diff --git a/vllm/config/parallel.py b/vllm/config/parallel.py index bb1d2095de..bd2532b1cc 100644 --- a/vllm/config/parallel.py +++ b/vllm/config/parallel.py @@ -393,8 +393,15 @@ class ParallelConfig: in (rank i+1, block j) only after (rank i, block j) is fully occupied. Block_size should be greater than or equal to cp_kv_cache_interleave_size. Block_size should be divisible by cp_kv_cache_interleave_size. + + When --cp-kv-cache-interleave-size is omitted (None), the interleave size + is resolved automatically based on NIXL transfer requirements. + Explicit settings take priority. """ + _allow_auto_resolve_cp_interleave_size: bool = True + """Whether NIXL may select the interleave size automatically.""" + data_parallel_index: int = Field(init=False) """Equal to the data parallel rank but not used for torch process groups and not overridden for dense models.""" diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py index f5bee6fa2c..893acb00cd 100644 --- a/vllm/config/vllm.py +++ b/vllm/config/vllm.py @@ -1237,11 +1237,6 @@ class VllmConfig: "PD with decode_context_parallel_size > 1 is only " "supported for MLA models." ) - assert not (self.model_config.is_hybrid and dcp_size > 1), ( - "PD with decode_context_parallel_size > 1 is not " - "supported for hybrid Mamba/SSM models." - ) - if self.lora_config is not None: self.lora_config.verify_with_model_config(self.model_config) @@ -2883,6 +2878,8 @@ class VllmConfig: "NixlConnector" ): return + if not self.parallel_config._allow_auto_resolve_cp_interleave_size: + return # Get the kernel block_size, but don't use resolve_kv_cache_block_size to avoid # scaling by dcp_size (we need the local block_size here). diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py index 5cb76932d4..ceef10e9fd 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py @@ -530,9 +530,6 @@ class NixlBaseConnectorWorker: # DCP support is scoped to MLA, with dcp_size in (1, tp_size): either fully # replicated or fully sharded. A DCP rank is always derivable this way. self.dcp_rank = self.tp_rank % self.dcp_size - if self._has_mamba and self.dcp_size > 1: - # Prefix-cache-aware DCP slicing isn't implemented for the Mamba group. - raise ValueError("DCP is not supported for hybrid MLA+Mamba models.") self.num_blocks = kv_cache_config.num_blocks self.enable_permute_local_kv = False @@ -2177,6 +2174,11 @@ class NixlBaseConnectorWorker: f"DCP sizes must divide one another: local={self.dcp_size}, " f"remote={remote_dcp_size} (engine {remote_engine_id})." ) + if self._has_mamba and self.dcp_size != remote_dcp_size: + raise RuntimeError( + "Hybrid MLA+Mamba NIXL transfers require matching DCP sizes, " + f"got local={self.dcp_size}, remote={remote_dcp_size}." + ) tp_ratio = self.transfer_topo.tp_ratio(remote_tp_size) block_size_ratio = self.transfer_topo.block_size_ratio( diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py index 758f5c9fab..d39a9fe287 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl/pull_worker.py @@ -17,6 +17,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.nixl.metadata import ( ) from vllm.distributed.kv_transfer.kv_connector.v1.nixl.tp_mapping import ( ReadSpec, + _is_attention_spec, ) from vllm.logger import init_logger @@ -213,7 +214,9 @@ class NixlPullConnectorWorker(NixlBaseConnectorWorker): local_ids = group_ids(meta.local_block_ids, rank) remote_ids = group_ids(remote_logical_block_ids, rank) for g in range(num_groups): - if not local_ids[g]: + if not local_ids[g] or not _is_attention_spec( + self._group_spec_types[g] + ): continue local_ids[g], remote_ids[g] = self._apply_dcp_prefix_caching( local_ids[g], diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py index bf5cc62d36..db398438d0 100644 --- a/vllm/engine/arg_utils.py +++ b/vllm/engine/arg_utils.py @@ -487,7 +487,7 @@ class EngineArgs: dcp_comm_backend: DCPCommBackend | None = ParallelConfig.dcp_comm_backend dcp_q_replicate: bool | None = ParallelConfig.dcp_q_replicate dcp_kv_cache_interleave_size: int = ParallelConfig.dcp_kv_cache_interleave_size - cp_kv_cache_interleave_size: int = ParallelConfig.cp_kv_cache_interleave_size + cp_kv_cache_interleave_size: int | None = None data_parallel_size: int = ParallelConfig.data_parallel_size data_parallel_rank: int | None = None data_parallel_start_rank: int | None = None @@ -1104,7 +1104,10 @@ class EngineArgs: ) parallel_group.add_argument( "--cp-kv-cache-interleave-size", - **parallel_kwargs["cp_kv_cache_interleave_size"], + **{ + **parallel_kwargs["cp_kv_cache_interleave_size"], + "default": None, + }, ) parallel_group.add_argument( "--prefill-context-parallel-size", @@ -2363,7 +2366,14 @@ class EngineArgs: dcp_comm_backend=self.dcp_comm_backend, dcp_q_replicate=self.dcp_q_replicate, dcp_kv_cache_interleave_size=self.dcp_kv_cache_interleave_size, - cp_kv_cache_interleave_size=self.cp_kv_cache_interleave_size, + cp_kv_cache_interleave_size=( + self.cp_kv_cache_interleave_size + if self.cp_kv_cache_interleave_size is not None + else ParallelConfig.cp_kv_cache_interleave_size + ), + _allow_auto_resolve_cp_interleave_size=( + self.cp_kv_cache_interleave_size is None + ), _api_process_count=self._api_process_count, _api_process_rank=self._api_process_rank, assigned_physical_gpu_ids=self._resolve_device_ids(),