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, 13:30 UTC
[BugFix] Pass in correct VLLM config in FlashInfer backend (#13207) (#16973)
Signed-off-by: 苏政渊 <suzhengyuan@moonshot.cn> Co-authored-by: 苏政渊 <suzhengyuan@moonshot.cn>
Zhengyuan Su (苏政渊) committed
Apr 22, 2025 at 21:44 UTC
f961d7f6ef1441b9e3674cd9df5ddaee984ce3ac
1 file changed
+3
-3
vllm/attention/backends/flashinfer.py
+3
-3
@@ -37,7 +37,7 @@ from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping,
37
is_block_tables_empty)
38
from vllm.attention.layer import Attention
39
from vllm.attention.ops.paged_attn import PagedAttention
40
-from vllm.config import VllmConfig, get_current_vllm_config
40
+from vllm.config import VllmConfig
41
from vllm.logger import init_logger
42
from vllm.utils import (async_tensor_h2d, get_kv_cache_torch_dtype,
43
make_tensor_with_pad)
@@ -187,7 +187,7 @@ class FlashInferState(AttentionState):
187
# Global hyperparameters shared by all attention layers
188
self.global_hyperparameters: Optional[PerLayerParameters] = None
189
190
- self.vllm_config = get_current_vllm_config()
190
+ self.vllm_config = self.runner.vllm_config
191
192
def _get_workspace_buffer(self):
193
if self._workspace_buffer is None:
@@ -613,7 +613,7 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
613
# Global hyperparameters shared by all attention layers
614
self.global_hyperparameters: Optional[PerLayerParameters] = None
615
616
- self.vllm_config = get_current_vllm_config()
616
+ self.vllm_config = self.runner.vllm_config
617
618
def prepare(self):
619
self.slot_mapping: List[int] = []