GitLab

* Fused attention kernel only supports sm80 and sm90
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update transformer_engine/jax/csrc/modules.cpp
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* arbitary fused kernel supports sm86/sm89 after 8.9.3
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Skip sm70
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Forward is_fused_attn_kernel_available to cpp backend
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove cpp is_fused_attn_available API
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

* miscellenous fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add back pytorch csrc extensions.h
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add unit tests for dpa checkpointing
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove seqlen%32/64 checks for now
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix tests for core attn bias
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add tests for changes regarding rng_state in aux_ctx_tensor
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reuse rng tracker from numerics in fused attn; skip checkpointing if FAv2 in numerics
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* uncomment comments used for testing
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix pre/post scale bias
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Update transformer_engine/pytorch/attention.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* remove skipifs for FAv2 check after PR366
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove checkpointing tests for transformer layer; dpa tests still provide coverage
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* adjust random number range for tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Add upper bound to FA version
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Check backend only when using FusedAttention
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* remove imports/variables related to FAv2 checks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* further fix random number ranges for tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix variable referenced before assignment error
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix flash attention dropout probability with inference
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add output as the fused attention ctx tensor
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add rng_state as the fused attention ctx tensors
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add flash attention supported lengths to the fused attention
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Refactor attention primitive to reuse abstract shaped array
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Detect backend type to allocate appropriate ctx size
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Skip dropout correctness instead of return success
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use cudaMemsetAsync and enhance the error handling
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add flash attention kernel elts_per_thread update
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove redundant max 512 suffix
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Keep only DType and remove NVTEDType from python
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix a float32_attention_logits bugs
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Re-calculate workspace size for self attention
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance bias/dbias shape guard
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the seed/rng_state checker
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use jax.core.ShapedArray as jax.abstract_arrays is deprecated
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the unittest docs
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix bprop for cuDNN 8.9.3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Update cuDNN version requirement to 8.9.3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* debug paddle CI
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* debug paddle CI; force LD_LIBRARY
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* debug paddle CI; force LD_LIBRARY to /opt
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove debug info for paddle
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change cudnn requirement to 8.9.1 for v1 and 8.9.0 for v2; add batch size 32 for unit test; add LD library path for paddle tests temporarily
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove printf line in fused_attn.cpp
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add batch size 32 for unit test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update cudnn-frontend to 0.9.2
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove temporary LD library path used for testing pre-released cudnn 8.9.3
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add long sequence support and unify three backends for fused attention
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update cudnn-frontend to v0.9.1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace cpu_float2half_rn with __float2half_rn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix backend selection and NVTEDType
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

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

* make cudnn plan caches thread_local
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix CI
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace cuDNN throw with NVTE_CHECK
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix replacement of cuDNN throw with NVTE_CHECK
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* force dropout probablity to 0 in inference mode
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change negInfinity to be consistent with m512 fused attn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove float2half conversion for scale_dropout
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add back runtime api for sm detection
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add gemm3 to enums FP8Fwd/BwdTensors
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* change dropout from no to yes for fmha_v1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove output_rng_state in m512 kernels
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix elts_per_thread calculation in kvpacked fwd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove dropout=0.0 restriction for m512 fused attn
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove output_rng_state completely from m512 kernels
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Enable fused attention dropout
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Cast the uint32 key/counter to int64
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update dropout support in fused attention docs
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revise devPtrCuSeqlen* to align the naming
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Support different Jax PRNG impls
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revert CastAsync since it is not used
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Implement is_training for 16-bit fused attn
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add fused attn with dropout sanity unit tests
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the comments readability and rng_state checker
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change the attention dropout shape to align other frameworks
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Make encoder tests deterministic
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change the default seed for the jax encoder tests
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Maintain offset in TE
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the resource safety
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revert rng_state type to allow only i64
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Handle the corner case for elts_per_threads calculation
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Populate rng state by kernels
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename rng_state as seed in cpp_extensions
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update the attention dropout comment
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* fix headers for doxygen
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix description f16 and use half precision instead
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* fix lint
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Unfused scale+softmax if bias is present
Signed-off-by: Reese Wang <rewang@nvidia.com>

* WAR a causal masking + no_bias bug and add the unittests
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix the optional args (bias) sharding
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Disable fused attn in JAX by default, enable it with NVTE_USE_FUSED_ATTN
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add thread local for the plan cache
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename dbeta to dbias for the readability
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add scaled softmax with dropout test cases
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Updated NVTE_FUSED_ATTN variable name
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add fused attention unit tests
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use NVTE_* enums
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use NVTE_Mask_Type and remove FMHADescriptor
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Move common functions to utils
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change namespace to fused_attn
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Move fused_attn_max_512_fwd_qkvpacked under the general APIs
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add fused_attn_max_512_bwd_qkvpacked
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Move fused_attn_max_512_bwd_qkvpacked under the general APIs
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove redundant blank line
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix a potential bug for cu_seqlen converter
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Reformat fused_attn_max_512
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Refine the unfused attention warning message
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename to fused_attn_max_512
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove the deprecated header
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix flax import
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename to fused attn
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add attention related mask
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add attn_mask_type and attn_bias_type
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Refactor jax primitive API
* Merge q_cu_seqlen and kv_cu_seqlen
* Remove is_causal_masking
* Replace seed with rng_state
* Add is_training argument
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove dsoftmax from the customcall
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add None guard for bias and dropout_rng
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add version guard
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add is_fused_attn_kernel_available() to correctly dispatch the attention impl
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix the merge conflict
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Adjust the code style
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add the missing blank lines
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change the order of FADescriptor members
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance the readability of fused_attn_max_512.cu
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Generalize the input dimension unpacking
Signed-off-by: Reese Wang <rewang@nvidia.com>

* 16 bits fused attention requires 8.9.1
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Update fused attention support matrix
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Handle None type when sharding
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change to the padding ratio
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Performance optimization for non-bias cases
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revert the cudnn-frontend PRIVATE keyword which was used for debugging
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Revert "Update fused attention support matrix"

This reverts commit 4effe67d0f08f733919a329ce5ab421958740f4a.
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Treat b * s as total_seqs to align ragged cases
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add FP16/BF16 max_seqlen <= 512 fused attention to the support matrix
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Refine test_fused_attn.py

* Replace reference code with flax.linen
* Remove unnecessary comments
* Use AttnMaskType
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Unify the cuDNN compile version
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add dropout to the support matrix
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Slightly adjust the headers
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Typo fix: remove redundant either
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Consolidating fused attention requirements
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Replace cudnn_frontend::throw_if with NVTE_CHECK for the better error line report
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename to fused_attn_fp16_bf16_max_seqlen_512 for the better readability
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove CUDNN_FRONTEND_UNUSED
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add more annotations to the custom calls
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>

* move dbias from input list to output list for bwd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* split asserts into three for bias checks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Update transformer_engine/pytorch/cpp_extensions.py
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>

* fix asserts for bias checks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* another fix for asserts for bias checks
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Signed-off-by: cyanguwa <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

remove used function ternary_pw_op_create
Signed-off-by: Charlene Yang <charleney@nvidia.com>
Co-authored-by: Charlene Yang <charleney@nvidia.com>

* Add FP8 fused attention to TE for PyTorch
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add license for cudnn-frontend, modify installation requirements, and refactor some headers for aesthetics
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add c api docs for fused attention
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add exception for unsupported precision/sequence length combinations
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix installation requirement for non fused attn use cases
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix docs for fused-attn
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* prefix enums with NVTE_ and replace old MHA_Matrix with NVTE_QKV_Matrix
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* minor fixes based on PR comments
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix description for kvpacked fwd
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix description of Bias in C api
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* minor fixes for cudnn requirement and description for QKV tensors
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix QKV layout description and support matrix for C api
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add asserts to cpp_extensions for qkv layout/bias type/attn mask type
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix typo precision
Signed-off-by: Charlene Yang <charleney@nvidia.com>

---------
Signed-off-by: Charlene Yang <charleney@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Charlene Yang <charleney@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>