Khanin/merged pool emb opt #4977
Conversation
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
| with torch.cuda.device(dst_device): | ||
| inputs = [torch.randn(10, 20) for _ in range(num_inputs)] | ||
| pitch = True | ||
| if pitch: |
There was a problem hiding this comment.
@kudomcho This logic is strange, we set pitch and immediately check its value..
There was a problem hiding this comment.
@q10 This is to check the assertion allclose in case pitching is enabled. Currently it forces to True to test the all to one for the pitch condition. Any preference on passing the pitch condition on the test argument?
There was a problem hiding this comment.
@q10 This is to check the assertion allclose in case pitching is enabled. Currently it forces to True to test the all to one for the pitch condition. Any preference on passing the pitch condition on the test argument?
Yes, could you make this an argument to the test method, and use hypothesis @given(...) to pass the selection in?
There was a problem hiding this comment.
I added use_pitched on the given at test_all_to_one_device and moved the modified the merged_embeddings_benchmark.py to do pitch calculation by adding --use_pitched.
kudomcho
force-pushed
the
khanin/merged_pool_emb_opt
branch
from
bd261b7 to
3a36197
Compare
…h calculation for better memory alignment on merged_embeddings_benchmark.py
Summary: X-link: facebookresearch/FBGEMM#1998 implemented pitch size on tensor allocation for better memory alignment. Differential Revision: D83994225 Pulled By: q10
implemented pitch size on tensor allocation for better memory alignment.