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, 10:45 UTC
[MISC] fix pin_memory=torch.cuda.is_available(), use is_pin_memory_available (#37415)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Mar 19, 2026 at 17:23 UTC
ca21483bf900b269dd1876352ee335ab62df2ebb
2 files changed
+4
-2
vllm/v1/structured_output/backend_lm_format_enforcer.py
+2
-1
@@ -11,6 +11,7 @@ from transformers import PreTrainedTokenizerBase
11
12
from vllm.sampling_params import SamplingParams
13
from vllm.utils.import_utils import LazyLoader
14
+from vllm.utils.platform_utils import is_pin_memory_available
15
from vllm.v1.structured_output.backend_types import (
16
StructuredOutputBackend,
17
StructuredOutputGrammar,
@@ -138,7 +139,7 @@ class LMFormatEnforcerBackend(StructuredOutputBackend):
139
(max_num_seqs, (self.vocab_size + 31) // 32),
140
-1,
141
dtype=torch.int32,
141
- pin_memory=torch.cuda.is_available(),
142
+ pin_memory=is_pin_memory_available(),
143
)
144
145
def destroy(self):
vllm/v1/structured_output/backend_outlines.py
+2
-1
@@ -15,6 +15,7 @@ from regex import escape as regex_escape
15
16
from vllm.sampling_params import SamplingParams
17
from vllm.utils.import_utils import LazyLoader
18
+from vllm.utils.platform_utils import is_pin_memory_available
19
from vllm.v1.structured_output.backend_types import (
20
StructuredOutputBackend,
21
StructuredOutputGrammar,
@@ -96,7 +97,7 @@ class OutlinesBackend(StructuredOutputBackend):
97
(max_num_seqs, (self.vocab_size + 31) // 32),
98
-1,
99
dtype=torch.int32,
99
- pin_memory=torch.cuda.is_available(),
100
+ pin_memory=is_pin_memory_available(),
101
)
102
103
def destroy(self):