Question about rope_scaling.factor (as null) in config.json
Dear inclusionAI Team,
Thank you for bringing out this model! During reviewing the architecture and using this model, I found that in config.json, the rope scaling part for the backbone llm is written as
"rope_scaling": {
"factor": null,
"type": "video_rope"
},
where factor is given null. I wonder if this "factor": null created by design and actually used somewhere? And shall we remove this attribute from config.json if it's not functioning anywhere. I thought this factor does not participate in model init or computation when we have video_rope as rope type.
Hey Jonathan1909,
Thank you for your question and for your interest in our model!
Regarding the rope_scaling field in config.json, you are correct that the "factor": null is intentionally set. In our implementation, the type field ("video_rope") is used to select the specific RoPE scaling method, while the factor parameter is not utilized in this context. We set it to null as a placeholder to maintain compatibility with certain codebases that expect this field to exist, even though it does not affect the model's initialization or computation.
For more details on how video_rope is implemented, you can refer to the code in modeling_bailing_moe_v2.py from lines 314 to 474. There you will see the actual logic that handles the RoPE scaling for video modalities.
We hope this clarifies your question. If you have any further inquiries, feel free to ask!
Ming team