From 0854bd263844cb2a58c2b06ebce7cad96306b8e2 Mon Sep 17 00:00:00 2001 From: Ethanfel Date: Fri, 10 Apr 2026 00:25:04 +0200 Subject: [PATCH] fix: cast discriminators to model dtype to match vocoder output Discriminators are constructed as float32 but receive bfloat16 tensors from the vocoder. Cast to model dtype on load to prevent conv dtype mismatch in feature matching loss. Co-Authored-By: Claude Opus 4.6 --- nodes/selva_bigvgan_trainer.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/nodes/selva_bigvgan_trainer.py b/nodes/selva_bigvgan_trainer.py index 25e096a..4130a17 100644 --- a/nodes/selva_bigvgan_trainer.py +++ b/nodes/selva_bigvgan_trainer.py @@ -954,8 +954,8 @@ def _do_train(vocoder, mel_converter, clips, f"MPD loaded={mpd_loaded}, MRD loaded={mrd_loaded}. " f"Available keys: {list(ckpt_d.keys())}" ) - mpd.to(device).eval() - mrd.to(device).eval() + mpd.to(device, dtype).eval() + mrd.to(device, dtype).eval() for p in mpd.parameters(): p.requires_grad_(False) for p in mrd.parameters():