[Bugfix][Elastic EP] Reject scale below the minimum data parallel size (#52702)
Signed-off-by: Itay Etelis <itay.etelis@ibm.com> Co-authored-by: Itay Etelis <itay.etelis@ibm.com>
This commit is contained in:
co-authored by
Itay Etelis
parent
b05ae5dc00
commit
d36cc4254e
@@ -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
|
||||
|
||||
@@ -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:
|
||||
|
||||
Reference in New Issue
Block a user