TypeError "Got unsupported ScalarType BFloat16" in validation when training with amp_dtype="bfloat16" #940

What happened

Training GTR-X-seg (task="segment") with amp=True, amp_dtype="bfloat16". Training steps run fine, but validation fails after every epoch, so no metrics are logged and best.pt is never selected. The validator runs under bf16 autocast, and postprocess/ec.py calls .cpu().numpy() on bf16 tensors, which NumPy does not support. postprocess and postprocess_pose in the same file have the same pattern. fp16 and fp32 are not affected.

Reproduce

import torch
from libreyolo.postprocess.ec import postprocess, postprocess_seg

out = {
    "pred_logits": torch.randn(1, 300, 1).bfloat16(),
    "pred_boxes": torch.rand(1, 300, 4).bfloat16(),
    "pred_masks": torch.randn(1, 300, 160, 160).bfloat16(),
}
postprocess(out, conf_thres=0.0, original_size=(640, 640))
postprocess_seg(out, conf_thres=0.0, original_size=(640, 640))

End to end: LibreYOLO("LibreGTRx-seg.pt").val(data=..., half=True, amp_dtype="bfloat16")

Traceback / output

ERROR | Validation failed: Got unsupported ScalarType BFloat16

  File "libreyolo/models/gtr/model.py", line 395, in _postprocess
    return postprocess_seg(
  File "libreyolo/postprocess/ec.py", line 82, in postprocess_seg
    "scores": scores.cpu().numpy()
TypeError: Got unsupported ScalarType BFloat16

Environment

libreyolo 1.6.0 (same code on dev e11cae7), Python 3.14, torch 2.12.0+cu126, Linux, CUDA
Custom dataset, ~26k images, 1 class, imgsz 640, batch 4

Already tried

• val with half=False works
• fp16 autocast works
• calling .float() on the bf16 tensors before .numpy() makes the snippet above pass