search

ai-forever / ru-gpts

Russian GPT3 models.
Apache License 2.0
2.08k stars 445 forks source link

CUDA: Error- invalid ptx #68

Closed AlexanderKozhevin closed 2 years ago

AlexanderKozhevin commented 3 years ago
Screenshot 2021-07-12 at 19 08 35
Artyrm commented 3 years ago

I guess maybe it is connected with --fp16 use: https://github.com/openai/triton/issues/187#issuecomment-893563590 But I can't figure out still how to disable it in this example.

Artyrm commented 3 years ago

Yup. If you're running ruGPT3XL_generate.ipynb just try to comment out this string in ru_gpts/src/xl_wrapper.py:

    # Fp16 conversion.
    model = FP16_Module(model)

    return model

Like that:

    # Fp16 conversion.
    # model = FP16_Module(model)

    return model

Worked for me.

From my understanding it is because of incompatibility of FP16 with some types of cards, to say Tesla P100 in my case.

UPD looks like I'm wrong - since fp16 work for me with Transformers. DK why it helps.

king-menin commented 2 years ago

Add fixes for last updates on colab for rugpt3xl