[BUG] SadTalker incompatible with PyTorch versions required for NVIDIA RTX 50xx (Blackwell) GPUs – C++ ops failing + keypoint extraction errors

[manucabrio]

Hello,
I am reporting this issue because SadTalker currently cannot run in GPU mode on NVIDIA RTX 50xx (Blackwell) laptops, due to incompatibilities between SadTalker’s dependencies (torchvision / facelib / retinaface) and the only versions of PyTorch that support this new architecture.

This is not a configuration error:

• The issue comes from deprecated or removed torchvision C++ ops required by SadTalker.
• These ops do not exist anymore in current PyTorch builds compatible with RTX 5080 / CUDA 12.8.

Below is a complete technical report.

---

1. System Information

---

Laptop: MSI Vector 16 HX AI
GPU: NVIDIA GeForce RTX 5080 Laptop GPU (Blackwell architecture)
CPU: Intel Ultra 9
RAM: 32 GB
Storage: 1 TB NVMe SSD
OS: Windows 11 64-bit
Python: 3.10.11
CUDA requirement: RTX 5080 requires CUDA >= 12.8

---

2. PyTorch Requirements for RTX 50xx

---

Stable PyTorch releases do not support the RTX 50xx series yet.

The only PyTorch builds that successfully detect and use the RTX 5080 GPU are NIGHTLY wheels:

torch==2.10.0.dev (CUDA 12.8)
torchvision==0.25.0.dev
torchaudio==2.10.0.dev

Running:

python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

returns:

CUDA dispo = True
Nombre de GPU = 1
Nom GPU = NVIDIA GeForce RTX 5080 Laptop GPU

Meaning the GPU is correctly detected with nightly PyTorch.

---

3. Why SadTalker Breaks on RTX 50xx

---

SadTalker internally relies on:

• old torchvision.ops APIs
• legacy C++ ops (NMS, bbox ops)
• RetinaFace + facelib layers compiled against older ABI
• functionality removed in torchvision >= 0.25

These APIs no longer exist in current nightly builds of PyTorch.

Attempting to downgrade torchvision to a SadTalker-compatible version (0.15.x) fails because:

• its C++ ops cannot load under torch nightly
• ABI is incompatible
• CUDA versions mismatch

---

4. Steps to Reproduce the Issue

---

1. Create a new venv:

   python -m venv venv
   venv\Scripts\activate

2. Install PyTorch nightly:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

3. Install SadTalker dependencies:

   pip install -r requirements.txt

4. Run SadTalker:

   python app_sadtalker.py

The UI loads normally.

5. Attempt to generate an animation → SadTalker crashes.

---

5. Error Logs

---

ERROR A — C++ ops cannot load:

RuntimeError: Couldn't load custom C++ ops.
This can happen if your PyTorch and torchvision versions are incompatible,
or if you had errors while compiling torchvision from source.

Traceback:

torchvision/ops/boxes.py
torchvision/extension.py
assert_has_ops()

This indicates that torchvision.ops.nms and related ops are not available anymore.

---

ERROR B — Keypoint extraction fails:

UnboundLocalError: local variable 'keypoints' referenced before assignment

Traceback:

extract_kp_videos_safe.py → extract_keypoint()
retinaface → detect_faces()
py_cpu_nms()
torchvision.ops.nms()

Since torchvision.ops.nms is removed from nightly builds, the detection pipeline crashes and keypoints never initialize.

---

6. What Has Already Been Attempted

---

• Installing old torchvision:
  pip install torchvision==0.15.2 --no-deps
  → fails (C++ extensions cannot load)

• Using latest nightly torchvision (0.25.0.dev)
  → breaks SadTalker: missing ops, incompatible APIs

• Testing all pre-processing modes
  → same issue (NMS fails)

• Testing multiple images
  → same issue

• CPU mode
  → works perfectly

---

7. Diagnosis (Root Cause)

---

SadTalker currently depends on:

• deprecated torchvision.ops APIs
• C++ ops built for CUDA 11.x
• RetinaFace detection relying on outdated interfaces

But RTX 5080 requires:

• PyTorch >= 2.10 nightly
• CUDA >= 12.8
• torchvision >= 0.25.x

These environments are not compatible.

Thus:
→ SadTalker cannot currently run in GPU mode on RTX 50xx GPUs.

---

8. Requested Feedback / Possible Solutions

---

Option A — Update SadTalker’s face detection pipeline

• Replace torchvision.ops.nms with a pure Python implementation
• Update RetinaFace / facelib to modern torch APIs
• Remove dependency on torchvision C++ ops

Option B — Provide a “torchvision-free” mode
Useful for environments where torchvision cannot load C++ ops.

Option C — Publish compatibility matrix
To inform users which PyTorch versions are expected to work with SadTalker.

Option D — Provide updated wheels
SadTalker-specific builds for CUDA 12.8 / Blackwell GPUs.

---

9. Conclusion

---

SadTalker works flawlessly in CPU mode, but GPU mode is currently impossible on NVIDIA RTX 50xx (Blackwell) GPUs because:

• PyTorch stable does not support these GPUs
• PyTorch nightly removes the legacy ops SadTalker depends on
• torchvision incompatible C++ ops break the detection pipeline

I am happy to test any patches, branches, or experimental wheels.

Thank you for maintaining this excellent project.

[johndpope]

try this new repo IMTalker - my branch has tensorrt / 5090 support + 58 fps
https://github.com/johndpope/IMTalker