File size: 7,560 Bytes
fec7158 | 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 | """
Unified Sparse-AST / BWM Architecture & Universal Subfolder Loader
Domain: Blender 3D Mathematics & Blender Python Libraries (bpy, mathutils, bmesh, numpy, gpu)
"""
import os
import json
import torch
import torch.nn as nn
import torch.nn.functional as F
import safetensors.torch as st
class RMSNorm(nn.Module):
def __init__(self, d):
super().__init__()
self.w = nn.Parameter(torch.ones(d))
def forward(self, x):
return x * torch.rsqrt(x.float().square().mean(-1, keepdim=True) + 1e-6).to(x.dtype) * self.w
class SparseASTBlock(nn.Module):
def __init__(self, d=800, h=1600, al_hidden=24, n_heads=8):
super().__init__()
self.n1 = RMSNorm(d)
self.n2 = RMSNorm(d)
self.a = nn.MultiheadAttention(d, n_heads, batch_first=True)
self.up = nn.Linear(d, h)
self.r = nn.Linear(h, 3)
self.down = nn.Linear(2*h, d)
self.al = nn.Sequential(nn.Linear(3*h + 3, al_hidden), nn.SiLU(), nn.Linear(al_hidden, 1))
self.m = nn.Linear(d, d, bias=False)
def forward(self, x):
t = x.shape[1]
q = self.n1(x)
mask = torch.ones(t, t, device=x.device, dtype=torch.bool).triu(1)
x = torch.clamp(x + self.a(q, q, q, attn_mask=mask, need_weights=False)[0], -50, 50)
z = torch.clamp(self.up(self.n2(x)), -12, 12)
g = F.gumbel_softmax(self.r(z), tau=2.0, hard=True, dim=-1)
e = z.sign() * torch.expm1(z.abs().clamp(max=2.0))
u = (torch.stack((z, z.sign() * torch.log1p(z.abs()), e), -1) * g.unsqueeze(-2)).sum(-1).clamp(-8, 8)
p, n = F.relu(u), F.relu(-u)
d = torch.exp(-F.softplus(self.al(torch.cat((p, n, u, g), -1))).squeeze(-1)).clamp(1e-4, 0.9999)
y = torch.clamp(x + self.down(torch.cat((p, n), -1)), -50, 50)
state = torch.zeros_like(y[:, 0])
o = []
for j in range(t):
state = torch.clamp(state * d[:, j:j+1] + y[:, j] * (1 - d[:, j:j+1]), -50, 50)
o.append(torch.clamp(y[:, j] + self.m(state), -50, 50))
return torch.stack(o, 1)
class SparseAST(nn.Module):
def __init__(self, d=800, h=1600, layers=28, seq_len=32, vocab=512, al_hidden=24, n_heads=8):
super().__init__()
self.d = d
self.h = h
self.layers = layers
self.seq_len = seq_len
self.vocab = vocab
self.e = nn.Embedding(vocab, d)
self.p = nn.Embedding(seq_len, d)
self.b = nn.ModuleList([SparseASTBlock(d=d, h=h, al_hidden=al_hidden, n_heads=n_heads) for _ in range(layers)])
self.n = RMSNorm(d)
self.h_out = nn.Linear(d, vocab, bias=False)
self.h_out.weight = self.e.weight # Tied word embeddings
def forward(self, i):
if i.shape[1] > self.seq_len:
i = i[:, -self.seq_len:]
seq = i.shape[1]
pos = torch.arange(seq, device=i.device)[None]
x = self.e(i) + self.p(pos)
for b in self.b:
x = b(x)
return self.h_out(self.n(x))
@classmethod
def from_pretrained(cls, repo_or_dir, subfolder="100M-32", device="cpu", target_context=None):
target_dir = os.path.join(repo_or_dir, subfolder) if subfolder else repo_or_dir
config_path = os.path.join(target_dir, "config.json")
with open(config_path, "r", encoding="utf-8") as f:
cfg = json.load(f)
native_seq = cfg.get("native_seq_len", 32)
effective_seq = max(native_seq, target_context) if target_context else native_seq
model = cls(
d=cfg["d_model"],
h=cfg["d_hidden"],
layers=cfg["num_layers"],
seq_len=effective_seq,
vocab=cfg["vocab_size"],
al_hidden=cfg.get("al_hidden", 24),
n_heads=cfg.get("num_attention_heads", 8)
)
weights_path = os.path.join(target_dir, "model.safetensors")
state_dict = st.load_file(weights_path, device="cpu")
if "h.weight" in state_dict and "h_out.weight" not in state_dict:
state_dict["h_out.weight"] = state_dict["h.weight"]
elif "h_out.weight" not in state_dict and "e.weight" in state_dict:
state_dict["h_out.weight"] = state_dict["e.weight"]
if effective_seq > native_seq and "p.weight" in state_dict:
old_p = state_dict["p.weight"]
new_p = F.interpolate(
old_p.T.unsqueeze(0),
size=effective_seq,
mode="linear",
align_corners=True
).squeeze(0).T
state_dict["p.weight"] = new_p
model.load_state_dict(state_dict)
model.to(device)
model.eval()
return model
class TopKRouter(nn.Module):
def __init__(self, vocab_size=512, hidden_dim=64, num_experts=4):
super().__init__()
self.embed = nn.Embedding(vocab_size, hidden_dim)
self.norm = nn.LayerNorm(hidden_dim)
self.mlp = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, num_experts)
)
def forward(self, x):
emb = self.norm(self.embed(x))
return self.mlp(emb)
class TopKSparseASTEnsemble(nn.Module):
def __init__(self, router, experts=None, expert_names=None, k=2, tau=1.0, device="cpu"):
super().__init__()
self.router = router
self.experts = nn.ModuleList(experts if experts else [])
self.expert_names = expert_names or []
self.k = k
self.tau = tau
self.device = device
@classmethod
def from_pretrained(cls, repo_or_dir, subfolder="TopK-MoE", device="cpu", load_experts=True):
target_dir = os.path.join(repo_or_dir, subfolder) if subfolder else repo_or_dir
cfg_path = os.path.join(target_dir, "config.json")
with open(cfg_path, "r", encoding="utf-8") as f:
cfg = json.load(f)
router = TopKRouter(
vocab_size=cfg["vocab_size"],
hidden_dim=cfg["hidden_dim"],
num_experts=cfg["num_experts"]
)
st_path = os.path.join(target_dir, "model.safetensors")
router.load_state_dict(st.load_file(st_path, device=device))
router.to(device)
router.eval()
experts = []
names = cfg.get("expert_names", ["3M-32", "10M-32", "100M-32", "200M-32"])
if load_experts:
for exp_name in names:
exp_sub = exp_name if exp_name.endswith("-32") else f"{exp_name}-32"
exp_path = os.path.join(repo_or_dir, exp_sub)
if os.path.exists(exp_path):
print(f"Loading MoE expert [{exp_name}] from {exp_sub}...")
m = SparseAST.from_pretrained(repo_or_dir, subfolder=exp_sub, device=device)
experts.append(m)
return cls(router=router, experts=experts, expert_names=names, k=cfg.get("k", 2), tau=cfg.get("tau", 1.0), device=device)
def forward(self, x, k=None):
if k is None:
k = self.k
router_logits = self.router(x)
topk_scores, topk_indices = torch.topk(router_logits, k=k, dim=-1)
gate_weights = F.softmax(topk_scores / self.tau, dim=-1)
return router_logits, topk_indices, gate_weights
|