In Jax TE, multiple threads in the same process are spawn to train for each GPU. Therefore hipModueLoadData, hipGetFunction need to be run for each GPU in each corresponding threads.
* Update license year
* Add initial code to override decode problem
* Fix splitkv traits/args overriding error
* Reshape and transpose lse for decode
* Remove debug code
* Prettify example code
* Use better function name
* Add kMergeNumHeadGroupsSeqLenQ flag
Kernel user can use this switch to turn on/off optimization for
some problem sizes
* Add missing flag declarations
* Default turn off kMergeNumHeadGroupsSeqLenQ in codegen
* Group similar statements together
* Remove assumption of seqlen_q=1
* Remove kMergeNumHeadGroupsSeqLenQ from splitkv combine kernel
* Support kMergeNumHeadGroupsSeqLenQ=true in fmha splitkv kernel
* Run kMergeNumHeadGroupsSeqLenQ=true kernels when need
* Fix group mode block skip logics
* Undo changes of normal fwd kernel
* Update in GridSize() and using GridSize() for splitkv kernel (#1799)
---------
Co-authored-by: Qianfeng <qianfeng.zhang@amd.com>
* Remove using tile partitioner for fmha_fwd_kernel
* Remove using tile partitioner for fmha_fwd_splitkv and splitkv-combine kernels
* Remove using tile partitioner for fmha_fwd_appendkv kernel
* Unify the format of GetTileIndex
* Add check for zero values
* Add static assertions
* Remove invalid option '-e' in smoke_test.sh
* Use correct path of smoke_test.sh
* Avoid zero-sized shared memory array
* Add warning comment
* Replace expr by integer_divide_ceil() call
* Use more readable constant names
* Write down assumption as static assertion
* Add more diagnostic error messages
* Fix wrong BlockWarps when using default pipeline policy
* Add more static assertions for A LDS desc
* Allow using vector size < 8 for data type fp16/bf16
* Align vector size between DRAM dist & LDS desc
* Remove no-longer used func decl
* Fix wrong displayed piepline name
* Undo policy template changes for tile_example_gemm_basic
* Add missing space and make error message stands out
* Unify print precision
* Add missing include directive <iomanip>
* Replace constant 64 by get_warp_size() call
* Replace constant 128 by named variable: BankLength
* Add kAMBlock/kBNBlock attributes
* Allow usig different A/B warp dist for multiple blocks
* Add helper function to get warp dist encodings
* Add 4x64x4 fp16 warp gemm attribute impl
* Complete the A/B warp dist encoding logic
* Fix wrong thread mapping for C matrix
* Use smaller vector size for small tile
* Add static assert to block unsupported warp gemm impl
* Extract common code out as helper method
* Add 4x64x16 fp16 warp gemm type alias
* Add comment to warning developers
* Undo WarpGemmAtrributeMfma<> changes
* Use more clear static assertion error message
* Add trivial wrapper to get warp dstr encodings
* Only transpose warp gemm result if it's square
* Fix compilation error
* Support multi-block warp gemm (on N direction)
* Remove duplicated code
* Fix output encoding of warp gemm
* Fix wrong shape of WarpGemmAtrributeMfmaIterateK<>
* Remove unused code
* Fix wrong shape of WarpGemmAttributeMfmaImplF16F16F32M4N64K4
* Add type config for bf16_t
* Add 4x64x16 bf16 warp gemm
* Update WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
* Add 64x4x4 fp16/bf16 warp gemm impl
* Add 64x4x16 fp16/bf16 warp gemm
* Add static assertion for better error diagnostic
* Get Q dram dstr directly form block gemm
* Add missing header: fused_moe.hpp
* Allow specifying different warp-gemm for gemm0 & gemm1
* Store P matrix into LDS before gemm1
* Fix inconsistant kernel name
* Remove constraint on gemm0 & gemm1 block warps
* Remove unsupported vector size from checking list
* Allow using 4x64x16 warp gemm for gemm0
* Finish policy customization
* Finish pipeline modification
F#
* Use block warps in codegen
* Fix wrong rank of m_lds_window origin
* Use better distributed tensor
* Make P-store earlier
* Remove duplicated experssions
* Remove unnecessary tile window
* Create new files for new splitkv pipeline
* Separate old/new pipeline codegen logic
* Sync changes form develop
* Undo gemm kernel/pipeline changes
* Undo gemm example changes
* Remove blank lines
* Fix typo
* Use new warp gemm interface
* Fix link error
* Fix wrong pipeline tag
* Fix more link error
* Avoid unnecessary padding
* Always use vector load for K
* Padding on fastest dimension when necessary
* Force padding Q on hdim_q
* Set high dimension padding flag to false
* Re-format headers
* Use warps=<1, 4, 1> for both gemm0 & gemm1
* Fix complilation errors
* Remove m/l shuffle logics
* Ignore duplicate data when write lse_acc
* Use gemm0 block warps as lds tile width
* Remove hard-coded numbers
* Fix wrong distribution width
* Remove unnecessary code
* Add s_barrier before writing to LDS
* Store Q into LDS before gemm0
* Fix wrong Q tile size
* Use simple Q lds descriptor for debuging
* Use more realistic Q lds descriptor
* Add comment & use better variable name
* Make Q lds space not overlapped with others
* Remove unnecessary block_tile_reduce_sync() call
* Move Q load statements
* Move block_sync_lds() right before use
* Re-order instructions
* Remove necessary lambda expression
* Use 8 threads on kMaxSplits direction while doing reduction
* Tiny correction for using 8 threads on kMaxSplits direction for combine kernel
* Padding num_split direction of o_acc tile window to 4x
* Update splitkv combine pipeline design
* Add kN1 back to splitkv combine pipeline problem
* Fix compilation errors
* Add missing template parameter
* Fix wrong splitkv combine kernel name
* Fix wrong origin
* Fix wrong LDS descriptor shape
* Fix sync & reduction logics
* Remove unnecessary static assertions
* Extract tile size computation logics
* Make sure we can reuse padding flags in combine kernels
* Rename variables
* Use OaccDataType in BlockFmhaSplitKVCombinePipelineTileSizes<>
* Remove unnecessary static assertion
* Fix function name typo
* Add constraint on kN1 template parameter
* Hide K tile loading latency in earlier iteration
* Fix wrong splitkv kernel name
* Use s_shuffling to replace p_shuffling which removes the needs of cross-warp reduction
* Rename pipeline
* Fix wrong pipeline name attribute
* Add GetAlignmentQ() for NWarpSShuffle pipeline
* Separate Q tile into dram tile & register tile concepts
* Remove non-squre warp gemm transpose c type alias
* Fallback tile size changes for fmha fwd splitkv
* Remove redundant change
* Refine naming for the S tile
* Use better naming of the S tile dstr (read from lds)
* Share Q lds with K lds
* Tiny change
* Fix with using static_for for passing CI checking
---------
Co-authored-by: Qianfeng Zhang <Qianfeng.Zhang@amd.com>
* Change in fwd-splitkv kernel to support num_splits=1 case
* Update in codegen fwd-splitkv to make num_splits > 1 cases pass
* Specify instance traits in dispatch
* Fix link error for fp8 kernels
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Generate group mode paged-attn kernel
* Enable paged-kvcache + group mode support
* Add missing header: fused_moe.hpp
* Add comment to explain kernel arg usage
* Make error message more clear
* Add comment for confusing data member names
* Add more comment for confusing variable names
* Fix typo in option description
* Add ceil_to_qualified_tile_length()
* Rename kK0BlockLength to kQKHeaddim
* Add kSubQKHeaddim concept to support headdim96
* Fix in math.hpp to avoid using __half interfaces
* Add LdsBufferSequence instance for headdim96
* Update in fmha_fwd/fmha_fwd_splitkv codegen to support hd96 testing
* Disable hd96 instance generation in codegen fmha_fwd and fmha_fwd_splitkv to save compiling time
* Reformat one file
* Fix text alignment in fmha_fwd_splitkv.py
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Use pre-defined constants for readability
* Use vector write for o_acc tensor
* Remove no-longer used policy method
* Deprecate no-longer used policy/pipeline
* Specify gemm0/gemm1 block warps separately in codegen
* Fix wrong ps_idx creation logic
* Add single-warp block gemm
* Supoprt single-warp gemm0
* Make MakeCBlockTile() as static method
* Use MakeCBlockTile() to get underlying tile distribution
* Use kNumGemm1Warps to compute # threads for gemm1
* Put normal case in the if clause
* Refine fmha splitkv block mapping
* Refine & fix the lse_acc/o_acc layout
* Fix wrong LDS size for K tile
* Use kK0=64 for hdim=128,256 fmha splitkv kernels
* Use kK1=64 for hdim=32,64,128 fmha splitkv kernels
* Undo kK0/kK1 changes
* Use more reasonable GetAlignmentV() computation
* Using store_tile() in fmha splitkv kernel epilogue
* Simplify the codes in splitkv_combine pipeline
* Always set kPadSeqLenK=true for fmha splitkv kernels
* Change in Oacc Alignment and TileDistribution to be more adaptable to tile sizes
---------
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
* Use dictionary to config all the functions
* Add init codegen logic for fmha fwd appendkv
* Call HIP_CHECK_ERROR() macro to get real source info
* Setup meaningfull arguments
* Sync kernel name with the codegen
* Add knew/vnew tensors to the kernel argument
* Fix wrong K values after appending
* Fix vnew append errro
* Extract common logics
* Fix Vnew tile dstr for row major case
* Conditionally add fwd_splitkv API in fmha_fwd example
* Conditionally add call to fmha_fwd_splitkv()
* Remove "EXAMPLE_" prefix of cmake variables
* Regsiter API handlers automatically
* Early return if 0 < s_k_new is not supported
* Show message if we are ignoring option
* Unify CMakeLists.txt coding style
* Set num_splits=1 if split-kv is not supported
* Add length/stride getters for HostTensor
* Add RoPE example utilities
* Add reference_rotary_position_embedding() (not implemented)
* Finish reference_rotary_position_embedding() impl
* Fix typo of HostTensor<>::get_length()
* Fix compilation errors
* Fix wrong answer when interleaved=false
* Fix wrong answer when interleaved=true
* Append K/V in the host verification code
* Simplify K appending logics
* Simplify v_host_ref definition
* Reduce input/output dimensions
* Rename function: add "batched" prefix
* Apply RoPE on host side
* Rename RoPE utility function
* Fix wrong tensor size
* Avoid invoking deprecated method 'find_module'
* Pass RoPE kernel args
* Create Rotary Cos/Sin tile windows in kernel
* Add compute data type alias for RoPE
* Randomly generate seqlen_knew if needed
* Fix seqlen_knew enabling check logic
* Add minimum seqlen_k to generate compliance kvcache
* Fix compilation error in debug mode
* Fix wrong boundaries
* Fix wrong seqlen_k for kvcache
* Rename variables used in distributio encoding
* Fix rotary cos/sin tensor/tile size
* Add constraint to the rotary_dim option
* Remove unused inner namespace
* Add dram distribution for rotary_cos/rotary_sin (interleaved)
* Only apply interleaved RoPE on Knew for now
* Fix wrong thread starting offset
* Instantiate multiple kernels for RoPE approaches
* Clean-up pipeline
* Fix error in RoPE host reference
* Handle RoPE half-rotated logics
* Support 8x rotary_dim under half-rotated RoPE
* Add comment
* Apply elementwise function to the loaded tiles
* Unify parameter/variable naming style
* Remove constness from q_ptr
* Add code blocks for q_tile
* Apply RoPE to q_tile
* Remove debug print code in kernel
* Fix wrong knew/vnew appending positions
* Use better naming for tile indices
* Add make_tile_window() for adding distribution only
* Skip code if # of block is more than needed
* Move thread locating logics into policy
* Remove always true static_assert()
* Rename header
* Rename RotaryEmbeddingEnum
* Extract rotary embedding logic out
* Re-order parameters
* Align naming of some tile size constants
* Rename more tile size constants
* Fix wrong grid size
* Fix wrong shape of knew_host/vnew_host
* Fix wrong index into knew_host/vnew_host
* Fix wrong rotary_cos/rotary_sin memory size for Q
* Extract Q/Knew vector size to helper methods
* Use different rotary_cos/rotary_sin distr for Q/Knew
* Update host/device specifiers
* Fix wrong data type for Q rotary_cos/rotary_sin
* Remove RoPEComputeDataType type alias
* Shift rotary_cos/rotary_sin by cache_seqlen_k
* Add comment for why I just 't' for all padding flags
* Align commit message to the real comment
* Fix wrong pipeline
* Rename utility function
* Disable host verification if API not exist
* Fix wrong rope key for fp8 pipeline
* Allow only apply RoPE on Q (without append KV)
* Add append-kv smoke tests
* Remove debug statements
* Remove more debug statements
* Re-arrange the 'set +x' command
* Remove no-longer used method in pipeline
* Add missing init code
* Refine pipeline padding settings
* Enlarge rotary_dim limit (8 -> 16)
* Enlarge KPerThread for rotary_interleaved=false
* Update rotary_dim range in smoke_test_fwd.sh
* Add template argument 'kIsPagedKV' for splitkv kernels
* Launch splitkv kernel if given page_block_size
* Fix wrong kernel name
* Fix seqlen_k_min for pre-fill case (1 -> 0)
* Add copy_const<> type trait
* Add another make_tile_window()
* Introduce 'TileWindowNavigator' types
* Simplify TileWindowNavigator interfaces
* Fix tile window navigation bugs
* Disable calling fmha_fwd()
* Remove ununnecessary data members
* Simplify more make_tile_window() overloads
* Move V tile through TileWindowNavigator
* Fix uneven split checking logic
* Move code after decide seqlen_q/seqlen_k
* Make sure we always start reading complete tile
* Use 128 as minimus page_block_size
* Fix wrong origin for bias
* Add batch_stride_k/batch_stride_v in group mode
* Unify origin
* Add missing kernel arguments for group mode
* Add paged-kv codegen logic for appendkv kernels
* Add block_table kernel args for appendkv kernel
* Add tile navigators to the appendkv kernel
* Fix wrong tensor descriptor lengths
* Pass re-created tile window to pipeline
* Fix wrong strides for appendkv kernel
* Allow transit tile_window to another page-block
* Handle cross-page-block write
* Donot perform write again if already in last page-block
* Always add fmha_fwd() api
* Add missing group mode argument
* Remove debug macro usages
* Rename option s_k_new to s_knew
* Separate splitkv/non-splitkv args/traits
* Remove fmha_fwd_dispatch()
* Fix compilation errors
* Remove dropout code in splitkv kernel
* Allow problem types without define kHasDropout attr
* Use generic lambda to init traits objects
* Separate more non-splitkv & splitkv traits/args
* Display more info for specific kernels
* Show more detailed warning message
* Rename 'max_num_blocks' to 'max_num_page_blocks'
* Remove no-longer used pipeline files
* Wrap code by #if directives
* Move functors to the begining of validation code
* Use generic lambda to init all the api traits/args
* Fix wrong seqlen for kvcache
* Add missing comment
* Rename TileWindowNavigator to PageBlockNavigator
* Only expose necessary methods (not attributes)
* Re-order pipeline paremeters
* Refine smoke_test_fwd.sh
* Fix wrong arugment count
* Make tile window directly via PageBlockNavigator
* Remove unused template paremeter
* Remove group mode from appendkv kernel
* Fix skcheck logic
* Fix wrong syntax in skcheck expr
* Use meaningful options in smoke test
* Remove options
* Fix formatting
* Fix more format
* Re-organize bash functions
* Pass cache_batch_idx to kernels
* Support cache_batch_idx in example
* Fix compilation error
* Add more appendkv test
* Add more case for appendkv
* Fix unexisted attribute
* Remove 0 < seqlen_knew constraint
* Clarify the case in warning message
* Remove macro checking
* Force batch mode when invoking appendkv & splitkv apis
* Fix mode overriding logics
* Fix wrong parameter name
* Randomize seqlen_k if use kvcache
* Use randomized seqlen_k for kvcache
* Avoid using too small rotary_cos & rotary_sin
* Rename parameter
* Add seqlen_q & seqlen_k rules
* Add comment
* Add more comments
* Fix compilation errors
* Fix typo in comment
* Remove type argument
* Avoid seqlen_k=0 for kvcache
* Revert "Avoid seqlen_k=0 for kvcache"
This reverts commit 21c4df89e4.
* Fix wrong uneven split checking logics
* Only randomize kvcache seqlen_k if 1 < batch
* Return earlier if split is empty
* Revert "Only randomize kvcache seqlen_k if 1 < batch"
This reverts commit b9a4ab0d7e.
* Re-order seqlen_k_start adjustment logics
* Fix compilation errors
* Re-format script
* Find executable from folder automatically
* Fix kvcache seqlen_k generating logic
* Make comment more clear
* Fix wrong knew/vew appending logic on host
* Add s_barrier to sync threads
* Revert "Add s_barrier to sync threads"
This reverts commit d3f550f30c.
* Support only using 1 row of rotary_cos/rotary_sin
* Rotate Q in different way
* Unify tensor view creation logics
* Fix wrong argument
* Add mask to switch how we use the rotary_cos/sin
* Move attr from traits to problem
* Move has_mask to fmha_fwd_appendkv_args
* Support use uint32_t as SAD operand in Alibi<>
* Use sad_u32() in splitkv kernels
* Store tensor views in PageBlockNavigator
* Use stored tensor view to update tile windows
* Enlarge tensor view size
* Remove debug code
* Fix wrong tensor view size
* Wrap tensor view into PageBlockNavigator
* Add DataType member to PageBlockNavigator
* Remove unnecessary member functions
* Refind macro use
* Fix typo
* Add blank line between directives and actual code
* Re-format files
* Remove type in comment
---------
Co-authored-by: carlushuang <carlus.huang@amd.com>
Co-authored-by: rocking <ChunYu.Lai@amd.com>
