[RT-DETRv2] MPS crash: build_2d_sinusoidal_position_embedding hardcodes torch.float64, breaking Apple Silicon / MPS inference

[shubhammr21]

System Info

- `transformers` version: 5.9.0
- Platform: macOS (Apple Silicon, MPS backend)
- Python version: 3.13
- PyTorch version: (MPS-enabled build)
- `docling` version: 2.95.0
- `docling-ibm-models` version: 3.13.2

Reproduction

Run RTDetrV2ForObjectDetection inference on any Apple Silicon Mac (MPS device). Triggered in practice via docling → docling-ibm-models → transformers, but reproducible with:

import torch
from transformers import RTDetrV2ForObjectDetection, RTDetrImageProcessor
from PIL import Image
import requests

device = torch.device("mps")
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

processor = RTDetrImageProcessor.from_pretrained("PekingU/rtdetr_v2_r18vd")
model = RTDetrV2ForObjectDetection.from_pretrained("PekingU/rtdetr_v2_r18vd").to(device)

inputs = processor(images=image, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.no_grad():
    outputs = model(**inputs)  # crashes here

Error:

TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

Traceback points to modeling_rt_detr_v2.py line 988, inside build_2d_sinusoidal_position_embedding:

omega = torch.arange(pos_dim, dtype=torch.float64, device=device) / pos_dim

Expected behavior

Inference should succeed on MPS devices.

Root cause: build_2d_sinusoidal_position_embedding already accepts a dtype parameter (defaulting to torch.float32), and the caller at line 1077 correctly passes dtype=hidden_states.dtype. However, all internal tensor allocations hardcode torch.float64, ignoring the parameter entirely.

Note: the same function body is the canonical source in modeling_vit_mae.py and gets inlined into the generated modeling_rt_detr.py and modeling_rt_detr_v2.py, so all three files (or the source + a regeneration) would need updating.

Verified locally on Apple Silicon (MPS, transformers 5.9.0): crash without the fix, logits shape: torch.Size([1, 300, 80]) with it.

[shubhammr21]

Suggested fix

The function build_2d_sinusoidal_position_embedding already accepts a dtype parameter (defaulting to torch.float32), but all internal tensor allocations still hardcode torch.float64, bypassing it entirely. The fix is to replace every torch.float64 inside the function body with dtype:

 def build_2d_sinusoidal_position_embedding(..., dtype: torch.dtype = torch.float32) -> torch.Tensor:
     ...
-        dtype: Output dtype; frequency arithmetic uses float64 internally.
+        dtype: Target dtype for all tensors.
     ...
-    omega  = torch.arange(pos_dim, dtype=torch.float64, device=device) / pos_dim
+    omega  = torch.arange(pos_dim, dtype=dtype,         device=device) / pos_dim
-    grid_h = torch.arange(height,  dtype=torch.float64, device=device)
+    grid_h = torch.arange(height,  dtype=dtype,         device=device)
-    grid_w = torch.arange(width,   dtype=torch.float64, device=device)
+    grid_w = torch.arange(width,   dtype=dtype,         device=device)
     ...
-        pos_embed = torch.cat([torch.zeros(1, embed_dim, dtype=torch.float64, device=device), pos_embed], dim=0)
+        pos_embed = torch.cat([torch.zeros(1, embed_dim, dtype=dtype,         device=device), pos_embed], dim=0)

The same function body appears verbatim in the generated modeling_rt_detr.py (imported from vit_mae) and modeling_rt_detr_v2.py, so all three files would need updating (or the generated ones regenerated after fixing the source in modeling_vit_mae.py).

Happy to open a PR if this direction looks right to you.

[Sriniketh24]

Hi @shubhammr21, thanks for the detailed analysis! I'd like to take this if you haven't started on a PR yet. The fix is straightforward — replace all hardcoded torch.float64 with the dtype parameter in build_2d_sinusoidal_position_embedding. Happy to step aside if you're already working on it.

[shubhammr21]

Thanks @Sriniketh24! Going with a slightly different approach than what I sketched above — PR is up at #46174.