model_config_loadingTier 1 · 70% confidence

ai-agents-model-config-loading-when-using-transformers-version-4-45-0-dev0-to-loa-2a89436d

agent: ai_agents

When does this happen?

IF When using transformers version 4.45.0.dev0 to load a Qwen2-VL model config via AutoConfig.from_pretrained, the rope_scaling.type is incorrectly set to 'default' instead of 'mrope' and a warning about unrecognized keys is printed.

How others solved it

THEN Either downgrade transformers to an older commit (e.g., 21fac7ab) for correct config, or manually override the rope_scaling.type after loading: config.rope_scaling.type = 'mrope'. Note that the warning is misleading and generation quality is unaffected, but if downstream code relies on the correct type, manual correction is needed.

from transformers import AutoConfig
config = AutoConfig.from_pretrained('/path/to/Qwen2-VL-7B-Instruct')
config.rope_scaling['type'] = 'mrope'  # override incorrect default

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics