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 7, 11:30 UTC
clean unused cudagraph_batch_sizes (#35552)
Signed-off-by: Boyuan Feng <boyuan@meta.com>
Boyuan Feng committed
Mar 2, 2026 at 14:00 UTC
c42dc402c14817c1c329aa5488d78eb204d4b4c1
1 file changed
-13
vllm/v1/worker/gpu_model_runner.py
-13
@@ -597,15 +597,6 @@ class GPUModelRunner(
597
self.async_output_copy_stream = torch.cuda.Stream()
598
self.prepare_inputs_event = torch.Event()
599
600
- # self.cudagraph_batch_sizes sorts in ascending order.
601
- if (
602
- self.compilation_config.cudagraph_capture_sizes
603
- and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
604
- ):
605
- self.cudagraph_batch_sizes = sorted(
606
- self.compilation_config.cudagraph_capture_sizes
607
- )
608
-
600
# Cache the device properties.
601
self._init_device_properties()
602
@@ -5705,10 +5696,6 @@ class GPUModelRunner(
5696
self.compilation_config.adjust_cudagraph_sizes_for_spec_decode(
5697
self.uniform_decode_query_len, self.parallel_config.tensor_parallel_size
5698
)
5708
- capture_sizes = self.compilation_config.cudagraph_capture_sizes
5709
- self.cudagraph_batch_sizes = (
5710
- capture_sizes if capture_sizes is not None else []
5711
- )
5699
5700
# Trigger cudagraph dispatching keys initialization after
5701
# resolved cudagraph mode.