From 1e2f2bfbca4e64e852713550ab547ab1665c3f2c Mon Sep 17 00:00:00 2001 From: k1kong Date: Wed, 10 Jun 2026 02:44:08 +0000 Subject: [PATCH] fix fir transformer 5.10.2 --- .../pipelines/basic/glm_image/stages/before_denoising.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/fastvideo/pipelines/basic/glm_image/stages/before_denoising.py b/fastvideo/pipelines/basic/glm_image/stages/before_denoising.py index cce0b9f7d..f3a0331d2 100644 --- a/fastvideo/pipelines/basic/glm_image/stages/before_denoising.py +++ b/fastvideo/pipelines/basic/glm_image/stages/before_denoising.py @@ -243,7 +243,9 @@ class GlmImageBeforeDenoisingStage(PipelineStage): image_grid_thw = inputs["image_grid_thw"] num_condition_images = image_grid_thw.shape[0] - 1 source_grids = image_grid_thw[:num_condition_images] - embed = torch.cat(self.vision_language_encoder.get_image_features(inputs["pixel_values"], source_grids), dim=0) + image_features = self.vision_language_encoder.get_image_features(inputs["pixel_values"], source_grids) + image_feature_parts = getattr(image_features, "pooler_output", image_features) + embed = torch.cat(image_feature_parts, dim=0) src_ids_d32 = self.vision_language_encoder.get_image_tokens(embed, source_grids) split_sizes = source_grids.prod(dim=-1).tolist() upsampled = [