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
[BUG] Reorder model config creation (#26124)
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
ahao-anyscale committed
Oct 2, 2025 at 23:59 UTC
c4b48d3c0f86d355c55c28a31029443df3727bc0
1 file changed
+4
-1
vllm/engine/arg_utils.py
+4
-1
@@ -1131,6 +1131,10 @@ class EngineArgs:
1131
device_config = DeviceConfig(
1132
device=cast(Device, current_platform.device_type))
1133
1134
+ model_config = self.create_model_config()
1135
+ self.model = model_config.model
1136
+ self.tokenizer = model_config.tokenizer
1137
+
1138
(self.model, self.tokenizer,
1139
self.speculative_config) = maybe_override_with_speculators(
1140
model=self.model,
@@ -1139,7 +1143,6 @@ class EngineArgs:
1143
trust_remote_code=self.trust_remote_code,
1144
vllm_speculative_config=self.speculative_config,
1145
)
1142
- model_config = self.create_model_config()
1146
1147
# * If VLLM_USE_V1 is unset, we enable V1 for "supported features"
1148
# and fall back to V0 for experimental or unsupported features.