GitLab

* Add generic stripe_height support for load balancing
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix imports in test for deprecated jax.experimental.pjit
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add test case for stripe_height greater than 1. Add stripe_height arg to reordering methods
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add Striped 1 and 4 test cases. Refactor the Load Balancing test case. Fix the incorrect shape in striping inverser reordering
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Modify test code for CP + AG + THD + stripe height greater than 1
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add stripe_height arg to fused attn and fused attn fwd API. Add appropriate mask checks for AG+THD+CP and pick BRCM to be executed per rank. Add Fused Attn Primitive for CP + THD +AG + Striping. Add a method to reorder and all gather segment ids and offsets for kv
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* TMP: Throwaway testing commit
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add comments in primitive registration process
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* TMP: Throwaway test commit
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Undoing incorrect rebase/merge leftovers
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* TMP: Throwaway test commits
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Add support for calculating q and kv seqlens and offsets per rank for CP+THD+AG+SW+Striped>1 primitive
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Augment jax primitive register code comments
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>

* Fix the array sizes and padding values returned for seqlens and offsets to fit what the fused attn primitive non cp computation
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add support in new primitive for softmax_offset related changes. Put in missing primitive registering line in again. Increase the seqoffsets arrays lengths by 1
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Add new set of helper functions for seqlens and seqoffsets fo AG+THD+CP+Stripe>1 which accounts for batching and seq offsets size b+1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add backward primitive for CP+THD+AG+Striped>1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Modify tests for backward primitive for CP+THD+AG+Striped>1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Move stripe_height along with other static args in fused_attn_bwd rule. Fix typo in CP+AG+TH+Striped>1 primitive
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Code clean up: remove older version for calculating seqlens and offsets for CP+AG+THD+striped>1 primitive
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Add test for CP+THD+AG+Striped>1
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix missing var
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Add SWA tests for AG+Striped>1+CP+THD+SWA
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Restoring test code
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Remove assert preventing SWA code path in CP+AG+Striped primitive
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Parametrize num_segments_per_seq in tests
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Clean up test code
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Clean up test code in TE common
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Clean up debug statements
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Rename stripe_height to stripe_size
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Code clean up and add additional comments
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

nit: Apply suggestions from code review
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Kshitij Lakhani <33047503+KshitijLakhani@users.noreply.github.com>

Fix type on fused attn tests
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Fix seqoffsets length to be passed onto FusedAttn primitive as it is b and not b+1 needed by cuDNN
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Remove commented code
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Signed-off-by: Kshitij Lakhani <33047503+KshitijLakhani@users.noreply.github.com>

Fix linting issues
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

Fix incorrect greptile change
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Skip THD test cases for CP + AG + Dual chunk. Skip BSHD cases for CP + AG + Striped>1. Correct the layout and shapr parameters passed to the tests
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Pass stripe_size explicitly for ring attn tests for THD cases
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Remove TODO
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* Explicitly fail if THD + AG is being used with a non padding causal mask
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* nit: Correct the ID for the test dist fused attn tests to account for cp*2 which is done under the hood
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Set num_segments_per_seq defaults to None instead of 0
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Augment comments. Add ValueError for stripe_size=0
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Test only 1 num_segments_per_seq combination for CP+AG+THD+Striped>1+SWA instead of 2. Modify the num segments and window size to easily to debug values
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Default stripe_size to None instead of 0. Modify stripe_size check for <=0 instead of ==0
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Remove incorrectly added file
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Explicitly pass zero sized arrays for seg ids and pos in the CP + AG + Striped primitive rather than using the seqlens or the offsets as placeholders
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* Fix linting errors
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* Add a deep dive doc for CP+THD+AG+Stripe>1+SWA regarding design considerations and decisions
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Put docs and pngs into it's separate dir
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Replace png screenshots with markdown coe blocks for the attention patterns. Remove unecessary pngs
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Add doc file to index.rst. Fix grammatical errors
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

---------
Signed-off-by: Kshitij Lakhani <klakhani@nvidia.com>
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>
Signed-off-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>
Co-authored-by: Kshitij  Janardan Lakhani <klakhani@login-eos01.eos.clusters.nvidia.com>
Co-authored-by: Kshitij  Janardan Lakhani <klakhani@login-preos01.a51.clusters.nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>