[MISC] fix pin_memory=torch.cuda.is_available(), use is_pin_memory_available (#37415)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/v1/structured_output/backend_lm_format_enforcer.py

@@ -11,6 +11,7 @@ from transformers import PreTrainedTokenizerBase
 
 from vllm.sampling_params import SamplingParams
 from vllm.utils.import_utils import LazyLoader
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.structured_output.backend_types import (
     StructuredOutputBackend,
     StructuredOutputGrammar,
@@ -138,7 +139,7 @@ class LMFormatEnforcerBackend(StructuredOutputBackend):
             (max_num_seqs, (self.vocab_size + 31) // 32),
             -1,
             dtype=torch.int32,
-            pin_memory=torch.cuda.is_available(),
+            pin_memory=is_pin_memory_available(),
         )
 
     def destroy(self):

vllm/v1/structured_output/backend_outlines.py

@@ -15,6 +15,7 @@ from regex import escape as regex_escape
 
 from vllm.sampling_params import SamplingParams
 from vllm.utils.import_utils import LazyLoader
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.structured_output.backend_types import (
     StructuredOutputBackend,
     StructuredOutputGrammar,
@@ -96,7 +97,7 @@ class OutlinesBackend(StructuredOutputBackend):
             (max_num_seqs, (self.vocab_size + 31) // 32),
             -1,
             dtype=torch.int32,
-            pin_memory=torch.cuda.is_available(),
+            pin_memory=is_pin_memory_available(),
         )
 
     def destroy(self):