File size: 2,341 Bytes
3b2d368 | 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 | import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from lmr.models.lm_base import LMBase
from .components import ICLBlock, PhiBlock
class FST(LMBase):
def __init__(self, config, train_mode=False):
super().__init__()
# print(config)
self.config = config
self.train_mode = train_mode
self.embedding_f = nn.Embedding(config.vocab_size, config.embed_dim_f)
if not self.config.use_shared_embedding:
self.embedding_phi = nn.Embedding(config.vocab_size, config.embed_dim_phi)
self.phi_s = nn.Parameter(torch.randn(1, 1, config.embed_dim_phi))
self.phi_blocks = nn.ModuleList([PhiBlock(config) for _ in range(config.n_layers // 2)])
self.icl_blocks = nn.ModuleList([ICLBlock(config) for _ in range(config.n_layers // 2)])
self.ln_out = nn.LayerNorm(config.embed_dim_f)
self.lm_head = nn.Linear(config.embed_dim_f, config.vocab_size, bias=False)
self.apply(self.init_weights)
self.lm_head.weight = self.embedding_f.weight
def forward(self, input_ids):
batch_size, seq_len = input_ids.shape
device = input_ids.device
context_embeddings = self.embedding_f(input_ids)
if self.config.use_shared_embedding:
phi = context_embeddings
if self.config.embed_dim_f != self.config.embed_dim_phi:
phi = phi[:, :, :self.config.embed_dim_phi]
else:
phi = self.embedding_phi(input_ids)
context_embeddings = torch.cat([
context_embeddings,
torch.zeros(batch_size, 1, self.config.embed_dim_f, device=device)
], dim=1)
phi_s = self.phi_s.expand(batch_size, -1, -1)
phi = torch.cat([phi_s, phi], dim=1)
f = torch.zeros(batch_size, seq_len + 1, self.config.embed_dim_f, device=device)
for phi_block, icl_block in zip(self.phi_blocks, self.icl_blocks):
phi = phi_block(phi)
_, _, f = icl_block(phi, context_embeddings, f)
f_NP1 = f[:, 1:, :]
f_NP1 = self.ln_out(f_NP1)
logits = self.lm_head(f_NP1)
return logits |