- Did you update?
pip install --upgrade unsloth unsloth_zoo
- Yes, latest version installed
Colab or Kaggle or local / cloud
- Number GPUs used, use
nvidia-smi
- 1 GPU, RTX 4080 with 16GB VRAM.
- Which notebook? Please link!
- Which Unsloth version, TRL version, transformers version, PyTorch version?
- unsloth: 2026.4.8, TRL: 0.24.0, transformers: 5.5.0, PyTorch: 2.10.0+cu130
- Which trainer?
SFTTrainer, GRPOTrainer etc
- GRPOTrainer with Qwen3-VL-8B on a multimodal ChartQA dataset
Description
Getting a shape mismatch in get_rope_index during GRPO's compute_loss forward pass with Qwen3-VL-8B on a multimodal dataset. The mismatch size varies per run (e.g. [3, 447] vs [3, 445], or [3, 454] vs [3, 447]), so it's not a fixed offset - it depends on batch contents. Crashes at the very first training step.
Minimal Code to Reproduce Error
from unsloth import FastVisionModel
from trl import GRPOTrainer, GRPOConfig
from datasets import load_dataset
import re
model, tokenizer = FastVisionModel.from_pretrained(
model_name="unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit",
max_seq_length=16384,
load_in_4bit=True,
)
def process(example):
image = example["image"].resize((512, 512)).convert("RGB")
prompt = [{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": example["query"].strip()},
],
}]
return {"prompt": prompt, "image": image, "answer": example["label"][0]}
dataset = load_dataset("HuggingFaceM4/ChartQA", split="train[:500]").map(process)
dataset = dataset.select_columns(["prompt", "image", "answer"])
def reward_func(completions, answer, **kwargs):
scores = []
for completion, ans in zip(completions, answer):
if isinstance(completion, list):
completion = completion[0]["content"] if completion else ""
match = re.search(r"<SOLUTION>(.*?)</SOLUTION>", completion, re.DOTALL)
predicted = match.group(1).strip() if match else ""
scores.append(1.0 if predicted == ans.strip() else 0.0)
return scores
training_args = GRPOConfig(
per_device_train_batch_size=4,
num_generations=4,
max_prompt_length=1024,
max_completion_length=1024,
num_train_epochs=1,
output_dir="outputs",
bf16=True,
)
training_args.unsloth_num_chunks = -1
training_args.unsloth_grpo_mini_batch = None
training_args.unsloth_logit_chunk_multiplier = None
training_args.vllm_sampling_params = None
trainer = GRPOTrainer(
model=model,
args=training_args,
processing_class=tokenizer,
reward_funcs=[reward_func],
train_dataset=dataset,
)
trainer.train() # crashes here Traceback
File "transformers/models/qwen3_vl/modeling_qwen3_vl.py", line 1117, in get_rope_index
position_ids[:, batch_idx, attention_mask[batch_idx].bool()] = llm_positions.to(position_ids.device)
RuntimeError: shape mismatch: value tensor of shape [3, 454] cannot be broadcast to indexing result of shape [3, 447]
pip install --upgrade unsloth unsloth_zooColaborKaggleor local / cloudnvidia-smiSFTTrainer,GRPOTraineretcDescription
Getting a shape mismatch in
get_rope_indexduring GRPO'scompute_lossforward pass with Qwen3-VL-8B on a multimodal dataset. The mismatch size varies per run (e.g.[3, 447]vs[3, 445], or[3, 454]vs[3, 447]), so it's not a fixed offset - it depends on batch contents. Crashes at the very first training step.Minimal Code to Reproduce Error
Traceback