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
Add deprecation warning for lora_extra_vocab_size (#23635)
Signed-off-by: Jinheng Li <ahengljh@gmail.com>
Jinheng committed
Aug 28, 2025 at 13:34 UTC
c8851a47235f5dfd3da3abf6c89453b3bdb41ad1
1 file changed
+8
-2
vllm/config/__init__.py
+8
-2
@@ -2439,8 +2439,8 @@ class LoRAConfig:
2439
lora_dtype: Union[torch.dtype, LoRADType] = "auto"
2440
"""Data type for LoRA. If auto, will default to base model dtype."""
2441
lora_extra_vocab_size: int = 256
2442
- """Maximum size of extra vocabulary that can be present in a LoRA adapter
2443
- (added to the base model vocabulary)."""
2442
+ """(Deprecated) Maximum size of extra vocabulary that can be present in a
2443
+ LoRA adapter. Will be removed in v0.12.0."""
2444
lora_vocab_padding_size: ClassVar[int] = current_platform\
2445
.get_lora_vocab_padding_size()
2446
@@ -2482,6 +2482,12 @@ class LoRAConfig:
2482
return hash_str
2483
2484
def __post_init__(self):
2485
+ # Deprecation warning for lora_extra_vocab_size
2486
+ logger.warning(
2487
+ "`lora_extra_vocab_size` is deprecated and will be removed "
2488
+ "in v0.12.0. Additional vocabulary support for "
2489
+ "LoRA adapters is being phased out.")
2490
+
2491
# Setting the maximum rank to 512 should be able to satisfy the vast
2492
# majority of applications.
2493
possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)