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
[BugFix] Fix pipeline parallel (#24621)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Sep 10, 2025 at 23:05 UTC
e2d8c27f68687476cb6c45723c44193b517308ea
3 files changed
+6
-2
vllm/executor/uniproc_executor.py
+4
@@ -71,6 +71,10 @@ class UniProcExecutor(ExecutorBase):
71
self.shutdown()
72
return
73
74
+ def shutdown(self) -> None:
75
+ if worker := self.driver_worker:
76
+ worker.shutdown()
77
+
78
79
UniProcExecutorAsync = UniProcExecutor
80
vllm/v1/worker/gpu_model_runner.py
-1
@@ -2070,7 +2070,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
2070
sampler_output = self._sample(logits, spec_decode_metadata)
2071
2072
with record_function_or_nullcontext("Bookkeep"):
2073
- assert isinstance(hidden_states, torch.Tensor)
2073
(
2074
num_nans_in_logits,
2075
logprobs_lists,
vllm/v1/worker/kv_connector_model_runner_mixin.py
+2
-1
@@ -45,7 +45,8 @@ class KVConnectorModelRunnerMixin:
45
46
@staticmethod
47
def ensure_kv_transfer_shutdown() -> None:
48
- if has_kv_transfer_group():
48
+ # has_kv_transfer_group can be None during interpreter shutdown.
49
+ if has_kv_transfer_group and has_kv_transfer_group():
50
ensure_kv_transfer_shutdown()
51
52
@staticmethod