[Bugfix][Model] Support LoRA on Qwen3 Output Embedding (#29816)

Signed-off-by: kurt <kurt@thinkingmachines.ai>
2026-02-06 04:25:31 -08:00
parent ac32e66cf9
commit 2991dd3d22
6 changed files with 132 additions and 13 deletions
--- a/vllm/model_executor/models/qwen3_moe.py
+++ b/vllm/model_executor/models/qwen3_moe.py
@@ -689,6 +689,11 @@ class Qwen3MoeForCausalLM(
        ]
    }

+    embedding_modules = {
+        "embed_tokens": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+
    fall_back_to_pt_during_load = False

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):