onnxruntime
Optimize python overhead of APEX amp
#9447
Merged

Optimize python overhead of APEX amp #9447

pengwa merged 12 commits into master from pengwa/opt_amp
pengwa
pengwa optimize python overhead of _post_amp_backward
f7bee7cd
pengwa pengwa requested a review from baijumeswani baijumeswani 4 years ago
pengwa pengwa requested a review from liqunfu liqunfu 4 years ago
pengwa pengwa requested a review from SherlockNoMad SherlockNoMad 4 years ago
pengwa pengwa requested a review from thiagocrepaldi thiagocrepaldi 4 years ago
pengwa pengwa requested a review from tlh20 tlh20 4 years ago
pengwa pengwa requested a review from xadupre xadupre 4 years ago
pengwa pengwa added training
pengwa gradually warm up the unscale kernel launches to reduce CUDA stream i…
fff81aa5
xadupre
xadupre commented on 2021-10-20
baijumeswani overwrite apex amp's zero_grad for faster implementation
e1db2615
baijumeswani baijumeswani force pushed from 1f372a38 to e1db2615 4 years ago
baijumeswani Update check_requirement functions
2001f7bb
baijumeswani
baijumeswani dismissed these changes on 2021-10-21
pengwa move unscale_fp16_grads_into_fp32_grads into C++ impl
7f9575c5
pengwa Merge branch 'pengwa/opt_amp' of https://github.com/microsoft/onnxrun…
4da10440
pengwa pengwa dismissed their stale review via 4da10440 4 years ago
pengwa refine a bit
4149d251
pengwa format
d5aeb6e9
pengwa make the heavy at::empty_like run in parallel
30e06728
pengwa improve the efficiency furthur, reducing 3.5ms to 1.7ms for unilm.
ba613669
pengwa unilm 1.7ms to 338us: 1). optimize python list <==> std::vector copy,…
0846a6c7
pengwa refine the logic a bit after validating
d97392d9
baijumeswani
baijumeswani approved these changes on 2021-10-25
baijumeswani
pengwa
pengwa pengwa merged b125446f into master 4 years ago
pengwa pengwa deleted the pengwa/opt_amp branch 4 years ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone