fam/nanochat-omni
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

use nn.init and initialize smear gate's weight as well

Browse Source
This commit is contained in:
Sofie Van Landeghem
2026-04-13 14:03:18 +02:00
committed by GitHub
parent 94b73ad29a
commit 9822cc7424
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nanochat/gpt.py
+3 -2
View File
@@ -237,8 +237,9 @@ class GPT(nn.Module):
# Decaying x0 init: earlier layers get more input embedding blending # Decaying x0 init: earlier layers get more input embedding blending
for i in range(n_layer): for i in range(n_layer):
self.x0_lambdas.data[i] = 0.20 - (0.15 * i / max(n_layer - 1, 1)) self.x0_lambdas.data[i] = 0.20 - (0.15 * i / max(n_layer - 1, 1))
self.smear_lambda.fill_(0.0) torch.nn.init.zeros_(self.smear_lambda)
self.backout_lambda.fill_(0.2) torch.nn.init.constant_(self.backout_lambda, 0.2)
torch.nn.init.uniform_(self.smear_gate.weight, 0.0, 0.02)
# Value embeddings (init like c_v: uniform with same std) # Value embeddings (init like c_v: uniform with same std)
for ve in self.value_embeds.values(): for ve in self.value_embeds.values():