diff --git a/vllm/entrypoints/serve/elastic_ep/api_router.py b/vllm/entrypoints/serve/elastic_ep/api_router.py index 02a2425090..52c68d3321 100644 --- a/vllm/entrypoints/serve/elastic_ep/api_router.py +++ b/vllm/entrypoints/serve/elastic_ep/api_router.py @@ -75,6 +75,8 @@ async def scale_elastic_ep(raw_request: Request): detail="Scale failed due to request drain timeout " f"after {drain_timeout} seconds", ) from e + except ValueError as e: + raise HTTPException(status_code=400, detail=str(e)) from e except Exception as e: logger.error("Scale failed: %s", e) raise HTTPException(status_code=500, detail="Scale failed") from e diff --git a/vllm/v1/engine/core_client.py b/vllm/v1/engine/core_client.py index d692a13d65..a626af89e1 100644 --- a/vllm/v1/engine/core_client.py +++ b/vllm/v1/engine/core_client.py @@ -1637,9 +1637,21 @@ class DPLBAsyncMPClient(DPAsyncMPClient): ) parallel_config = self.vllm_config.parallel_config num_experts = self.vllm_config.model_config.get_num_experts() - num_redundant_experts = ( + num_physical_experts = ( num_experts + parallel_config.eplb_config.num_redundant_experts - ) * new_data_parallel_size // cur_data_parallel_size - num_experts + ) + num_redundant_experts = ( + num_physical_experts * new_data_parallel_size // cur_data_parallel_size + - num_experts + ) + if num_redundant_experts < 0: + # Scaling keeps physical experts per engine fixed, so below this + # size the logical experts no longer fit. + raise ValueError( + f"Cannot scale to data_parallel_size {new_data_parallel_size}, " + f"minimum is " + f"{-(-num_experts * cur_data_parallel_size // num_physical_experts)}" + ) if new_data_parallel_size < cur_data_parallel_size: await self._prepare_scale_down_elastic_ep(new_data_parallel_size) else: