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, 14:45 UTC
[Misc] Raise error for V1 not supporting Long LoRA. (#16415)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Apr 11, 2025 at 16:51 UTC
a26f59ccbc0ad4d72acce6595db0f128e83180f1
2 files changed
+7
-1
vllm/config.py
+6
@@ -2573,6 +2573,11 @@ class LoRAConfig:
2573
logger.warning("LoRA with chunked prefill is still experimental "
2574
"and may be unstable.")
2575
2576
+ def verify_lora_support(self):
2577
+ if self.long_lora_scaling_factors is not None and envs.VLLM_USE_V1:
2578
+ raise ValueError(
2579
+ "V1 LoRA does not support long LoRA, please use V0.")
2580
+
2581
2582
@dataclass
2583
class PromptAdapterConfig:
@@ -3672,6 +3677,7 @@ class VllmConfig:
3677
self.lora_config.verify_with_model_config(self.model_config)
3678
self.lora_config.verify_with_scheduler_config(
3679
self.scheduler_config)
3680
+ self.lora_config.verify_lora_support()
3681
if self.prompt_adapter_config:
3682
self.prompt_adapter_config.verify_with_model_config(
3683
self.model_config)
vllm/lora/models.py
+1
-1
@@ -364,7 +364,7 @@ class LoRAModelManager(AdapterModelManager):
364
self._last_mapping: Optional[LoRAMapping] = None
365
self._create_lora_modules()
366
self.model.lora_manager = self
367
- self.adapter_type = 'LoRa'
367
+ self.adapter_type = 'LoRA'
368
369
@property
370
def capacity(self) -> int: