forked from EngineX-Cambricon/enginex-mlu370-vllm
65 lines
2.7 KiB
Python
65 lines
2.7 KiB
Python
from typing import Callable, List, Optional, Tuple, Union, Type
|
|
|
|
from vllm.executor.executor_base import ExecutorAsyncBase
|
|
from vllm.logger import init_logger
|
|
from vllm.model_executor.layers.sampler import SamplerOutput
|
|
from vllm.sequence import ExecuteModelRequest, PoolerOutput
|
|
from vllm.utils import make_async
|
|
from vllm.executor.gpu_executor import GPUExecutor
|
|
from vllm.worker.worker_base import WorkerBase
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
class MLUExecutor(GPUExecutor):
|
|
|
|
def _init_executor(self) -> None:
|
|
"""Initialize the worker and load the model.
|
|
"""
|
|
assert self.parallel_config.world_size == 1, (
|
|
"MLUExecutor only supports single MLU.")
|
|
|
|
self.driver_worker = self._create_worker()
|
|
self.driver_worker.init_device()
|
|
self.driver_worker.load_model()
|
|
|
|
def _get_worker_module_and_class(
|
|
self) -> Tuple[str, str, Optional[Callable[[], Type[WorkerBase]]]]:
|
|
worker_class_fn = None
|
|
if self.scheduler_config.is_multi_step:
|
|
worker_module_name = "vllm.worker.mlu_multi_step_worker"
|
|
worker_class_name = "MLUMultiStepWorker"
|
|
elif self.speculative_config:
|
|
worker_module_name = "vllm.spec_decode.mlu_spec_decode_worker"
|
|
worker_class_name = "create_mlu_spec_worker"
|
|
else:
|
|
worker_module_name = "vllm.worker.mlu_worker"
|
|
worker_class_name = "MLUWorker"
|
|
return (worker_module_name, worker_class_name, worker_class_fn)
|
|
|
|
def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks) -> None:
|
|
"""Initialize the KV cache by invoking the underlying worker.
|
|
"""
|
|
# NOTE: This is logged in the executor because there can be >1 worker
|
|
# with other executors. We could log in the engine level, but work
|
|
# remains to abstract away the device for non-GPU configurations.
|
|
logger.info("# MLU blocks: %d, # CPU blocks: %d", num_gpu_blocks,
|
|
num_cpu_blocks)
|
|
max_concurrency = (num_gpu_blocks * self.cache_config.block_size /
|
|
self.model_config.max_model_len)
|
|
logger.info("Maximum concurrency for %s tokens per request: %.2fx",
|
|
self.model_config.max_model_len, max_concurrency)
|
|
|
|
self.driver_worker.initialize_cache(num_gpu_blocks, num_cpu_blocks)
|
|
|
|
|
|
class MLUExecutorAsync(MLUExecutor, ExecutorAsyncBase):
|
|
|
|
async def execute_model_async(
|
|
self,
|
|
execute_model_req: ExecuteModelRequest,
|
|
) -> List[Union[SamplerOutput, PoolerOutput]]:
|
|
output = await make_async(self.driver_worker.execute_model
|
|
)(execute_model_req=execute_model_req)
|
|
return output
|