A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit floating point (FP8) precision on Hopper and Ada GPUs, to provide better performance with lower memory utilization in both training and inference.
This PR exposes underly THD format API to JAX side and wrap it to be supported on multi-gpus.
Fixes # (issue)
Type of change
[ ] Documentation change (change only to the documentation, either a fix or a new content)
[ ] Bug fix (non-breaking change which fixes an issue)
[x] New feature (non-breaking change which adds functionality)
[ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
[ ] Infra/Build change
[ ] Code refractor
Changes
Unify fused_attn_qkvpacked, fused_attn_kvpacked, fused_attn to be a single function
Add experimental fused_attn_thd API with the corresponding unit tests, this API requires users to configure a static max_segments_per_seq argument and ensure the "number of segments" in a seqeunce less and equal than this value
Description
This PR exposes underly THD format API to JAX side and wrap it to be supported on multi-gpus.
Fixes # (issue)
Type of change
Changes
fused_attn_qkvpacked
,fused_attn_kvpacked
,fused_attn
to be a single functionfused_attn_thd
API with the corresponding unit tests, this API requires users to configure a staticmax_segments_per_seq
argument and ensure the "number of segments" in a seqeunce less and equal than this valuetest_fused_attn
runtimeChecklist: