From 1f232f43efac7056d7c30d7914f32cf3bce971ff Mon Sep 17 00:00:00 2001 From: CalamitousFelicitousness Date: Sat, 8 Aug 2026 03:56:56 +0100 Subject: [PATCH] fix(video): keep the conditioner vision tower unquantized The modular text encoder quant config now excludes the vision tower. The default skip pool covers diffusion module names, so qwen-vl vision blocks quantized alongside the language layers; vision blocks have no validated quantization precedent and run only for keyframe workflows. --- modules/video_models/video_modular.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/video_models/video_modular.py b/modules/video_models/video_modular.py index ec39c79b9..3424695aa 100644 --- a/modules/video_models/video_modular.py +++ b/modules/video_models/video_modular.py @@ -45,7 +45,7 @@ def load_modular_pipe(repo_cls, repo: str, workflow: str | None = None, revision if 'quantization_config' in quant_args: quant_config['transformer'] = quant_args['quantization_config'] quant_config['transformer_ref'] = quant_args['quantization_config'] - te_args = model_quant.create_config(module='TE') + te_args = model_quant.create_config(module='TE', modules_to_not_convert=['.model.visual']) # the conditioner's vision tower stays unquantized: quantized vision blocks have no validated precedent and only run for keyframe workflows if 'quantization_config' in te_args: quant_config['text_encoder'] = te_args['quantization_config'] if quant_config: