vlm: enable radix cache for qwen-vl models (#5349)

Co-authored-by: Xinyuan Tong <justinning0323@outlook.com>
2025-04-24 12:35:05 +09:00
parent 7d0edf3cae
commit c998d04b46
26 changed files with 429 additions and 331 deletions
--- a/test/srt/test_vlm_accuracy.py
+++ b/test/srt/test_vlm_accuracy.py
@@ -229,9 +229,9 @@ class TestMiniCPMVLogits(VisionLLMLogitsBase):
                input_ids=input_ids,
                input_embedding=model.get_input_embeddings(),
                image_data_embedding_func=model.get_image_feature,
-                placeholder_token_ids=[
-                    self.processor.tokenizer.unk_token_id,
-                ],
+                placeholder_tokens={
+                    Modality.IMAGE: self.processor.tokenizer.unk_token_id,
+                },
            )

        self.compare_outputs(sglang_output, hf_output)