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 15, 04:45 UTC
[Misc] Fix import (#20233)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Woosuk Kwon committed
Jun 29, 2025 at 20:34 UTC
19108ef31191e217766ffe52e8e382ddbec20fdb
1 file changed
+1
-2
vllm/v1/worker/gpu_model_runner.py
+1
-2
@@ -45,7 +45,7 @@ from vllm.sequence import IntermediateTensors
45
from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
46
GiB_bytes, LazyLoader, async_tensor_h2d, cdiv,
47
check_use_alibi, get_dtype_size,
48
- is_pin_memory_available)
48
+ is_pin_memory_available, round_up)
49
from vllm.v1.attention.backends.mamba_attn import Mamba2AttentionBackend
50
from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
51
CommonAttentionMetadata)
@@ -1308,7 +1308,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
1308
tp_size = self.vllm_config.parallel_config.tensor_parallel_size
1309
if self.compilation_config.pass_config. \
1310
enable_sequence_parallelism and tp_size > 1:
1311
- from vllm.utils import round_up
1311
num_input_tokens = round_up(num_scheduled_tokens, tp_size)
1312
else:
1313
num_input_tokens = num_scheduled_tokens