viiccwen opened a new issue, #1004:
URL: https://github.com/apache/mahout/issues/1004

   ### Summary
   Extend QDP to support zero-copy basis encoding from GPU pointers when the 
indices are provided as float32 (e.g. PyTorch `torch.float32` CUDA tensor of 
basis indices).
   
   - **Amplitude encoding** has zero-copy float32 GPU APIs 
(`encode_from_gpu_ptr_f32`, `encode_from_gpu_ptr_f32_with_stream`); the Python 
bindings use them for float32 CUDA tensors.
   - **Basis encoding** today uses the generic `encode_from_gpu_ptr` / 
`encode_batch_from_gpu_ptr` path, which expects the pointer to point to 
**integer** data (e.g. `usize` / int64) on the device. When the user passes a 
float32 CUDA tensor (e.g. indices stored as `float32` for consistency with 
other model tensors), the API does not support it and may require `.cpu()` or a 
cast, losing zero-copy.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to