1. Industry AI Model Platform / 
  2. v1.x / 
  3. Resource Management / 
  4. Model Fine-tuning / 
  5. Model Fine-tuning FAQ

Documentation

Last updated:
    Edit

    Model Fine-tuning FAQ

    FAQ

    The fine-tune button is grayed out with message "Fine-tuning framework does not support this model"

    Cause: The current fine-tuning framework (LLaMA-Factory or MS-Swift) does not yet support this model architecture.

    Solution: Contact the platform administrator with the model name and relevant details. The administrator will evaluate and add support as soon as possible.

    The fine-tune button is grayed out with message "Model metadata not recognized"

    Cause: The model files are incomplete, or the model's configuration information (architecture info in config.json) cannot be automatically recognized.

    Solution:

    1. Confirm the model repository contains a complete config.json file.
    2. Contact the platform administrator to manually trigger a model metadata scan.

    Training loss is not decreasing

    Possible Causes:

    • Learning rate is too high or too low.
    • Dataset format is incorrect, causing samples to be skipped.
    • Training data volume is too small.

    Solution:

    1. Check that the dataset format matches the selected framework's requirements.
    2. Adjust the learning rate (starting from 1e-4 is recommended).
    3. Ensure sufficient training samples (at least a few hundred is recommended).

    Details: The create page estimates weight VRAM from parameter count and precision, then multiplies by 2.0 (KV cache plus optimizer and runtime overhead). If the scanned LoRA estimate is larger, the larger value is used. That recommendation is compared with the GPU Memory (GB) advertised on each specification; cards below it are marked Low Memory.

    Fine-tuning needs more headroom than inference. A ~27.78B BF16 model has about 55.56 GB of weights, so the fine-tune recommendation is about 111 GB (about 83 GB for inference).

    Older metadata for some models may still store 1 GB (the scanner missed nested text_config). The page now recalculates from parameter count; an administrator can also rescan the model.

    Out of memory (OOM) during training

    Solution:

    1. Prefer a specification at or above Recommended Minimum GPU Memory. Cards marked Low Memory are more likely to OOM.
    2. Switch to a compute configuration with more VRAM.
    3. Enable quantized fine-tuning in the framework settings (e.g., QLoRA with 4-bit quantization).
    4. Reduce the batch size.
    5. Use gradient accumulation to achieve an equivalent larger batch size.