diff --git a/tests/models/test_registry.py b/tests/models/test_registry.py index 14d85fc09d..b3ba172822 100644 --- a/tests/models/test_registry.py +++ b/tests/models/test_registry.py @@ -129,8 +129,6 @@ def test_registry_model_property(model_arch, is_mm, init_cuda, score_type): # in V1. # ("MLPSpeculatorPreTrainedModel", False, False), ("DeepseekV2ForCausalLM", True, False), - ("KimiLinearForCausalLM", True, False), - ("KimiK3ForConditionalGeneration", True, False), ("Qwen2VLForConditionalGeneration", True, True), ], ) diff --git a/vllm/model_executor/layers/attention/mla_attention.py b/vllm/model_executor/layers/attention/mla_attention.py index 25eef7462e..f9b86bd810 100644 --- a/vllm/model_executor/layers/attention/mla_attention.py +++ b/vllm/model_executor/layers/attention/mla_attention.py @@ -2543,16 +2543,6 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]): if num_decodes > 0: dcp_tot_seq_lens_device = None if self.dcp_world_size > 1: - assert seq_lens is not None, ( - "MLA DCP decode requires seq_lens on CommonAttentionMetadata" - ) - if dcp_local_seq_lens is None: - dcp_local_seq_lens = get_dcp_local_seq_lens( - seq_lens, - dcp_size=self.dcp_world_size, - dcp_rank=get_dcp_group().rank_in_group, - cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size, - ) dcp_tot_seq_lens_device = seq_lens[:num_decodes] seq_lens = dcp_local_seq_lens