Symptom: AutoModel.from_pretrained("GSAI-ML/LLaDA-8B-Instruct", trust_remote_code=True, ...) hard-segfaults (no Python exception) during weight loading under transformers 5.x. Checkpoint integrity verified first (all six safetensors shards full-size; re-download did not change the behavior).
Environment: transformers 5.12.1, torch 2.11.0+cu128, Python 3.10, Windows 11, RTX 5060 Ti 16 GB. Crash occurs with and without 4-bit quantization (bitsandbytes 0.49.2).
Faulthandler stack at the crash (transformers' new weight-loading path):
Current thread (most recent call first):
...\torch\storage.py", line 471 in __getitem__
...\transformers\core_model_loading.py", line 989 in _materialize_copy
...\transformers\core_model_loading.py", line 1005 in _job
...\transformers\core_model_loading.py", line 797 in materialize_tensors
...\transformers\core_model_loading.py", line 830 in convert
...\transformers\core_model_loading.py", line 1466 in convert_and_load_state_dict_in_model
...\transformers\modeling_utils.py", line 4457 in _load_pretrained_model
...\transformers\modeling_utils.py", line 4327 in from_pretrained
Notes: standard architectures (e.g. the gemma-2 family) load fine in the same environment under transformers 5.12, so this looks specific to the interaction between LLaDA's trust_remote_code modeling files and the transformers ≥5 loader rewrite, rather than a general environment problem. Reproduces 3/3 attempts.
Workaround that fully resolves it: pin transformers>=4.46,<5 — the same environment then loads and generates normally (verified end-to-end, 4-bit on a 16 GB consumer GPU).
Might be worth a note in the README/model card until the remote code is updated for the v5 loader. Happy to provide more details or test a fix. Also reported on the Hugging Face model page's community tab. (Context: we use LLaDA in a public research episode on text-diffusion emergence — write-up with the documented pin here.)
Symptom:
AutoModel.from_pretrained("GSAI-ML/LLaDA-8B-Instruct", trust_remote_code=True, ...)hard-segfaults (no Python exception) during weight loading under transformers 5.x. Checkpoint integrity verified first (all six safetensors shards full-size; re-download did not change the behavior).Environment: transformers 5.12.1, torch 2.11.0+cu128, Python 3.10, Windows 11, RTX 5060 Ti 16 GB. Crash occurs with and without 4-bit quantization (bitsandbytes 0.49.2).
Faulthandler stack at the crash (transformers' new weight-loading path):
Notes: standard architectures (e.g. the gemma-2 family) load fine in the same environment under transformers 5.12, so this looks specific to the interaction between LLaDA's
trust_remote_codemodeling files and the transformers ≥5 loader rewrite, rather than a general environment problem. Reproduces 3/3 attempts.Workaround that fully resolves it: pin
transformers>=4.46,<5— the same environment then loads and generates normally (verified end-to-end, 4-bit on a 16 GB consumer GPU).Might be worth a note in the README/model card until the remote code is updated for the v5 loader. Happy to provide more details or test a fix. Also reported on the Hugging Face model page's community tab. (Context: we use LLaDA in a public research episode on text-diffusion emergence — write-up with the documented pin here.)