RecurrentGemma not compatible with autocast / AMP training

[xplip]

System Info

• transformers version: 4.40.2
• Platform: Linux-5.15.0-92-generic-x86_64-with-glibc2.31
• Python version: 3.10.13
• Huggingface_hub version: 0.20.3
• Safetensors version: 0.4.3
• Accelerate version: 0.30.1
• Accelerate config: not found
• PyTorch version (GPU?): 2.1.2+cu121 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: True
• Using distributed or parallel set-up in script?: False

Who can help?

@ArthurZucker

Information

• [X] The official example scripts
• [X] My own modified scripts

Tasks

• [X] An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• [X] My own task or dataset (give details below)

Reproduction

import torch
from transformers import RecurrentGemmaConfig, RecurrentGemmaForCausalLM

def main():
    V = 288
    B = 16
    T = 300
    device = "cuda"

    config = RecurrentGemmaConfig(
        vocab_size=V,
        num_hidden_layers=12,
        hidden_size=1024,
        num_attention_heads=8,
        intermediate_size=6144,
        attention_window_size=T,
    )
    model = RecurrentGemmaForCausalLM._from_config(config, torch_dtype=torch.float32).to(device)

    autocast_settings = [
        {"dtype": torch.float16, "enabled": True},
        {"dtype": torch.bfloat16, "enabled": True},
        {"enabled": False},
    ]

    for autocast_setting in autocast_settings:
        print(f"\nRunning with autocast setting: {autocast_setting}:")
        try:
            with torch.cuda.amp.autocast(**autocast_setting):
                outputs = model(input_ids=torch.randint(0, V, (B, T), device=device))
                print(outputs.logits.shape)
        except RuntimeError as e:
            print(e)

if __name__ == "__main__":
    main()

Expected behavior

The script should run without errors with autocast enabled, as otherwise training with AMP is not available.

Output of the script above:

Running with autocast setting: {'dtype': torch.float16, 'enabled': True}
Index put requires the source and destination dtypes match, got Float for the destination and Half for the source.

Running with autocast setting: {'dtype': torch.bfloat16, 'enabled': True}
Index put requires the source and destination dtypes match, got Float for the destination and BFloat16 for the source.

Running with autocast setting: {'enabled': False}
torch.Size([16, 300, 288])

Expected output:

Running with autocast setting: {'dtype': torch.float16, 'enabled': True}
torch.Size([16, 300, 288])

Running with autocast setting: {'dtype': torch.bfloat16, 'enabled': True}
torch.Size([16, 300, 288])

Running with autocast setting: {'enabled': False}
torch.Size([16, 300, 288])

[ArthurZucker]

Will merge the PR!