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 6, 09:30 UTC
Fix `SpeculatorsConfig` now that `PreTrainedConfig` is a `dataclass` in Transformers (#37574)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Mar 19, 2026 at 18:04 UTC
e5d96dc8fce24043dd50d1671a7674bda1728f7f
1 file changed
+7
vllm/transformers_utils/configs/speculators/base.py
+7
@@ -14,6 +14,13 @@ from vllm.transformers_utils.utils import without_trust_remote_code
14
class SpeculatorsConfig(PretrainedConfig):
15
model_type = "speculators"
16
17
+ def __init__(self, **kwargs):
18
+ """In Transformers v5, `PretrainedConfig` is decorated with `dataclass` and
19
+ `huggingface_hub.dataclasses.strict(accept_kwargs=True)`.
20
+ Inheriting classes do not inherit the `accept_kwargs=True` behaviour so we must
21
+ explicitly pass any kwargs to `PretrainedConfig.__init__`."""
22
+ super().__init__(**kwargs)
23
+
24
@classmethod
25
def from_pretrained(
26
cls,