From 6e3bbef56895951a8831562bac9440b7598581e7 Mon Sep 17 00:00:00 2001 From: Opher Lieber Date: Fri, 17 Apr 2026 02:35:13 -0700 Subject: [PATCH] expose num_embeddings in VocabParallelEmbeddingWithLoRA (#22547) --- python/sglang/srt/lora/layers.py | 1 + 1 file changed, 1 insertion(+) diff --git a/python/sglang/srt/lora/layers.py b/python/sglang/srt/lora/layers.py index 7b184e71a..8b596c7ef 100644 --- a/python/sglang/srt/lora/layers.py +++ b/python/sglang/srt/lora/layers.py @@ -70,6 +70,7 @@ class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA): self.weight = base_layer.weight self.embed_dim = base_layer.embedding_dim self.vocab_size = base_layer.org_vocab_size + self.num_embeddings = base_layer.num_embeddings # Embedding LoRA with TP > 1 keeps weights fully replicated # (unsharded) on every rank. This works correctly because the