diff --git a/tests/v1/kv_offload/test_file_mapper.py b/tests/v1/kv_offload/test_file_mapper.py index 0e462f8de2..6f6e0d6619 100644 --- a/tests/v1/kv_offload/test_file_mapper.py +++ b/tests/v1/kv_offload/test_file_mapper.py @@ -64,6 +64,7 @@ def make_mapper_from_offloading_spec(**kwargs) -> FileMapper: "dcp_size", 1 ) mock_vllm_config.parallel_config.rank = kwargs.get("rank", 0) + mock_vllm_config.use_v2_model_runner = kwargs.get("use_v2_model_runner", False) mock_kv_cache_config = MagicMock() mock_kv_cache_config.kv_cache_groups = kwargs.get("kv_cache_groups", []) @@ -210,3 +211,16 @@ def test_parallel_agnostic_excludes_mla(): ) assert fm.fields["tp_size"] == 2 assert fm.rank == 1 + + +def test_parallel_agnostic_disabled_on_v2_model_runner(): + # V2's KV layout is not known to be parallelism-invariant: don't collapse. + fm = make_mapper_from_offloading_spec( + tp_size=2, + rank=1, + kv_cache_groups=[_full_attention_group()], + use_v2_model_runner=True, + parallel_agnostic=True, + ) + assert fm.fields["tp_size"] == 2 + assert fm.rank == 1 diff --git a/tests/v1/kv_offload/tiering/test_obj_tier.py b/tests/v1/kv_offload/tiering/test_obj_tier.py index bac5729eaf..aae3c60c53 100644 --- a/tests/v1/kv_offload/tiering/test_obj_tier.py +++ b/tests/v1/kv_offload/tiering/test_obj_tier.py @@ -37,6 +37,7 @@ def _make_vllm_config(): decode_context_parallel_size=1, rank=0, ), + use_v2_model_runner=False, ) diff --git a/vllm/v1/kv_offload/file_mapper.py b/vllm/v1/kv_offload/file_mapper.py index c19f07ff51..d8fadb0998 100644 --- a/vllm/v1/kv_offload/file_mapper.py +++ b/vllm/v1/kv_offload/file_mapper.py @@ -84,10 +84,13 @@ class FileMapper: ] # Only a single full-attention group is parallelism-invariant. MLA is # excluded: its latent KV is replicated per rank, never head-sharded. + # The V2 model runner is excluded: its KV layout is not known to be + # parallelism-invariant. groups = kv_cache_config.kv_cache_groups spec = groups[0].kv_cache_spec if len(groups) == 1 else None parallel_agnostic = ( parallel_agnostic + and not vllm_config.use_v2_model_runner and isinstance(spec, FullAttentionSpec) and not isinstance(spec, MLAAttentionSpec) )