Ethanfel/ComfyUI-SelVA
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(bigvgan-trainer): clone vocoder parameters to strip inference tensor flag

Browse Source 
The vocoder is loaded inside ComfyUI's torch.inference_mode(), making all
its parameters inference tensors. Autograd cannot save inference tensors
for backward even with requires_grad=True. Clone all parameters inside
torch.inference_mode(False) before training to get normal tensors.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Ethanfel
2026-04-09 01:55:16 +02:00
parent f04d59fe63
commit c86306bde8
1 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nodes/selva_bigvgan_trainer.py
+6
View File
@@ -222,6 +222,12 @@ class SelvaBigvganTrainer:
try: try:
with torch.inference_mode(False): with torch.inference_mode(False):
with torch.enable_grad(): with torch.enable_grad():
# Vocoder parameters were loaded inside ComfyUI's inference_mode()
# and are inference tensors. Autograd cannot save them for backward.
# Clone inside inference_mode(False) to get normal tensors.
for param in vocoder.parameters():
param.data = param.data.clone()
vocoder.train() vocoder.train()
for step in range(steps): for step in range(steps):