h2o4gpu
h2o4gpu copied to clipboard
h2oai
H2Oai GPU Edition
https://github.com/h2oai/h2o4gpu/commit/7a7424f7253ea650989b9b5174e8597801d95222#diff-1a13759f63589530199630e07535c1a7R89
Noticed that conda install of h2o4gpu installs lot more packages in the created environment than the corresponding pip install method. ``` (env) hemen@mr-dl7:~/Workspace/Office/h2o4gpu-issue$ pip list --local Package Version ----------------- -----------...
``` pytest tests/python/open_data/glm/ ``` `WrappedPointer.__del__` rises an error sometimes that lib.modelfree1_double is NoneType Exception ignored in: Traceback (most recent call last): File "/home/sh1ng/dev/.venv/lib/python3.6/site-packages/h2o4gpu/solvers/utils.py", line 37, in __del__
Computation like `rows / n_gpu` may ignore a few records at the end of an array. https://github.com/h2oai/h2o4gpu/blob/master/src/gpu/kmeans/kmeans_h2o4gpu.cu
- **OS platform, distribution and version (e.g. Linux Ubuntu 16.04)**: Linux Ubuntu 18.04 - **Installed from (source or binary)**: binary - **Version**: 0.2.0 - **Python version (optional)**: 3.6 - **CUDA/cuDNN...
I thought cuda9 would improve this, but seems they didn't improve this aspect. Even MAGMA package has the same issue. I understand some stuff is inefficient for the GPU, but...
https://s3.console.aws.amazon.com/s3/buckets/h2o-release/h2o4gpu/snapshots/ai/h2o/h2o4gpu/0.1-nonccl-cuda8/?region=us-east-1&tab=overview wheels are here, while docker are as in the other github issue.
The `calculate_centroids` kernel from `kmeans_centroids.h` is non-deterministic for `float`! The method basically goes through all the records and sums them up by labels, this summed records will be used to...
Like in POGS we need a ptr method for SVD. Will put this as 0.2.0 but can be pushed for later.