GitLab

Fuses dropout and softmax in backward pass, add bias support to CPP MHA, add additive mask support, separate Q/K/V parameters (#854)
Co-authored-by: Sukru Eryilmaz <seryilmaz@computelab-dgx1v-32.nvidia.com>

* add CUDAGenerator guard

* fix generator_flag

* add guards for gen pointer/ref issue

* change mutex_ to mutex()

* add check_generator
Co-authored-by: pbialecki <pbialecki@nvidia.com>

* disable ninja for multihead_attn

* fix getCurrentStream in multihead_attn
Co-authored-by: pbialecki <pbialecki@nvidia.com>

This reverts commit 92b3b9a9.

* NHWC support for multi tensor apply

* compilation fix for version<=1.4

* Adding C++ Multihead Attention implementation to contrib.

* Add reference test that at least works for forward.

* Remove CublasLt support from multihead attention.

* Add new Python version of self attention.

* Update python model of MHA with backward pass.

* Fixed Output Linear connection in MHA.

* Clean up compiles and add documentation to PySelfAttention.

* Add Encdec Python version of multihead attention.  Cleanup files.

* Tests for self and encdec multihead attention.

* Add reference pytorch implementation of attention with norm and add.

* Add cutlass branch definition.

* Add cutlass download to compile.

* Add norm/add tests.

* Add biases to pytorch python versions.

* Add tests and fix issues with python version of attention masking.

* Create README.md

* Update README.md

* Update README.md

* Update perf test parameters.

* Update README.md

* Update README.md

* Update README.md

* Add files via upload

* Update README.md

* Update README.md

* Update README.md

* Fix matmul1 output tensor size.  Fix tests that missed issue.

* Allow for Z dimensions of 64K and greater on batched GEMMs.

* remove redundant imports

* general cleanup, remove deprecated or unused functions

* Adding C++ Multihead Attention implementation to contrib.

* Add reference test that at least works for forward.

* Remove CublasLt support from multihead attention.

* Add new Python version of self attention.

* Update python model of MHA with backward pass.

* Fixed Output Linear connection in MHA.

* Clean up compiles and add documentation to PySelfAttention.

* Add Encdec Python version of multihead attention.  Cleanup files.

* Tests for self and encdec multihead attention.

* Add reference pytorch implementation of attention with norm and add.

* Add cutlass branch definition.

* Add cutlass download to compile.

* Add norm/add tests.

* Add biases to pytorch python versions.

* Add tests and fix issues with python version of attention masking.

* Create README.md

* Update README.md

* Update README.md

* Update perf test parameters.

* Update README.md

* Update README.md

* Update README.md

* Add f...

* add WAR for pip>=19.3.1

* remove pipmain, use extras_require instead

* move previous fused_adam and fp16_optimizer to contrib

* make build contrib.fused_adam optional

* change build option name

* remove unnecessary try import

* Pushing for build tests

* Contrib files

* Removing deprecated checks

Co-authored-by: Aditya Agrawal <aditya.iitb@gmail.com>
Co-authored-by: Marek Kolodziej <mkolod@gmail.com>

* First draft, for discussion

* Fix mistakes in LAMB equations

* Add loop over chunk

* Bug fix

* Bug fix

* Bug fix

* Undo bug fix

* Bug fix

* Add multi tensor LAMB optimizer to setup.py

* Rename step_size to learning_rate

* Fix compilation errors

* Add softmax cross entropy loss with label smoothing support.

* Fix deprecation of AT_DISPATCH_XXX and several minor issues.

* Fix issues commented by reviewers.

* Add FB license.

* Remove code generation constraints.

* Add a simple unittest for label smoothing.

* Persistent group batchnorm added

Added persistent grouped batch norm for performance run on strong scaling case:
currently only supporting:

  1. nhwc layout
  2. fp16
  3. synchronization only within a node!

Environment variable is used to tune LAUNCH_MARGIN that limits the CTAs usage
by the persistent kernel.

Documentation and examples will follow.

* updating type().scalarType() to scalar_type()

* moving launch margin to be defined at layer creation, adding a knob cap max ctas per sm

* fixing the cta computation

* review comment:

set device_id through cudaGetDevice()
move cudaMemset to cudaMemsetAsync
updated __threadfence() to __threadfence_system() inter device write

* Adding Torch + bare-metal nvcc version check and container build tests

* Putting a canary in the coalmine

* canary proved elusive

* Trying direct setup.py install

* this should work

* Removing canary

* hopefully this works

Initial implementation, all fp32
Tested against torch.optim.sgd