KeyError: 'default' issue

[anr2me]

I was using the latest transformers (5.9.0) with python 3.13, and got this KeyError: 'default' error at:

Lance/modeling/qwen2_5_vl/modeling_qwen2_5_vl.py

Lines 587 to 594 in 3c757d2

	# HACK: 强制设置为default
	self.rope_type = "default"
	self.max_seq_len_cached = config.max_position_embeddings
	self.original_max_seq_len = config.max_position_embeddings
	self.config = config
	self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

Where self.rope_type is hardcoded to "default", and ROPE_INIT_FUNCTIONS imported with:

from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS

However, according to Claude, "default" key only exist on transformers ~4.40–4.42:

transformers 4.43+ use "llama3" or "dynamic"
Some other builds use "linear", "yarn"

A few suggestions from Claude are:

• Using fallback when "default" doesn't exist:

# Try one of these based on what print() shows you
self.rope_init_fn = ROPE_INIT_FUNCTIONS.get("default") \
    or ROPE_INIT_FUNCTIONS.get("llama3") \
    or list(ROPE_INIT_FUNCTIONS.values())[0]

• Bypass ROPE_INIT_FUNCTIONS entirely and define the init function directly:

def default_rope_init(config, device):
    base = config.rope_theta if hasattr(config, "rope_theta") else 10000.0
    dim = config.hidden_size // config.num_attention_heads
    inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device).float() / dim))
    return inv_freq, 1.0

self.rope_init_fn = default_rope_init

[LutsyferII]

my solution (Write under self.config = config):

    if self.rope_type in ROPE_INIT_FUNCTIONS:
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
    else:
        def custom_rope_init_fn(cfg, dev):
            h_dim = getattr(cfg, "head_dim", cfg.hidden_size // cfg.num_attention_heads)
            base_theta = getattr(cfg, "rope_theta", 1000000.0)
            inv_frq = 1.0 / (base_theta ** (torch.arange(0, h_dim, 2).float().to(dev) / h_dim))
            return inv_frq, 1.0
            
        self.rope_init_fn = custom_rope_init_fn

    inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)