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 6, 14:30 UTC
[Bugfix] Fix DP MTP Dummy Run (#35243)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Benjamin Chislett committed
Mar 17, 2026 at 11:16 UTC
f63ed7b5aca634b23d070b9cd9f654f0c74b65ad
1 file changed
+2
-1
vllm/v1/worker/gpu_worker.py
+2
-1
@@ -903,7 +903,8 @@ class Worker(WorkerBase):
903
self.profiler.stop()
904
905
def execute_dummy_batch(self) -> None:
906
- self.model_runner._dummy_run(1, uniform_decode=True)
906
+ num_tokens = getattr(self.model_runner, "uniform_decode_query_len", 1)
907
+ self.model_runner._dummy_run(num_tokens, uniform_decode=True)
908
909
def add_lora(self, lora_request: LoRARequest) -> bool:
910
return self.model_runner.add_lora(lora_request)