GitHub · Where software is built

Labels Milestones New issue

[Bug] ValueError: passed inputs_embeds to .generate() when using LlamaForCausalLM with Unsloth

#3779

· Pioneer-Weirdo opened

on Dec 25, 2025

[Bug] NameError: name 'psutil' is not defined in UnslothSFTTrainer._prepare_dataset (Windows environment)

#3777

· midorin-Linux opened

on Dec 25, 2025

[Bug] FBGEMM on the L40s cannot load

#3776

· karsarobert opened

on Dec 25, 2025

[Bug]AttributeError: GptOssExperts has no attribute 'down_projs' when loading gpt-oss-20b on RTX 5090

#3775

· BryceSWNS opened

on Dec 25, 2025

[Bug] GGUF conversion fails for large models: Unsloth fails to recognize sharded output files

#3773

· dtometzki opened

on Dec 24, 2025

How to Ensure Loss Is Computed Only on Assistant Responses During VLM SFT

#3772

· Zuozhuo opened

on Dec 24, 2025

For GRPO training in Unsloth, which uses less VRAM—FP8 or 4-bit?

#3771

· l-besiege-l opened

on Dec 24, 2025

[Bug] partial multimodal token full attention not supported

#3770

· dipta007 opened

on Dec 24, 2025

[Bug] UnslothFusedLossBackward tensor breaks with Transformers default <code>average_tokens_across_devices=True</code>, causing loss to become int

#3769

· wang-7 opened

on Dec 23, 2025

[Bug] RuntimeError: Boolean value of Tensor with more than one value is ambiguous

currently fixing

#3766

· qtcg opened

on Dec 23, 2025

[Bug] Error when install unsloth docker in ASUS Ascent GX10

#3764

· PizzaSnow opened

on Dec 23, 2025

[Bug] RuntimeError: Direct module loading failed for unsloth_compiled_module_gpt2: unexpected indent (unsloth_compiled_module_gpt2.py, line 730)

#3763

· zhi-lu opened

on Dec 23, 2025