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
Lora bias(enable_lora_bias) deprecate warning (#24339)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Ashwin Phadke committed
Sep 6, 2025 at 22:12 UTC
6024d115cdd4d23b117d4ba9bd27d7e0311fed19
1 file changed
+7
-2
vllm/config/__init__.py
+7
-2
@@ -2458,7 +2458,6 @@ class LoRAConfig:
2458
LoRA adapter. Will be removed in v0.12.0."""
2459
lora_vocab_padding_size: ClassVar[int] = current_platform\
2460
.get_lora_vocab_padding_size()
2461
-
2461
default_mm_loras: Optional[dict[str, str]] = None
2462
"""Dictionary mapping specific modalities to LoRA model paths; this field
2463
is only applicable to multimodal models and should be leveraged when a
@@ -2470,7 +2469,8 @@ class LoRAConfig:
2469
will be automatically assigned to 1-n with the names of the modalities
2470
in alphabetic order."""
2471
bias_enabled: bool = False
2473
- """Enable bias for LoRA adapters."""
2472
+ """[DEPRECATED] Enable bias for LoRA adapters. This option will be
2473
+ removed in v0.12.0."""
2474
2475
def compute_hash(self) -> str:
2476
"""
@@ -2503,6 +2503,11 @@ class LoRAConfig:
2503
"in v0.12.0. Additional vocabulary support for "
2504
"LoRA adapters is being phased out.")
2505
2506
+ # Deprecation warning for enable_lora_bias
2507
+ if self.bias_enabled:
2508
+ logger.warning("`enable_lora_bias` is deprecated "
2509
+ "and will be removed in v0.12.0.")
2510
+
2511
# Setting the maximum rank to 512 should be able to satisfy the vast
2512
# majority of applications.
2513
possible_max_ranks = (8, 16, 32, 64, 128, 256, 320, 512)