Torch not compiled with CUDA enabled

[sumitsharansatsangi]

I am running it on mac with following specification

Chipset Model: Apple M2
Type: GPU 
Bus: Built-In 
Total Number of Cores: 8 
Vendor: Apple (0x106b) 
Metal Support: Metal 4

My output

When using '--method hf', ensure that the batch size is set correctly.  We will default to batch size of 1.
Chandra CLI - Starting OCR processing
Input: input.pdf
Output: output
Method: hf

Loading model with method 'hf'...
Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:17<00:00,  4.48s/it]
Some parameters are on the meta device because they were offloaded to the disk.
Model loaded successfully.

Found 1 file(s) to process.

[1/1] Processing: input.pdf
  Loaded 1 page(s)
  Processing pages 1-1...
  Error processing input.pdf: Torch not compiled with CUDA enabled

Processing complete. Results saved to: output

My command was

chandra  input.pdf ./output --method hf  --page-range 1 

[fatemeh-mohseni-AI]

The error “Torch not compiled with CUDA enabled” means your current PyTorch installation does not support GPU acceleration — it was built for CPU-only use.

run :

import torch
print(torch.__version__)
print(torch.version.cuda)
print(torch.cuda.is_available())

If it prints:

torch.version.cuda → None

torch.cuda.is_available() → False

then you’re using a CPU-only build.

1. Find your CUDA version
2. Install a PyTorch build that matches your CUDA version
3. Verify GPU is working

import torch
print(torch.cuda.is_available())  # should be True
print(torch.cuda.get_device_name(0))

[omaratef3221]

Hi,

Does it support Apple MAC MPS Torch ?

[joedanields]

Hi,

Does it support Apple MAC MPS Torch ?

Hey 👋

Yes, Chandra should work on Apple Silicon (M1/M2/M3) systems using PyTorch’s MPS (Metal Performance Shaders) backend.

The error you’re seeing —

Torch not compiled with CUDA enabled

just means that your current PyTorch build is trying to use CUDA (which is only for NVIDIA GPUs). On macOS, GPU acceleration works through MPS, not CUDA.

You can check if MPS is available on your system by running:

import torch
print(torch.backends.mps.is_available())  # should be True
print(torch.backends.mps.is_built())      # should be True

If both return True, Chandra will automatically use your M2 GPU through Metal.
If not, make sure you have a recent PyTorch version (≥ 1.12) installed:

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

(PyTorch will enable MPS automatically when running on Apple Silicon.)

[omaratef3221]

Even though both prints return true. When I run through CLI using hf as method I get this:

Found 1 file(s) to process.

[1/1] Processing: RETREAT PALM DUBAI.pdf
  Loaded 1 page(s)
  Processing pages 1-1...
  Error processing RETREAT PALM DUBAI.pdf: Torch not compiled with CUDA enabled

Processing complete. Results saved to: output

and file is not saved and not generated

[FelipeBerrios]

Hi, perhaps it's because the hf.py file is forcing the use of "cuda" for input.

inputs = inputs.to("cuda")

[joedanields]

Hi, perhaps it's because the hf.py file is forcing the use of "cuda" for input.

inputs = inputs.to("cuda")

maybe we can try

[tabtablabs-dev]

working on my m4 max now with #21

[gasparschott]

Worked for me, too, on my M1 MBP.
I just changed inputs = inputs.to("cuda") to inputs = inputs.to("cpu").