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
[V0 deprecation] Clean up LoRA (#25686)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Sep 26, 2025 at 02:12 UTC
0fa673af4c2aeaf6d21a0163dc55218f5ee1daa6
1 file changed
+1
-8
vllm/lora/punica_wrapper/punica_gpu.py
+1
-8
@@ -11,7 +11,6 @@ from typing import Optional, Union, final
11
12
import torch
13
14
-import vllm.envs as envs
14
from vllm.lora.layers import LoRAMapping
15
from vllm.triton_utils import HAS_TRITON
16
@@ -41,14 +40,8 @@ class PunicaWrapperGPU(PunicaWrapperBase):
40
max_num_batched_tokens,
41
device=device)
42
44
- # When cudagraph capture size is greater than max_num_seqs (max_batches,
45
- # here), V0 captures the graph as if max_num_seqs is set to
46
- # the capture size.
47
- # V1 doesn't have this problem and always respects max_num_seqs.
48
- max_num_prompts = (max_batches
49
- if envs.VLLM_USE_V1 else max_num_batched_tokens)
43
self.prompt_mapping_meta = LoRAKernelMeta.make(self.max_loras,
51
- max_num_prompts,
44
+ max_batches,
45
device=device)
46
47
def update_metadata(self, mapping: LoRAMapping,