1. Did you update? pip install --upgrade unsloth unsloth_zoo
   yes (unsloth updated to 2025.7.8)

2. Colab or Kaggle or local / cloud
   local Python script
   OS: Linux

3. Number GPUs used, use nvidia-smi
   GPU(s): 1× NVIDIA RTX A6000 (compute 8.6, 47.44 GB VRAM)5. Number GPUs used, use nvidia-smi

4. Which notebook? Please link!
   Local

5. Which Unsloth version, TRL version, transformers version, PyTorch version?
   Unsloth: 2025.7.8
   TRL (Transformers Reinforcement Learning): 0.19.1
   Transformers: 4.53.2
   PyTorch: 2.7.0+cu126 (CUDA 12.6)

6. Which trainer? SFTTrainer, GRPOTrainer etc```pythonPut Minimal code to reproduce error here ###Remove Hugging Face Using GRPOTrainer from trl, in a plain Python script (train_deepseek_grpo_custom-dataset_v2_long_context.py).

Configs:

Script

script: train_deepseek_grpo_custom-dataset_v2.py

Model & data

model_name: "unsloth/DeepSeek-R1-0528-Qwen3-8B"
revision: "main"
dataset_name: "custom"
train_split: "train"
test_split: "test"

Output & logging

output_dir: "deepseek_r1_grpo"
wandb_project: "deepseek_qwen3_avy"
wandb_run_name: "training_5percent_dataset_new_reward_func"

Training hyperparameters

batch_size: 2
accum_steps: 1
total_steps: 30
lr: 2e-5

Context window & LoRA

max_length: 55000
max_new_tokens: 256
max_completion_len: 256
lora_rank: 16

Error:
AssertionError: seq_len <= cos.shape[0]
File “…/unsloth/kernels/rope_embedding.py”, line 91, in forward
assert(seq_len <= cos.shape[0])