Benson Ma

Results 80 comments of Benson Ma

Latest FBGEMM doesn't build with latest PyTorch

Closing, as the issue appears to be resolved. Please feel free to open a new ticket if the issue is observed again.

[Question] What does device / embedding_specs.compute_device parameter in ctor of TBE mean?

Hi @JacoCheung `ComputeDevice` specifies the TBE kernel variant (i.e. will the kernel execute on the CPU or CUDA) that will be applied to each embedding table. `EmbeddingLocation` specifies the target...

[Question] What does device / embedding_specs.compute_device parameter in ctor of TBE mean?

@JacoCheung If `ComputeDevice==CPU`, then `EmbeddingLocation` must be `CPU `; see https://github.com/pytorch/FBGEMM/blob/main/fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_training.py#L422

How to install this on Windows x64

Hi @Elllllllvin unfortunately FBGEMM_GPU doesn't support Windows at the moment, as there is Linux-dependent code in parts of the codebase. It may be possible to install FBGEMM_GPU on Windows through...

version compatiblity with import lib

Hi @jerryzh168 this looks strange, how are you installing fbgemm_gpu?

File Limit Request: fbgemm-gpu-nightly - 1500 MiB

Replied in https://github.com/pypi/support/issues/5902, please see my comments

fbgemm_gpu doesn't build for CPU because impl_abstract_pystub is not found

Hi @gshimansky, thanks for pointing out the bug. This is being addressed in https://github.com/pytorch/FBGEMM/pull/2363

add ckpt and restore with feature evict metaheader

Hi @lalala-2 I believe @bobbyliujb is working on this over at https://github.com/pytorch/FBGEMM/pull/4429

[ROCm] Building from source for ROCm results in fbgemm_gpu_config.so: cannot open shared object file: No such file or directory

Hi @nikhil-tensorwave Starting from 1.2.0, FBGEMM_GPU ROCm is now properly distributed as a variant on download.pytorch.org: ``` pip install torch --index-url https://download.pytorch.org/whl/rocm6.3/ pip install fbgemm-gpu --index-url https://download.pytorch.org/whl/rocm6.3/ ```

Remove unused AVX{2,512}_FLAGS

@cyyever the tests are failing, but we have fixed a number of CI issues over the last 3 days. Could you rebase and try again?