From ff5d1a3dc01e7d26c9ccc18fc2c734efbe928f5c Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Tue, 29 Apr 2025 10:09:44 -0700 Subject: [PATCH] duplicate input embeddings --- model/models/qwen25vl/model_text.go | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/model/models/qwen25vl/model_text.go b/model/models/qwen25vl/model_text.go index 18ba79e84..25c140ed4 100644 --- a/model/models/qwen25vl/model_text.go +++ b/model/models/qwen25vl/model_text.go @@ -146,7 +146,7 @@ func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Ten func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor, batch input.Batch, cache kvcache.Cache) (ml.Tensor, error) { // Initial token embedding - hiddenState := m.TokenEmbedding.Forward(ctx, inputs) + hiddenState := m.TokenEmbedding.Forward(ctx, inputs).Duplicate(ctx) for _, image := range batch.Multimodal { visionOutputs := image.Multimodal.(ml.Tensor)