Missing custom rope-scaling
Hi! I’m trying to load this Llama-3.1-WhiteRabbitNeo-2-8B model via AutoModelForCausalLM.from_pretrained(...) in Transformers, but I get a rope_scaling error:
ValueError: rope_scaling must be a dictionary with two fields, type and factor,
got {'factor': 8.0, 'high_freq_factor': 4.0, 'low_freq_factor': 1.0,
'original_max_position_embeddings': 8192, 'rope_type': 'llama3'}
It looks like the config uses a custom rope-scaling method (“llama3”) that vanilla Transformers doesn’t recognize. I don’t see a custom Python loader file in the repo (like modeling_llama3.py), so is there a specialized script or fork needed to load this model as intended?
If so, could you please share that loader or instructions on how to enable the “llama3” rope scaling? I’d prefer not to remove or override the rope_scaling block, as I’d like to keep the extended-context benefits.
Thanks in advance—excited to try out your model once this is sorted!