[ROCm] Remove BF16 workaround now that Paddle framework supports HIP BF16

[Lau-JW]

Description

Remove BF16 workarounds for ROCm now that Paddle framework supports HIP BF16 precision type.

Changes

1. paddlex/inference/utils/misc.py: Add "dcu" device to is_bfloat16_available() function

2. paddlex/inference/models/doc_vlm/modeling/paddleocr_vl/_paddleocr_vl.py: Set _keep_in_fp32_modules = None instead of ["visual", "mlp_AR"]

Motivation

Previously PaddleX had workarounds because Paddle framework did not support HIP BF16:

• is_bfloat16_available() returned False for ROCm
• Visual encoder forced to FP32 via _keep_in_fp32_modules

Now that Paddle framework supports HIP BF16 (PR: PaddlePaddle/Paddle#78665), these workarounds can be removed.

Related Issues

Fixes #5081

Checklist

• Code is formatted per PaddleX coding conventions
• Unit tests pass
• Documentation updated if needed

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[paddle-bot]

Thanks for your contribution!