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 7, 16:30 UTC
[Bugfix] Remove erroneous lower bound on LoRA vocab size constraint (#35354)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Feb 26, 2026 at 13:44 UTC
5e58bdc7113a2c62a9bfb71304d0d1563b0da7f3
2 files changed
+4
-6
tests/lora/test_layers.py
+2
-2
@@ -469,7 +469,7 @@ def test_lm_head_logits_processor(
469
470
471
@torch.inference_mode()
472
-@pytest.mark.parametrize("vocab_size", [512, 32000, 258049, 300000])
472
+@pytest.mark.parametrize("vocab_size", [258049, 300000])
473
@pytest.mark.parametrize("device", DEVICES)
474
def test_lm_head_logits_processor_invalid_vocab_size(
475
default_vllm_config, dist_init, vocab_size, device
@@ -489,7 +489,7 @@ def test_lm_head_logits_processor_invalid_vocab_size(
489
logits_processor, 1024, torch.float16, device, None
490
)
491
492
- with pytest.raises(ValueError, match="vocab size must be > 32000 and <= 258048"):
492
+ with pytest.raises(ValueError, match="vocab size must be <= 258048"):
493
lora_logits_processor.create_lora_weights(max_loras, lora_config)
494
495
vllm/lora/layers/logits_processor.py
+2
-4
@@ -88,10 +88,8 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):
88
model_config: PretrainedConfig | None = None,
89
) -> None:
90
# TODO: Verify if this condition can be further relaxed
91
- if self.base_layer.vocab_size <= 32000 or self.base_layer.vocab_size > 258048:
92
- raise ValueError(
93
- "When using LoRA, vocab size must be > 32000 and <= 258048"
94
- )
91
+ if self.base_layer.vocab_size > 258048:
92
+ raise ValueError("When using LoRA, vocab size must be <= 258048")
93
self.lora_a_stacked = torch.zeros(
94
(
95
max_loras,