A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI] Fix mypy for `vllm/executor` (#26845)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 15, 2025 at 04:23 UTC
71557a5f7c221b63759a0d87c0b175b1bab243e6
4 files changed
+23
-11
tools/pre_commit/mypy.py
+1
-1
@@ -28,6 +28,7 @@ FILES = [
28
"vllm/assets",
29
"vllm/distributed",
30
"vllm/entrypoints",
31
+ "vllm/executor",
32
"vllm/inputs",
33
"vllm/logging_utils",
34
"vllm/multimodal",
@@ -44,7 +45,6 @@ SEPARATE_GROUPS = [
45
"vllm/attention",
46
"vllm/compilation",
47
"vllm/engine",
47
- "vllm/executor",
48
"vllm/inputs",
49
"vllm/lora",
50
"vllm/model_executor",
vllm/executor/executor_base.py
+4
-3
@@ -18,7 +18,7 @@ from vllm.lora.request import LoRARequest
18
from vllm.sequence import ExecuteModelRequest
19
from vllm.tasks import SupportedTask
20
from vllm.utils import make_async
21
-from vllm.v1.outputs import PoolerOutput, SamplerOutput
21
+from vllm.v1.outputs import SamplerOutput
22
from vllm.v1.worker.worker_base import WorkerBase
23
24
logger = init_logger(__name__)
@@ -54,7 +54,7 @@ class ExecutorBase(ABC):
54
self._init_executor()
55
self.is_sleeping = False
56
self.sleeping_tags: set[str] = set()
57
- self.kv_output_aggregator = None
57
+ self.kv_output_aggregator: KVOutputAggregator | None = None
58
59
@abstractmethod
60
def _init_executor(self) -> None:
@@ -143,8 +143,9 @@ class ExecutorBase(ABC):
143
144
def execute_model(
145
self, execute_model_req: ExecuteModelRequest
146
- ) -> list[SamplerOutput | PoolerOutput] | None:
146
+ ) -> list[SamplerOutput]:
147
output = self.collective_rpc("execute_model", args=(execute_model_req,))
148
+ assert output[0] is not None
149
return output[0]
150
151
def stop_remote_worker_execution_loop(self) -> None:
vllm/executor/ray_distributed_executor.py
+12
-6
@@ -217,7 +217,9 @@ class RayDistributedExecutor(DistributedExecutorBase):
217
num_gpus=num_gpus,
218
scheduling_strategy=scheduling_strategy,
219
**ray_remote_kwargs,
220
- )(RayWorkerWrapper).remote(vllm_config=self.vllm_config, rpc_rank=rank)
220
+ )(RayWorkerWrapper).remote( # type: ignore[attr-defined]
221
+ vllm_config=self.vllm_config, rpc_rank=rank
222
+ )
223
else:
224
worker = ray.remote(
225
num_cpus=0,
@@ -225,7 +227,9 @@ class RayDistributedExecutor(DistributedExecutorBase):
227
resources={current_platform.ray_device_key: num_gpus},
228
scheduling_strategy=scheduling_strategy,
229
**ray_remote_kwargs,
228
- )(RayWorkerWrapper).remote(vllm_config=self.vllm_config, rpc_rank=rank)
230
+ )(RayWorkerWrapper).remote( # type: ignore[attr-defined]
231
+ vllm_config=self.vllm_config, rpc_rank=rank
232
+ )
233
worker_metadata.append(RayWorkerMetaData(worker=worker, created_rank=rank))
234
235
worker_ips = ray.get(
@@ -303,7 +307,7 @@ class RayDistributedExecutor(DistributedExecutorBase):
307
continue
308
worker_node_and_gpu_ids.append(
309
ray.get(worker.get_node_and_gpu_ids.remote())
306
- ) # type: ignore
310
+ ) # type: ignore[attr-defined]
311
312
node_workers = defaultdict(list) # node id -> list of worker ranks
313
node_gpus = defaultdict(list) # node id -> list of gpu ids
@@ -495,7 +499,9 @@ class RayDistributedExecutor(DistributedExecutorBase):
499
if async_run_tensor_parallel_workers_only:
500
ray_workers = self.non_driver_workers
501
ray_worker_outputs = [
498
- worker.execute_method.remote(sent_method, *args, **kwargs)
502
+ worker.execute_method.remote( # type: ignore[attr-defined]
503
+ sent_method, *args, **kwargs
504
+ )
505
for worker in ray_workers
506
]
507
@@ -715,7 +721,7 @@ class RayDistributedExecutor(DistributedExecutorBase):
721
tasks.append(
722
asyncio.create_task(
723
_run_task_with_lock(
718
- driver_worker.execute_method.remote,
724
+ driver_worker.execute_method.remote, # type: ignore[attr-defined]
725
self.pp_locks[pp_rank],
726
"execute_model",
727
execute_model_req,
@@ -733,7 +739,7 @@ class RayDistributedExecutor(DistributedExecutorBase):
739
"worker loop is disabled for VLLM_USE_RAY_SPMD_WORKER=1"
740
)
741
coros = [
736
- worker.execute_method.remote("start_worker_execution_loop")
742
+ worker.execute_method.remote("start_worker_execution_loop") # type: ignore[attr-defined]
743
for worker in self.non_driver_workers
744
]
745
return await asyncio.gather(*coros)
vllm/executor/ray_utils.py
+6
-1
@@ -90,14 +90,17 @@ try:
90
91
execute_model_req = self.input_decoder.decode(serialized_req)
92
93
+ assert self.worker is not None, "Worker is not initialized"
94
+
95
# TODO(swang): This is needed right now because Ray Compiled Graph
96
# executes on a background thread, so we need to reset torch's
97
# current device.
98
if not self.compiled_dag_cuda_device_set:
99
+ assert self.worker.device is not None
100
current_platform.set_device(self.worker.device)
101
self.compiled_dag_cuda_device_set = True
102
100
- output = self.worker._execute_model_spmd(
103
+ output = self.worker._execute_model_spmd( # type: ignore[attr-defined]
104
execute_model_req, intermediate_tensors
105
)
106
# Pipeline model request and output to the next pipeline stage.
@@ -119,6 +122,7 @@ try:
122
# Not needed
123
pass
124
else:
125
+ assert self.worker.device is not None
126
current_platform.set_device(self.worker.device)
127
128
self.compiled_dag_cuda_device_set = True
@@ -139,6 +143,7 @@ try:
143
scheduler_output, intermediate_tensors = scheduler_output
144
else:
145
scheduler_output, intermediate_tensors = scheduler_output, None
146
+ assert self.worker.model_runner is not None
147
output = self.worker.model_runner.execute_model(
148
scheduler_output, intermediate_tensors
149
)