4 Pull requests merged by 1 person

• Qwen3 inference fixes

  #2436 merged Apr 30, 2025

• Fixup qwen3 qk norm

  #2427 merged Apr 29, 2025

• Fixup qwen3

  #2423 merged Apr 29, 2025

• [WIP] Initial support for Qwen3. Will udpate when the model is released

  #2211 merged Apr 29, 2025

1 Pull request opened by 1 person

• Added missing code of conduct

  #2416 opened Apr 26, 2025

33 Issues closed by 12 people

• [Question]torch._dynamo.exc.Unsupported: 'inline in skipfiles: device.__init__ | __init__ /home/miniconda3/envs/unsloth_env/lib/python3.11/site-packages/torch/cuda/__init__.py, skipped according trace_rules.lookup SKIP_DIRS'

  #2429 closed Apr 30, 2025

• Continued Pre-Training Notebook not working with unsloth/Llama-3.2-1B-bnb-4bit

  #1210 closed Apr 29, 2025

• Is there proper attention masking done when applying packing=true?

  #1207 closed Apr 29, 2025

• NameError: name 'Unpack' is not defined

  #1181 closed Apr 29, 2025

• Error - 'OutOfMemoryError: CUDA out of memory.'

  #1214 closed Apr 29, 2025

• Please add the model: EleutherAI/polyglot-ko-5.8b

  #1209 closed Apr 29, 2025

• [Bug] unsloth `ImportError` when using `triton==3.3.0`

  #2403 closed Apr 29, 2025

• [Feature] When will Dynamic Quants 2.0 be available for custom models?

  #2421 closed Apr 29, 2025

• ORPO trainer not works after SFT

  #1203 closed Apr 28, 2025

• [Bug] NotImplementedError: Unsloth: llama.cpp GGUF conversion does not yet support converting model types of Gemma3ForCausalLM.

  #2422 closed Apr 28, 2025

• [Question] How to train the "pt" gemma model with text corpus ? any sample unsloth notebook ?

  #2420 closed Apr 28, 2025

• "FlashAttention only support fp16 and bf16 data type" error when using dora

  #1154 closed Apr 27, 2025

• raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'") AttributeError: 'LongRopeRotaryEmbedding' object has no attribute 'long_cos_cached'. Did you mean: 'short_cos_cached'?

  #1189 closed Apr 27, 2025

• Issue saving mistral-7b-instruct-v0.3-bnb-4bit to GGUF

  #1197 closed Apr 27, 2025

• [Feature] Implement finetunning for new gemma int4 models

  #2412 closed Apr 27, 2025

• [FEAT] integration of SmolVLM

  #2320 closed Apr 27, 2025

• [Bug] FastLanguageModel.from_pretrained() secretly changes the model I really want to fine-tune

  #2407 closed Apr 27, 2025

• [Docs]how to understand the R1-2.51bits or V3-2.71bits

  #2344 closed Apr 27, 2025

• [Docs]benchmark compare between 2.71 and 4.5 bit?

  #2345 closed Apr 27, 2025

• Train_on_completions cant handle eval_datasets as dictionary

  #1192 closed Apr 26, 2025

• [Bug] AttributeError: 'Mistral3ForConditionalGeneration' object has no attribute 'model'

  #2415 closed Apr 26, 2025

• how can i get the unsloth pro version？

  #2414 closed Apr 26, 2025

• does unlsoth support freeze tunning

  #1183 closed Apr 25, 2025

• pip install --upgrade --no-cache-dir unsloth BROKE CUDA packages. Inference slower.

  #1187 closed Apr 25, 2025

• ModuleNotFoundError : Failed to import transformers.models.falcon_mamba.configuration_falcon_mamba

  #1185 closed Apr 25, 2025

• Can't import unsloth when both the latest version of unsloth and transformers are installed

  #1179 closed Apr 25, 2025

• [Bug] PatchDPOTrainer - AttributeError: 'dict' object has no attribute 'logits'

  #2406 closed Apr 25, 2025

• [Bug] NameError: name 'tokenizer_call' is not defined

  #2400 closed Apr 24, 2025

• A bug in save.py

  #1170 closed Apr 24, 2025

• Fine tuning without GPU?

  #1132 closed Apr 24, 2025

• Gradient accumulation fix does change the max_steps value

  #1163 closed Apr 24, 2025

• [Question] Which notebook should I use for continue pretraining LLM with domain knowledge?

  #2402 closed Apr 24, 2025

• How to use this as the reference policy?

  #1167 closed Apr 23, 2025

22 Issues opened by 22 people

• [GRPO+LoRA] no attribute 'load_lora'

  #2437 opened Apr 30, 2025

• multi-GPU training

  #2435 opened Apr 30, 2025

• [Bug] GRPO Qwen3 Failed: Inference tensors cannot be saved for backward. To work around you can make a clone to get a normal tensor and use it in autograd.

  #2434 opened Apr 30, 2025

• [Bug] error installing and then importing on any currently supported image of SageMaker AWS

  #2433 opened Apr 29, 2025

• [Bug] PassManager::run failed when training on Google Colab

  #2432 opened Apr 29, 2025

• [Bug]RTX 5090 error

  #2431 opened Apr 29, 2025

• RuntimeError: Expected there to be 1 prompt updates corresponding to 1 image items, but instead found 0 prompt updates!

  #2430 opened Apr 29, 2025

• QWEN3 FINE-TUNING now in Unsloth!

  #2428 opened Apr 29, 2025

• Llama3_1_(3B)_GRPO_LoRA vllm error

  #2426 opened Apr 29, 2025

• [Question] Dose DDP support for Unsloth Open?

  #2425 opened Apr 29, 2025

• GRPO Training: Repeated Output After Initial Normal Output

  #2424 opened Apr 29, 2025

• AttributeError: 'HybridCache' object has no attribute 'float'

  #2419 opened Apr 27, 2025

• [Bug] Loss not decreasing with Qwen 2.5 32B

  #2417 opened Apr 26, 2025

• [Bug] Latest Unsloth cannot load local models. AttributeError: 'NoneType' object has no attribute 'count'

  #2413 opened Apr 25, 2025

• [Question] Adding several peft adapters and ensuring unsloth takes them into account

  #2411 opened Apr 25, 2025

• [Feature] DIA TTS model finetuning support

  #2410 opened Apr 25, 2025

• [Feature] TRL 0.17 support

  #2409 opened Apr 25, 2025

• "RuntimeError: CUDA driver error: unknown error" when Fine Tuning Llama-3.2-11B-Vision-Instruct

  #2408 opened Apr 25, 2025

• Unsloth models output gibberish on LONG inputs

  #2405 opened Apr 24, 2025

• [Question] Do not see 2x speed finetuning Qwen2.5-VL model

  #2404 opened Apr 24, 2025

• Jetson finetune load model out of memory

  #2401 opened Apr 24, 2025

• [Bug] Impossible to convert Gemma3 4b into gguf

  #2399 opened Apr 23, 2025

42 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• vLLM Windows CUDA support [tested]

  #2158 commented on Apr 28, 2025 • 1 new comment

• Error saving GGUF of vision model

  #1504 commented on Apr 27, 2025 • 0 new comments

• Error, when loading llama3: `AttributeError: module 'transformers.models.bit.modeling_bit' has no attribute 'Linear'`

  #2191 commented on Apr 27, 2025 • 0 new comments

• BackendCompilerFailed: backend='inductor' raised: SystemError: PY_SSIZE_T_CLEAN macro must be defined for '#' formats

  #2230 commented on Apr 28, 2025 • 0 new comments

• [Bug] Issue while evaluating on GRPO.

  #2367 commented on Apr 28, 2025 • 0 new comments

• FakeTensor Error in GRPO training

  #2079 commented on Apr 28, 2025 • 0 new comments

• [Question] Cannot install specific releases from source ?

  #2368 commented on Apr 28, 2025 • 0 new comments

• Unable to load unsloth trained model saved to a local directory.

  #934 commented on Apr 28, 2025 • 0 new comments

• [Question] Fine-tune Gemma3 OOM

  #2366 commented on Apr 28, 2025 • 0 new comments

• RuntimeError: Unsloth: Quantization failed! You might have to compile llama.cpp yourself, then run this again.

  #1781 commented on Apr 29, 2025 • 0 new comments

• Error while importing unsloth in databricks

  #1294 commented on Apr 29, 2025 • 0 new comments

• Add STT model for fine tune

  #2394 commented on Apr 29, 2025 • 0 new comments

• Unsloth currently does not support multi GPU setups in unsloth-2024.8

  #859 commented on Apr 29, 2025 • 0 new comments

• [Bug] unsloth_compiled_module_mamba2.py IndentationError: unexpected indent

  #2347 commented on Apr 29, 2025 • 0 new comments

• [BUG] Image features and image tokens do not match on full finetuning

  #2251 commented on Apr 29, 2025 • 0 new comments

• Maybe bug ? Cannot retrain fine tuning output from gemma3-4b

  #2304 commented on Apr 29, 2025 • 0 new comments

• 'unsloth/llava-v1.6-mistral-7b-hf' model inference ValueError: Image features and image tokens do not match: tokens: 1175, features 1176

  #2225 commented on Apr 29, 2025 • 0 new comments

• Unsloth On Mac

  #685 commented on Apr 30, 2025 • 0 new comments

• CUDA error: out of memory in WSL with 24G VRAM while 2/3 was still left unused

  #1797 commented on Apr 30, 2025 • 0 new comments

• OOM on WSL, GRPOTrainer RuntimeError: CUDA driver error: out of memory

  #1744 commented on Apr 30, 2025 • 0 new comments

• Added Support for Apple Silicon

  #1289 commented on Apr 28, 2025 • 0 new comments

• How to create or get the ollama modelFile of Unsloth tube square fine-tuning model?

  #1823 commented on Apr 23, 2025 • 0 new comments

• Is there a training method for GRPO using Qwen2.5-VL-3B-Instruct?

  #2324 commented on Apr 23, 2025 • 0 new comments

• [BUG] Gemma3 vision throws error: TypeError: 'int' object is not iterable

  #2258 commented on Apr 23, 2025 • 0 new comments

• [RunPod] - llama_model_load: error loading model: error loading model vocabulary: cannot find tokenizer merges in model file

  #1947 commented on Apr 24, 2025 • 0 new comments

• Gemma3ForCausalLM_forward() got multiple values for argument 'logits_to_keep'

  #2167 commented on Apr 24, 2025 • 0 new comments

• Improve Documentation on Loading Fine-Tuned Models with Added Tokens

  #1863 commented on Apr 24, 2025 • 0 new comments

• [Feature] Qwen 2.5-Omni Support?

  #2325 commented on Apr 24, 2025 • 0 new comments

• [FEAT] add Quantization Aware Training (QAT) support

  #2311 commented on Apr 25, 2025 • 0 new comments

• [Bug] AttributeError: 'LlamaForCausalLM' object has no attribute 'vllm_engine'

  #2384 commented on Apr 26, 2025 • 0 new comments

• [Bug] When use customized trl.trainer, there is a sharp increase in CUDA memory?

  #2397 commented on Apr 26, 2025 • 0 new comments

• [Bug]Encountered a RuntimeError: CUDA error: a memory access was encountered during training with Unsloth’s GRPOTrainer

  #2387 commented on Apr 26, 2025 • 0 new comments

• [Question] Unexpectable warnings during unsloth setup

  #2391 commented on Apr 26, 2025 • 0 new comments

• save_pretrained_merged Issue

  #2159 commented on Apr 26, 2025 • 0 new comments

• [Feature] Is it possible to support to train microsoft/bitnet-b1.58-2B-4T ?

  #2390 commented on Apr 26, 2025 • 0 new comments

• Comprehensive Report: 3-Day Installation Struggle on Windows 10/WSL Following All Official Methods

  #2395 commented on Apr 26, 2025 • 0 new comments

• DAPO Implementation

  #2141 commented on Apr 26, 2025 • 0 new comments

• [Bug]Error loading model on Colab A100

  #2380 commented on Apr 27, 2025 • 0 new comments

• [Bug] TinyLlama Finetune not learning

  #2385 commented on Apr 27, 2025 • 0 new comments

• [Question]After fine-tuning, the model performed well, but after merging into GGUF (without quantization), its performance dropped. Why might this happen?

  #2374 commented on Apr 27, 2025 • 0 new comments

• raise RuntimeError("mmap can only be used with files saved with "

  #570 commented on Apr 27, 2025 • 0 new comments

• [Feature] Can support THUDM/GLM-Z1-9B-0414, thanks

  #2398 commented on Apr 27, 2025 • 0 new comments