LibertAIDAI

/

GLM-5.3-Flash-NVFP4

like 61

Follow

LibertAI 45

same here on GB10

You need to download the weights beforehand and point vllm to the local folder. Then it’ll work.

@corgito @eugr-nv

I think the issue is caused by accessing the configuration file directory using the model name itself.

If you have mounted the Hugging Face cache, you can work around the problem with the following patch.

--- a/transformers_utils/processors/glm5next.py 2026-08-27 10:44:53.961431596 +0900
+++ b/transformers_utils/processors/glm5next.py 2026-08-27 10:46:56.517110367 +0900
@@ -49,7 +49,7 @@
     VideosKwargs,
 )
 from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
-from transformers.utils import TensorType, logging
+from transformers.utils import TensorType, logging, cached_file
 from transformers.video_processing_utils import BaseVideoProcessor
 from transformers.video_utils import (
     VideoInput,
@@ -850,7 +850,8 @@
             **{k: v for k, v in ip_cfg.items() if k != “image_processor_type”}
         )

-        with open(os.path.join(model_path, “processor_config.json”)) as f:
+        config_path = cached_file(model_path, “processor_config.json”)
+        with open(config_path) as f:
             vp_cfg = _cap_cfg(dict(json.load(f)[“video_processor”]), is_video=True)
         video_processor = Glm5NextVideoProcessor(
             **{k: v for k, v in vp_cfg.items() if k != “video_processor_type”}

Just want to say how awesome LibertAIDAI is at answering questions. For those who want the easy way to solve this on RunPod - download VLLM or whatever latest with no args and download it to your workspace. Then use HF_HUB_OFFLINE=1 to make it search for the downloaded version, then it loads.

Hey! Sorry for not answering directly :) I saw others did it accurately. Thanks for the info on the HF_HUB_OFFLINE flag.