diff --git a/vllm/v1/worker/gpu/model_runner.py b/vllm/v1/worker/gpu/model_runner.py index ea476d0c13..6be4ab146c 100644 --- a/vllm/v1/worker/gpu/model_runner.py +++ b/vllm/v1/worker/gpu/model_runner.py @@ -510,6 +510,7 @@ class GPUModelRunner(LoRAModelRunnerMixin): self.supports_mm_inputs, self.req_states, self.block_tables, + cls=self.pcp_manager_cls, ) initialize_mamba_ssu_backend( self.vllm_config.mamba_config, self.kv_cache_config @@ -1751,6 +1752,11 @@ class GPUModelRunner(LoRAModelRunnerMixin): ########### EPLB methods end ########### + # Out-of-tree hardware runners can select a PCP manager class. + @property + def pcp_manager_cls(self) -> type[pcp.PCPManager]: + return pcp.PCPManager + class ExecuteModelState(NamedTuple): input_batch: InputBatch diff --git a/vllm/v1/worker/gpu/pcp_manager.py b/vllm/v1/worker/gpu/pcp_manager.py index f50e487494..62bacf5314 100644 --- a/vllm/v1/worker/gpu/pcp_manager.py +++ b/vllm/v1/worker/gpu/pcp_manager.py @@ -654,19 +654,20 @@ def maybe_build_pcp_manager( supports_mm_inputs: bool, req_states: RequestState, block_tables: BlockTables, + cls: type[PCPManager] = PCPManager, ) -> PCPManager | None: parallel_config = vllm_config.parallel_config pcp_size = parallel_config.prefill_context_parallel_size if pcp_size <= 1: return None - PCPManager.validate_config(vllm_config, supports_mm_inputs) + cls.validate_config(vllm_config, supports_mm_inputs) pcp_rank = get_pcp_group().rank_in_group dcp_size = parallel_config.decode_context_parallel_size dcp_rank = get_dcp_group().rank_in_group if dcp_size > 1 else 0 - return PCPManager( + return cls( pcp_world_size=pcp_size, pcp_rank=pcp_rank, device=device,