4.1V Model and GLM-4.5V Model Conversion Code Updates #41784
Conversation
zRzRzRzRzRzRzR
commented
Oct 22, 2025
zRzRzRzRzRzRzR
commented
- Fixed weight conversion issues for some model providers and removed some debug logs
- Simplified some functions
zRzRzRzRzRzRzR
changed the title
|
Hi, is there a reference somewhere for the issues this is fixing? |
|
No, but I encountered some bugs when converting the open-source model at https://github.com/thu-coai/Glyph, so I fixed them here together. |
There was a problem hiding this comment.
Let's delete all the test files
| "n_shared_experts": text_config.get("n_shared_experts", 1), | ||
| "norm_topk_prob": text_config.get("norm_topk_prob", True), | ||
| "num_experts_per_tok": text_config.get("num_experts_per_tok", 8), | ||
| "rope_scaling": {"type": "default", "mrope_section": [8, 12, 12]}, |
There was a problem hiding this comment.
In the main branch it is now called rope_parameters and also includes the theta inside the dict. So maybe
"rope_scaling": {"rope_type": "default", "rope_theta": 10000.0, "mrope_section": [8, 12, 12]},
There was a problem hiding this comment.
I've pushed the updates, is this how I understood it?
zucchini-nlp
left a comment
zucchini-nlp
left a comment
There was a problem hiding this comment.
Thanks, I didn't review in detail and I will trust it converts correctly from the original format. Just left a few nits about config attributes naming
| @@ -1,4 +1,3 @@ | |||
| # coding=utf-8 | |||
| "rope_theta": model_config.get("rotary_base", 10000.0), | ||
| "image_token_id": model_config.get("image_token_id", 151363), | ||
| "video_token_id": model_config.get("video_token_id", 151364), | ||
| "tie_word_embeddings": False, |
There was a problem hiding this comment.
this one is also part of text config
| "use_cache": text_config.get("use_cache", True), | ||
| "vocab_size": text_config.get("vocab_size", 151552), | ||
| "partial_rotary_factor": 0.5, | ||
| "rope_scaling": {"rope_type": "default", "rope_theta": 10000.0, "mrope_section": [8, 12, 12]}, |
There was a problem hiding this comment.
lets' call this key rope_parameters to align with recent changes
| "video_token_id": model_config.get("video_token_id", 151344), | ||
| "image_token_id": model_config.get("image_token_id", 151363), | ||
| "video_token_id": model_config.get("video_token_id", 151364), | ||
| "tie_word_embeddings": False, |
There was a problem hiding this comment.
same here for tie_word_embeddings and rope_parameters
|
[For maintainers] Suggested jobs to run (before merge) run-slow: glm4v, glm4v_moe |
|
Thanks, lets merge! |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
) * update for new model convert * Update convert_glm4v_moe_mgt_weights_to_hf.py * restore * Update convert_glm4v_mgt_weights_to_hf.py * update * 1 * Update convert_glm4v_moe_mgt_weights_to_hf.py * Update convert_glm4v_mgt_weights_to_hf.py * finish * update * 2 * 2 * 1 * Update convert_glm4v_moe_mgt_weights_to_hf.py * update * update with tie_word_embeddings place
4 participants
