Implement per-page K/V quantization for paged attention:
- Add KV_BLOCKSCALE enum to BlockAttentionQuantScaleEnum
- Use exp2 shift trick to eliminate explicit P scaling overhead
- Prefetch physical pages offset for KV cache, overlaps with
computations
## Proposed changes
Please describe the motivation behind the pull request, whether it
enables a new feature or fixes a bug. If there are associated pull
requests or issues, please link them to the pull request.
## Checklist
Please put an `x` into the boxes that apply. You can also fill these out
after creating the PR. If you're not sure, please don't hesitate to ask.
- [ ] I have added tests relevant to the introduced functionality, and
the unit tests are passing locally
- [ ] I have added the test to REGRESSION_TESTS list defined at the top
of CMakeLists.txt in tests/CMakeLists.txt, **IF** the test takes more
than 30 seconds to run.
- [ ] I have added inline documentation which enables the maintainers
with understanding the motivation
- [ ] I have removed the stale documentation which is no longer relevant
after this pull request
- [ ] (If this change is user-facing) I have added release notes which
provide the end users with a brief summary of the improvement from this
pull request
- [ ] I have run `clang-format` on all changed files
- [ ] Any dependent changes have been merged
## Discussion
If this is a relatively large or complex change, feel free to start a
discussion by explaining why you chose the solution you did and what
alternatives you considered
---
🔁 Imported from
[ROCm/composable_kernel#3696](https://github.com/ROCm/composable_kernel/pull/3696)
🧑💻 Originally authored by @Jeff-Huang
---------
Co-authored-by: Jeff Huang <chiachi.huang@amd.com>
Co-authored-by: Illia Silin <Illia.Silin@amd.com>
* [Compiler] Addressing new compiler warnings
Clang enables new lifetime warnings in production and we see build
errors due to this with the staging compiler.
The attributes added in this PR are suggested by the compiler. However,
I'm not very familiar with the code base, so the changes may be
incorrect.
* Update some more instances
* Adds file-level ignores via clang diagnostic pragma
The number of instances was large, so I decided to use file-level scope
to disable the warning via pragma clang diagnostic ignored.
It also showed this warning coming from the gtest dependency. For that,
I did add the respective command line flag to the CMake variables. I
don't know if this is acceptable or not.
* This adds the remaining instances
For a build on gfx90a.
* fix clang format
* Adding couple more instances from gfx1200 build
* Fixed another few instances
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
[ROCm/composable_kernel commit: 069500464d]
* add block scale parameters to kernel
* add block scale to kernel
* add smoke test
* format
* Revert "format"
This reverts commit 356c3c9706.
* only format my code
* format py
* fix auto not allowd in function prototype
* change instance tttt to ttff
* fix structured binding issue
* change s_acc elementwise op
* async pipeline add block scale
* add quantation P using shift exp2
* precompute (m - shift) once per row
* change blk scale seqstrt ptr name
* fix some name
* fix for deduction guide
* fix some comments
* add P scale to qr_ksvs_pipeline
* add comment to idx_identity
* change the method of calculating descale block index
* unify naming style: use block_scale_ as name prefix
* unify naming style
* update the CHANGELOG.md
* Add FP8 block scale quantization support for FMHA forward kernel
---------
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com>
[ROCm/composable_kernel commit: dd0b4294af]
* solve compiler issue
* solve the gfx950 mfma shuffle regression
* refactor jenkinsfile to handle arch name better
* [CK TILE] set divisor to count of thread along k dimension
* fix the compiler error
* solve degradation
* Finish the multiplies fix
* fix the scales
* solve compilation error
* solve the composes
* solve the error of tile sweeper
* fix the test and example
* fix for gfx950
---------
Co-authored-by: Max Podkorytov <4273004+tenpercent@users.noreply.github.com>
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
Co-authored-by: Cong Ma <congma13@amd.com>
[ROCm/composable_kernel commit: 00c46785a8]
- Enable page_size=1 support in batch prefill codegen (linear layout only).
- Implement per-token page lookup in `kv_offset_array_transform` for page_size=1 to handle 3D input tensors correctly.
- Relax `kPageBlockSize` alignment assertion for the page_size=1 case.
[ROCm/composable_kernel commit: c9f112b026]
* add page_block_size parameter
* add is_sglang_layout to parameters
* add kv_offset_array_transform to batch async for page size 16
* add kv_last_page_lens to kernel
* change kv layout to [num_total_pages, page_block_size, hdim]
* format
* - enable codegen of batch_prefill kernels
- create new problem struct BlockFmhaBatchPrefillPipelineProblem for
batch prefill kernels
- generate different page sizes of batch prefill kernels (1, 16)
* 1. fix wrong calculation of page id in kv_offset_array_transform in gfx950
2. support page size 1024
* fix python format
* change kv cache layout to [num_blocks, num_kv_heads, head_size/x,
block_size, x] and [num_blocks, num_kv_heads, block_size/X, head_size, X]
* 1. Introduced `kVectorSize` in BlockFmhaBatchPrefillPipelineProblem instead of using hardcode values
2. Makes batch prefill kernel traits structures inherent from fmha fwd
traits
3. Add some static check for Page size, vector size, hdim, ..., etc.
* [Refactor] Replace is_sglang_layout with Enums for KV cache configuration
Refactored `fmha_batch_prefill` to use `BlockAttentionKVCacheMemoryLayoutEnum` (VECTORIZED/LINEAR) and `BlockAttentionKVCacheLookupTableEnum` (SGLANG_1D/VLLM_2D) instead of a single
boolean.
**Changes:**
* Added Enum definitions in `block_attention_kvcache_layout_enum.hpp`.
* Updated Kernel, Pipeline, and Traits to template on these Enums.
* Implemented `kv_offset_array_transform` logic based on `kKVMemoryLayout`.
* Refactored `PageBlockTableKargs` to adapt to `kKVLookupTable`.
* Updated CodeGen scripts to support new parameters.
This decouples memory layout from the paging mechanism, enabling flexible KV cache configurations.
* 1. remove batch prefill pipeline with sk_pad=false
2. correct some comments
3. add static assert to make sure v offsets is in same page within a tile.
* fix vgpr spill count
* remove unnecessary t2s functions
* add fp8 support for receipt 200 and 600 in fmha_bath_prefill.py
* support linear kv cache layout
* Remove block_table_ptr from fwd_batch_prefill_args. Instead, reuse
kv_page_indices as a pointer of the lookup table.
* 1. merge multiple transforms into single transform.
2. add static check to make sure vlayout is row-major.
* move FmhaFwdCommonKargs::seqlen_k_ptr to VllmPageTableKargs.
* update changelog
---------
Co-authored-by: ltqin <letaoqin@amd.com>
Co-authored-by: PoYen, Chen <PoYen.Chen@amd.com>
[ROCm/composable_kernel commit: cc75a1dc5f]
* replace qr with async pipeline
* Add fp8fp32 to DTYPE_BITS
* Add kAlignmentRandVal to avoid compile fail
* format
---------
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
[ROCm/composable_kernel commit: eb7f617713]
* remove EXCLUDE_FROM_ALL from ck-tile examples
-> +15 min build time w/ 64 threads for a single arch
* fix cpp17 compile error in the ck-tile examples
---------
Co-authored-by: khuagarw <khuagarw@amd.com>
Co-authored-by: Ding, Yi <yi.ding@amd.com>
[ROCm/composable_kernel commit: 79aae7c7f7]
* Pass hdim to tile_example_fmha_fwd in fp8 tests
* Add WMMA support to fwd FMHA pipelines
* Tune tile sizes a bit for less spilling
fp16 256 is still quite slow
* Fix Q grad tile distribution for warp size = 32 and hdim >= 256
With AccDataType = float and warp size = 32, K0 becomes 0, K repeat is required to correcty distribute the tile.
* Use code based on BlockDropout in BlockDropoutBwd
* Fix split KV combine kernel for gfx12 (warp size 32) and make it more universal
* Fix LSE LDS tensor descriptors: kMaxSplits and kM0 were swapped, it worked on gfx9
because they both equal to 8 while on gfx12 they are 8 and 4;
* Fix Oacc LDS tensor descriptor: it was transposed even though its shape=[4 * kM0, kN1],
it worked on gfx9 because 4 * kM == kN1 == 32;
* Removing these hidden dependecies allows to support:
* any number of warps (power-of-2), not only 4;
* kN1 = 16, not only 32;
* any number of splits;
* Rename ids like o_acc_4 and Oacc4 to eliminate confusion: kNumWarps doesn't have to be 4 now
* Replace hard-coded kN1 in dispatch code with the requested tile size
* Add gfx12-specific tile sizes for split KV
* Pass GPU architecture to kernel generation scripts
This is still a temporary solution.
* Build and run FMHA CI tests for gfx12
* Fix issue after merging
* Fix bwd tile sizes
The current pipelines always read only one tile K and V tile, this
requires bk0 == bhdq and bk2 == bhdv (kK0 == kQKHeaddim and
kK2 == kVHeaddim).
* Use hardware f32->f8 on gfx12, remove v_perm
__builtin_amdgcn_perm is not needed because
__builtin_amdgcn_cvt_pk_fp8_f32 allows to specify which word (16 bit of
32-bit dword) is used to store results (two f8 values).
* Update changelog
* Add WMMA support to pagedkv
* Fix scripts after rebasing
* Support 16x16 (MFMA, WMMA) and 32x32 (MFMA) tiles in fwd and bwd BlockDropout
Add comments with dropout implementation details
Fix performance regression of fwd+dropout
* Remove some usage of type punning (reinterpret_cast with ref or ptr) in Philox;
* "scalarize" seed and offset, they may come either from kernel args or from device memory
(presumably loaded with vector loads).
These changes help the compiler to procude more optimal code and reduce register spilling.
Use WarpGemmDispatcher instead of explicit WarpGemmMfma... to get CWarpDstrEncoding
Use code based on BlockDropout in BlockDropoutBwd
Refactor BlockDropout (fwd)
Implement BlockDropout (fwd) for WMMA
Originally BlockDropout only supported 32x32 tiles (IsWG32 = true),
this version supports 16x16 tiles.
If MPerBlock > MWarp * 16, it can generate numbers for two 16x16 tiles, similarly
to BlockDropoutBwd.
Implement BlockDropoutBwd for WMMA
Remove MakeRandValLds* functions unused in BlockDropoutBwd
Remove unused Run overload from BlockDropoutBwd
* Fix regression with philox seed and offset when they exceed 32-bit int
__builtin_amdgcn_readfirstlane works with 32-bit values, seed and offset
are 64-bit so they get truncated.
* Fix names after cherry-picking
* Fix selection of a fallback tile based on bm0
The assumption that the largest bm0 == 128 is not always true for
current fp32 tiles.
* Do not use filters related to qr_async_trload
They disable tiles/pipelines which are valid for gfx12.
* Use different dstr encoding when C is transposed
* Do not call GetQKBlockGemm (and hence WarpGemmDispatcher) in host code
Some WarpGemmDispatcher instantiations are defined only
for specific archs and undefined on host.
Calculations related to sched barriers are moved from Pipeline's public
fields into pipeline's operator().
* Fix incorrect name WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution
Correct name is WarpGemmMfmaFp8Fp8F32M32N32K32SwizzleBTransposedCDistribution
because it's 32x32x16 with IterateK = 2 so K = 32, also all tiles used
in codegen scripts are 32, 32, 32.
* Generalize usages of WarpGemmDispatcher for MFMA and WMMA
WarpGemmMfmaFp8Fp8F32M32N32K32SwizzleBTransposedCDistribution is still
used explicitly becaus of swizzle factor = 4.
* Mark has_load_tr as maybe_unused
There are no transpose loading for RDNA.
* Remove CK_TILE_USE_MFMA/WMMA from fmha-related code
* Detect BlockSize on host based on warp size of the current device
If kBlockSize == kNumWarps * get_warp_size(), the kernel is launched with
kBlockSize / 2 because on host get_warp_size() == 64 always.
* Fix calculation of grid size for combine kernel with warp size = 32
* Add missing includes and header
* Support multiple archs in one binary for fwd
* Support multiple archs in one binary for fwd_splitkv, fwd_appendkv, pagedkv_prefill
* Support multiple archs in one binary for bwd
* trload kernels are compiled only for gfx950;
* instances with padding are checked after instances without padding so
they can be used as fallbacks (similarly to fwd);
* Extract common code from register_traits
* Revert "Fix regression with philox seed and offset when they exceed 32-bit int"
To simplify merging , the proper fix is in develop already.
* Support new numerical d paddings in trait ordering checks
* Build fp32 tests only on gfx9
* Do not use hardcoded M0 = 64 for dot bwd kernel
* Use textwrap.indent from standard library
* Make fp8 pipelines on gfx12 consistent with gfx9
* Update tests for current pipelines
* Make ninja check more responsive in CI
ninja buffers output so this job looks hanging.
* Support fp8fp32 by limiting O vector size
The fp32 output type requires storing 8 * sizeof(float) = 32 bytes,
which is not implemented (here 8 is the number of C values per lane for
v_wmma_f32_16x16x16...).
* Remove unused cmake options
* Unify including amd_buffer_addressing.hpp/_builtins.hpp
* Temporarily use amd_buffer_addressing.hpp on >=gfx10
amd_buffer_addressing_builtins.hpp uses inline asm for loads/stores
which is not compatible with >=gfx10:
* 1 scalar for exec masks instead of 2,
* gfx12 uses different instruction names etc.
* Update asm in bf16 conversions to work with warp 32
* Do not generate splitkv/appendkv with vlayout=col for consistency with fwd
* Add arch tags to kernels/host funcs, compile for each arch separately
* Add kM0 to fmha_bwd_dot_do_o kernel name to match filename
* Add workaround for miscompilation of bwd with padded hdim
SWDEV-559729: v_wmma instructions can be incorrectly placed in divergent
branches used to store padded tensors (when some lanes are inactive due
to padding). Inline asm with dummy dependencies on VGPRs of the tensors
prevents the compiler doing this.
* Fix add_gtest_executable for absolute paths
Some tests (like gemm_tile_engine) pass absolute paths to source files.
In CI the branch name is a part of the root dir, and if the branch name
contains "wmma", "xdl" etc., files can be incorrectly excluded.
* Run only hdim 128 smoke tests for fp8fp32
There are no instances for hdim 64 and 256.
* Format py with ruff to simplify merging develop
* Fix incorrect var name
* Codegen for gfx9,gfx950 when --targets is not specified
Aiter and Pytorch require changes for passing their targets to the codegen scripts.
With this temporary solution the files are generated but not all of them
have to be really built (depending on the used --offload-arch=).
* Combine arch-related values into ArchTrait
This more centralized approach removes duplication of various formatting templates.
* Try a workaround for Jenkins error "groovyjarjarasm.asm.MethodTooLargeException: Method too large"
Some code is extracted into a function.
[ROCm/composable_kernel commit: 1e77695fe8]
* [CK_TILE] fmha: Add query padding support to backward pass
Introduces support for query sequence padding (q_padding) in the FMHA backward pass kernels.
- Passing `seqlen_q_ptr` to the backward kernels to distinguish logical from physical sequence lengths.
- Updating `OGradDotO`, `ConvertQGrad`, and `DQDKDV` kernels to respect logical lengths and handle zero-length sequences.
- Aligning LSE indexing in the forward kernel with the padded layout for consistency.
- Adding a new GTest suite (`test_fmha_bwd_kernel_padding.cpp`) with comprehensive tests for various padding scenarios, including zero-length
sequences and deterministic mode.
* fix clang format
* Adapt fmha_bwd_runner.cpp to new q, kv sequence padding
Add backward q/kv sequence padding unit tests.
* [CK_TILE] fmha: Unify sequence length and padding handling
Refactor the handling of sequence lengths and padding in the
FMHA forward and backward kernels to provide a more unified and flexible
interface.
- Replaced `seqstart_padded_*_ptr` with a more robust system that uses
`seqstart_*_ptr` for physical sequence lengths and introduces
`seqlen_*_ptr` and `cu_seqlen_*_ptr` for logical (unpadded) lengths.
- Established a clear order of precedence for determining sequence
length: cumulative lengths (`cu_seqlen_*_ptr`) take priority,
followed by per-sequence lengths (`seqlen_*_ptr`), and finally
physical lengths derived from `seqstart_*_ptr`.
- Clarified the distinction between "group mode" and "batch mode" and
how sequence lengths are handled in each case.
- Renamed `cu_seqlen_kv_ptr` to `cu_seqlen_k_ptr` for consistency.
- Updated comments and documentation to reflect the new argument
structure and usage.
---------
Co-authored-by: illsilin_amdeng <Illia.Silin@amd.com>
[ROCm/composable_kernel commit: 7c6430eca0]
* Support 16x16 (MFMA, WMMA) and 32x32 (MFMA) tiles in fwd and bwd BlockDropout
Add comments with dropout implementation details
Fix performance regression of fwd+dropout
* Remove some usage of type punning (reinterpret_cast with ref or ptr) in Philox;
* "scalarize" seed and offset, they may come either from kernel args or from device memory
(presumably loaded with vector loads).
These changes help the compiler to procude more optimal code and reduce register spilling.
Use WarpGemmDispatcher instead of explicit WarpGemmMfma... to get CWarpDstrEncoding
Use code based on BlockDropout in BlockDropoutBwd
Refactor BlockDropout (fwd)
Implement BlockDropout (fwd) for WMMA
Originally BlockDropout only supported 32x32 tiles (IsWG32 = true),
this version supports 16x16 tiles.
If MPerBlock > MWarp * 16, it can generate numbers for two 16x16 tiles, similarly
to BlockDropoutBwd.
Implement BlockDropoutBwd for WMMA
Remove MakeRandValLds* functions unused in BlockDropoutBwd
Remove unused Run overload from BlockDropoutBwd
* Fix regression with philox seed and offset when they exceed 32-bit int
__builtin_amdgcn_readfirstlane works with 32-bit values, seed and offset
are 64-bit so they get truncated.
* Add F32 MFMA warp gemms
* Support f32 in fwd FMHA
* Implement transpose_vectors for 4-byte types (float)
* Fix unexpected implicit f32->uint32 cast in buffer_store<4>
__builtin_amdgcn_raw_buffer_store_b32 expects unsigned int but float was passed (implicitly casted to uint).
mbuf_t types in other buffer_store<> are changed for consistency.
* Support F32 in bwd FMHA
hdim = 256 is disabled for now because it uses too much memory on gfx90a
* Support Headdim = 48 (divisible by 16) in fwd
* Add fp32-specific receipts (800 and 801)
* Tune fwd tiles
* Tune bwd tiles
* Use small tiles only for small seqlen_q
* Fix after rebasing
* Fix selection of a fallback tile based on bm0
The assumption that the largest bm0 == 128 is not always true for
current fp32 tiles.
* Remove constraints and adjust filtering for fp32
Custom constraints are no longer needed because now the smallest tile
is selected automtically based on seqlen_q.
Filters related to qr_async_trload disabled valid fp32 tiles.
* Add fp32 tests
* Make splitkv and appendkv compile for fp32 only
There are no instances yet, but API still must compile when only fp32 is
requested.
* Remove unimportant f32 instances
* Add test_ck_tile_fmha_*_fp32 to REGRESSION_TESTS
* Replace magic numbers with a constant, improve comments for dropout
* Update changelog
* Fix condition that dq_acc must be set to zero when mask is used
The change was introduced in #2799
* Replace warp_uniform with recently added amd_wave_read_first_lane
* Add hdim = 96 and 192 to fwd
[ROCm/composable_kernel commit: 1edd250115]
* Fix validation of rotary embedding with time_kernel_
When rotary embedding is used, the appendkv kernel modifies the q tensor
(multiple times when time_kernel_ is set). We need to reset the q buffer
and rerun all kernels.
* Fix synchronization issue in splitkv combine pipeline
Different warps can read and then rewrite the same values of lse_acc_lds.
Sometimes warps progress at different speeds, one warp can rewrite
values that are still being read by another warp.
Running the tests multiple times and, preferably, with multiple
processes on the same GPU helps to trigger this issue:
bin/test_ck_tile_fmha_fwd_fp16 --gtest_repeat=-1 --gtest_shuffle --gtest_throw_on_failure --gtest_filter="TestCkTileFmhaFwd/*KV*"
[ROCm/composable_kernel commit: c6bfd97c2d]
* * [CK_TILE] Add sequence padding and variable length support in fmha (and v3)
- Group Mode Padding: Introduces the `-s_qpad` argument to support
physically padded layouts. Kernels now use padded start pointers
(`seqstart_padded_*_ptr`) for memory addressing.
- Batch Mode Variable Length: Adds `-q_eff_lens` and `-kv_eff_lens`
arguments for efficient processing of variable-length sequences by
passing cumulative effective lengths (`cu_seqlen_*_ptr`) to the kernel.
- FMHA examples: Support padding and variable length both in
group and batch mode. Dispatcher is updated as well (dispatch to
kPadSeqLenK enabled pipeline).
- New padding test cases: Add padding test cases to `smoke_test_fwd.sh` and
`test_fmha_fwd.inc`, and add benchmarks to `benchmark_fwd.sh` and
`benchmark_fwd_v3.sh` as well. These test cases and benchmarks that
specifically validate/benchmark the new padding and variable-length
functionalities in both group and batch modes.
* [CK_TILE] Fix build error in fmha unit tests
* [CK_TILE] add mqa, gqa to sequence padding unit tests
* [CI_TILE] Reduce the number of padding seqlen unit tests in FMHA to avoid timeouts in CI
* [CK_TILE] remove unnecessary MageKArgs overload in FmhaFwdV3Kernel and FmhaFwdKernel
[ROCm/composable_kernel commit: 518d24e662]
