The PaddleOCR-VL pipeline previously needed two ROCm-specific escape hatches:

* `_keep_in_fp32_modules = ["visual", "mlp_AR"]` on
  PaddleOCRVLForConditionalGeneration kept the SigLIP vision tower and the
  multimodal projector in FP32 because BF16 layer_norm and BF16 softmax were
  not registered for HIP, so running the vision encoder in BF16 crashed.
* Four `paddle.is_compiled_with_rocm()` blocks in
  `paddlex/inference/models/runners/paddle_static/runner.py` (lines 406-408,
  462-464, 496-498, 505-507) called
  `delete_pass("conv2d_add_act_fuse_pass")` and
  `delete_pass("conv2d_add_fuse_pass")` because both PIR passes rewrite
  conv2d+add[+act] into the `fused_conv2d_add_act` op, which only has a
  cuDNN GPUDNN kernel — kernel dispatch then failed on ROCm.

These are addressed at the framework level by the upstream Paddle BF16 fix
(layer_norm + softmax registration on HIP, plus gating both PIR passes on
PADDLE_WITH_CUDA so they no longer run on HIP builds). With that wheel
installed, both PaddleX workarounds become unnecessary:

* Drop `_keep_in_fp32_modules` so the vision encoder + multimodal projector
  run natively in BF16 on ROCm. End-to-end output matches the FP32-fallback
  path on PaddleOCR-VL-1.5 (validated on MI300X / gfx942 / ROCm 7.2). This
  overlaps with PaddlePaddle#5077; if PaddlePaddle#5077 lands first, the conflict is trivial.
* Drop all four `delete_pass` blocks under `paddle.is_compiled_with_rocm()`.
  Once the framework PR lands, the two passes are no longer registered on
  HIP wheels, so `delete_pass` becomes a no-op there.

Requires the framework BF16 PR to be merged and released; with older Paddle
wheels the BF16 visual path will still crash on ROCm. CUDA behavior is
unchanged — both passes remain registered under PADDLE_WITH_CUDA, and the
vision encoder simply uses whatever dtype the model is loaded with.