File size: 8,748 Bytes
a6e2346 | 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 | import torch
import torch.nn as nn
import torch.nn.functional as F
from gguf import GGUFWriter
class ModelConfig:
def __init__(self):
# Core parameters
self.vocab_size = 32000
self.hidden_size = 768
self.num_hidden_layers = 4
self.num_attention_heads = 8
self.intermediate_size = 3072
# Expert parameters
self.num_experts = 4
# Efficiency parameters
self.chunk_size = 256
self.compression_ratio = 4
# Reasoning parameters
self.max_graph_nodes = 512
self.node_dim = self.hidden_size // 4 # 192
# Regularization
self.hidden_dropout_prob = 0.1
self.initializer_range = 0.02
class CoATGraphManager(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
# Node initialization
self.base_nodes = nn.Parameter(torch.randn(config.max_graph_nodes, config.node_dim))
self.projection = nn.Linear(config.hidden_size, config.node_dim)
# Update mechanism
self.update_gate = nn.Sequential(
nn.Linear(config.node_dim * 2, config.hidden_size),
nn.GELU(),
nn.Linear(config.hidden_size, 1),
nn.Sigmoid()
)
def forward(self, hidden_states, current_nodes):
batch_size = hidden_states.size(0)
# Clone to prevent in-place errors
current_nodes = current_nodes.clone()
# Aggregate sequence information
seq_aggregated = hidden_states.mean(dim=1) # [batch, hidden_size]
# Project to node space
projected = self.projection(seq_aggregated) # [batch, node_dim]
# Calculate similarity scores
similarity = torch.matmul(projected.unsqueeze(1), current_nodes.transpose(1, 2)) # [batch, 1, max_nodes]
# Get top-2 nodes
_, topk_indices = torch.topk(similarity.squeeze(1), k=2, dim=-1) # [batch, 2]
# Gather relevant nodes
selected_nodes = torch.gather(
current_nodes,
1,
topk_indices.unsqueeze(-1).expand(-1, -1, self.config.node_dim)
) # [batch, 2, node_dim]
# Calculate updates
combined = torch.cat([
selected_nodes,
self.base_nodes[topk_indices]
], dim=-1)
update_weights = self.update_gate(combined)
updated_nodes = selected_nodes * update_weights + self.base_nodes[topk_indices] * (1 - update_weights)
# Safe scatter update
current_nodes.scatter_(
1,
topk_indices.unsqueeze(-1).expand(-1, -1, self.config.node_dim),
updated_nodes
)
return current_nodes
class ChunkKVAttention(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.head_dim = config.hidden_size // config.num_attention_heads
# Projections
self.q_proj = nn.Linear(config.hidden_size, config.hidden_size)
self.k_proj = nn.Linear(config.hidden_size, config.hidden_size)
self.v_proj = nn.Linear(config.hidden_size, config.hidden_size)
# Compression
self.k_compress = nn.Linear(config.chunk_size, config.chunk_size//config.compression_ratio)
self.v_compress = nn.Linear(config.chunk_size, config.chunk_size//config.compression_ratio)
def forward(self, hidden_states):
batch_size, seq_len, _ = hidden_states.size()
# Process queries
q = self.q_proj(hidden_states)
# Process keys/values in chunks
k = self._process_chunk(self.k_proj, self.k_compress, hidden_states)
v = self._process_chunk(self.v_proj, self.v_compress, hidden_states)
# Reshape for attention
q = q.view(batch_size, -1, self.config.num_attention_heads, self.head_dim).transpose(1, 2)
k = k.view(batch_size, -1, self.config.num_attention_heads, self.head_dim).transpose(1, 2)
v = v.view(batch_size, -1, self.config.num_attention_heads, self.head_dim).transpose(1, 2)
# Attention calculation
attn = torch.matmul(q, k.transpose(-2, -1)) / torch.sqrt(torch.tensor(self.head_dim))
attn = F.softmax(attn, dim=-1)
output = torch.matmul(attn, v)
return output.transpose(1, 2).flatten(2)
def _process_chunk(self, proj, compress, x):
chunks = []
for i in range(0, x.size(1), self.config.chunk_size):
chunk = proj(x[:, i:i+self.config.chunk_size])
compressed = compress(chunk.transpose(1, 2)).transpose(1, 2)
chunks.append(compressed)
return torch.cat(chunks, dim=1)
class SelfMoA(nn.Module):
def __init__(self, config):
super().__init__()
self.experts = nn.ModuleList([
nn.Sequential(
nn.Linear(config.hidden_size, config.intermediate_size),
nn.GELU(),
nn.Linear(config.intermediate_size, config.hidden_size)
) for _ in range(config.num_experts)
])
self.gate = nn.Linear(config.hidden_size, config.num_experts)
def forward(self, x):
gate = F.gumbel_softmax(self.gate(x), hard=True, dim=-1)
return sum(expert(x) * gate[..., i].unsqueeze(-1) for i, expert in enumerate(self.experts))
class DeepSeekLiteBlock(nn.Module):
def __init__(self, config):
super().__init__()
self.attention = ChunkKVAttention(config)
self.moa = SelfMoA(config)
self.coat = CoATGraphManager(config)
self.norm = nn.LayerNorm(config.hidden_size)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, x, nodes):
# Attention path
attn_out = self.attention(self.norm(x))
x = x + self.dropout(attn_out)
# Update graph nodes
updated_nodes = self.coat(x, nodes)
# MOA path
moa_out = self.moa(self.norm(x))
x = x + self.dropout(moa_out)
return x, updated_nodes
class DeepSeekLite(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.embedding = nn.Embedding(config.vocab_size, config.hidden_size)
self.layers = nn.ModuleList([DeepSeekLiteBlock(config) for _ in range(config.num_hidden_layers)])
self.final_norm = nn.LayerNorm(config.hidden_size)
# Initialize graph nodes with cloning
self.graph_nodes = nn.ParameterList([
nn.Parameter(torch.randn(config.max_graph_nodes, config.node_dim).clone().detach().requires_grad_(True))
for _ in range(config.num_hidden_layers)
])
def forward(self, input_ids):
x = self.embedding(input_ids)
batch_size = input_ids.size(0)
for layer_idx, layer in enumerate(self.layers):
# Clone and expand nodes for each layer
nodes = self.graph_nodes[layer_idx].unsqueeze(0).expand(batch_size, -1, -1).clone()
x, _ = layer(x, nodes)
return self.final_norm(x)
def save_gguf(model, filename):
writer = GGUFWriter(filename, "deepseek-lite")
# Add model configuration
writer.add_uint32("vocab_size", model.config.vocab_size)
writer.add_uint32("hidden_size", model.config.hidden_size)
writer.add_uint32("num_hidden_layers", model.config.num_hidden_layers)
writer.add_uint32("num_attention_heads", model.config.num_attention_heads)
writer.add_uint32("num_experts", model.config.num_experts)
writer.add_uint32("max_graph_nodes", model.config.max_graph_nodes)
# Add all parameters
for name, param in model.named_parameters():
writer.add_tensor(name, param.detach().cpu().numpy())
writer.write_header_to_file()
writer.write_kv_data_to_file()
writer.write_tensors_to_file()
writer.close()
if __name__ == "__main__":
config = ModelConfig()
model = DeepSeekLite(config)
# Test forward pass
inputs = torch.randint(0, config.vocab_size, (2, 1024))
with torch.no_grad():
outputs = model(inputs)
print(f"Successful execution! Output shape: {outputs.shape}")
print(f"Parameter count: {sum(p.numel() for p in model.parameters())/1e6:.1f}M")
# Save model
save_gguf(model, "deepseek-lite.gguf")
print("Model saved in GGUF format") |