round((pos + 1) * n_attn / total)
if a_used < target_attn and a_used < n_attn:
result.append("attn")
a_used += 1
elif m_used < n_mamba:
result.append("mamba")
m_used += 1
else:
result.append("attn")
a_used += 1
return result
@staticmethod
def _make_layer(cfg, kind: str, layer_idx: int) -> nn.Module:
if kind == "mamba":
return MambaSubLayer(cfg, layer_idx=layer_idx)
if kind == "attn":
return AttnSubLayer(cfg, layer_idx=layer_idx)
raise ValueError(f"unknown layer type: {kind}")
def forward(
self,
x: torch.Tensor,
word_pos: torch.Tensor | None = None,
morph_depth: torch.Tensor | None = None,
attn_mask: torch.Tensor | None = None,
causal: bool = True,
step: int | None = None,
) -> torch.Tensor:
for layer in self.layers:
x = layer(
x,
word_pos=word_pos,
morph_depth=morph_depth,
attn_mask=attn_mask,
causal=causal,
step=step,
)
return x
def _check() -> None:
from Model.config import tiny_config
torch.manual_seed(0)
cfg = tiny_config()
bsz, seq_len = 2, 16
x = torch.randn(bsz, seq_len, cfg.d_model)
word_pos = torch.arange(seq_len).unsqueeze(0).expand(bsz, seq_len)
morph_depth = torch.zeros(bsz, seq_len, dtype=torch.long)
block = StandardBlock(cfg)
y = block(
x,
word_pos=word_pos,
morph_depth=morph_depth,
)
print("StandardBlock")
print(f" shape: {tuple(x.shape)} -> {tuple(y.shape)}")
recurrent = RecurrentBlock(cfg)
y2 = recurrent(
x,
word_pos=word_pos,
morph_depth=morph_depth,
)
print("RecurrentBlock")
print(f" layer_types: {recurrent.layer_types}")
print(f" shape: {tuple(x.shape)} -> {tuple(y2.shape)}")
print(f" params: {sum(p.numel() for p in recurrent.parameters()):,}")
x2 = torch.randn(bsz, seq_len, cfg.d_model, requires_grad=True)
loss = recurrent(
x2,
word_pos=word_pos,
morph_depth=morph_depth,
).sum()
loss.backward()
print(f" grad_norm: {x2.grad.norm().item():.6f}")
if __name__ == "__main__":
_check()
|