Checklist

• I searched related issues but found no solution.
• The bug persists in the latest version.
• Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
• If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Describe the bug

When we override enable_fused_set_kv_buffer to False and run test/registered/dllm/test_llada2_mini.py, this test will fail.

The root cause is that, the flashinfer attention backend skip the save_kv_cache, which is incorrect. After removing the following lines, the test will pass again.

Reproduction

1. Modify this function. Always return False
   

   sglang/python/sglang/srt/models/utils.py

   Lines 107 to 114 in d73f06f

   	def enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
   	"""Enable fused set_kv_buffer only on CUDA with bfloat16 KV cache."""
   	return (
   	_is_cuda
   	and hasattr(forward_batch.token_to_kv_pool, "dtype")
   	and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
   	and not isinstance(forward_batch.token_to_kv_pool, SWAKVPool)
   	)

2. Run pytest -xss test/registered/dllm/test_llada2_mini.py

Environment

H200