GitLab

Support flash attention 2 with causal masking when KV's seq length is longer than Q's seq length. (#436)

* Add RNG state to kernel launch params
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Save seed and offset for backward
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Single thread write to global mem
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* compute_dq_dk_dv_1colblock get seed and offset from launch params
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* compute_dq_dk_dv_1rowblock get seed and offset from launch params
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change forward c++ APIs to save RNG state for backward
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Change backward c++ APIs to set RNG state for bprop launcher
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bug fixes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Python side API changes
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Bug fix; only save seeds instead of full offset
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Account for 3D grid size
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

See also: https://stackoverflow.com/questions/55136414/constexpr-variable-captured-inside-lambda-loses-its-constexpr-ness

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

Follow xFormers's DISTPATCH_BOOL. Haven't tested it on Windows.

This is faster since we only need to do atomic adds on dq, instead of atomic
adds on both dk and dv.

Previously we were using block_size=256.

They don't have to have the same block size, number of threads, etc.

Theoretically this might have lower numerical error since the scaling is in
fp32 instead of fp16 (not sure, I haven't thought too carefully about it).
However, in practice, the numerical errors seem about the same.