[CK] Add FP8 KV_BLOCKSCALE support for batch prefill
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
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
CK CI migration.
## Motivation
Enable the CK CI after migration from standalone repo.
## Technical Details
Modify the jenkinsfile in projects/composablekernel to update the CI
workflow.
## Test Plan
This is for CK internal testing only.
## Test Result
Set up new CK CI pipeline/dashboard.
## Submission Checklist
- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
* Added two new failure patterns to detect. Including test function to verify if the patterns are detected
* Modifying pattern match to detect docker login failure. Removed passing tests.
* Removing passing tests. Modifying docker pattern to detect failure
* Removed passing tests
* Removing test logging function
This change improves the clang-format CI check to be faster and not
depend on git being available in the build environment.
Changes:
- Use `find` instead of `git ls-files` (no git dependency)
- Check all C++ files: *.h, *.hpp, *.cpp, *.h.in, *.hpp.in, *.cpp.in, *.cl
- Exclude build/ and include/rapidjson directories
- Use parallel processing with 8 cores (-P 8) for ~8x speedup
- Show only errors with unified diff format (-u)
- Clear error messages: "ERROR: <file> needs formatting"
- Preserve original logic: run clang-format only when RUN_CPPCHECK=false,
or run both clang-format and cppcheck when RUN_CPPCHECK=true
Performance:
- Sequential processing: ~93 seconds for 5,899 files
- Parallel with 8 cores: ~12 seconds for 5,899 files
- Per-file processing time: ~15ms
This reduces CI time while maintaining code formatting standards.
1. Added `-DCK_EXPERIMENTAL_BUILDER=OFF` to the `setup_args` to explicitly disable the experimental builder
2. Added a detailed comment explaining why this is necessary:
- SLES15 is a legacy platform with limited C++20 ecosystem support
- While the ROCm compiler supports C++20, the older system libraries and standard library implementation on SLES15 does not reliably support all C++20 features required by the experimental CK Builder
* memory op changes
* memory op changes
* Fixing TILE_ENGINE_BASIC in Tile Engine
* Removing gfx90a from Tile Engine Run
* [CK TILE ENGINE] increasing ci configs for BASIC case
* Setting RUN_TILE_ENGINE_BASIC_TESTS to ON by default
---------
Co-authored-by: Max Podkorytov <4273004+tenpercent@users.noreply.github.com>
* 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>
* Removing hard-coded trace filename
* Including stage name in notification
* Simplifying capture setup and tagging file names with arch
* Removed test property from notification message
* Fixing regex to get arch name
* Fixing error in notification and modified regex
* generate and visualize build traces for all archs
* generate build traces in all cases
* fix jenkins logic
* fix typo
* use more threads for parsing dependency map
* add script to parse ninja traces and issue warnings
* fix python script syntax and header
* fix python syntax one more time
* fix python syntax
* Parallelization in dataset generation
* Parallelizable tests for fwd, bwd data, bwd weight with datasets
* .gitignore generated datasets
* Test parallelization script with round-robin GPU scheduling
* Parallelization updates to test generation and running
* Dataset paths relative to executable
* Update output from test generation
* Default to one GPU in test generation
* Add small dataset tests to Jenkins
* Update copyright lines
* Update test_data/generate_test_dataset.sh
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* Move trap disable
* Common get path function
---------
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* [CK TILE STREAMK] Introduce initial support for tile engine in streamk GEMM.
- This commit lays the groundwork for integrating the tile engine into streamk GEMM.
It focuses on creating benchmark executables for streamk GEMM.
- Additional scripts like test_benchmark.sh and gemm_benchmark.py will be added once
the streamk implementation reaches stability.
* [CK TILE STREAMK] Enable CI to execute tile engine benchmarks for StreamK GEMM
* [CK TILE STREAMK] Refactor: Extract common utility functions.
* [CK TILE STREAMK] Revise tile engine of streamk to align with the updated implementation
* Add pre-commit
* [CK TILE STREAMK] Add 'dp_persistent' and 'reduction_strategy' in output of CK TILE STREAMK
* [CK TILE STREAMK] Fix a bug about value of 'dp_persistent' of CK TILE STREAMK
* [CK TILE STREAMK] Update Jenkinsfile
* [CK TILE Engine] Update StreamK tile engine help message
Remove default value messages as they are automatically printed
* [CK TILE Engine] Update StreamK tile engine
- Remove namespace reboot
* [CK TILE Engine] Update StreamK tile engine
- Fix merge error
* build and run ck_builder tests
* add test_ckb_all to targets
* fix syntax
* fix test path
* Update CMake targets for builder testing in CI (#3290)
Our existing CMake only had build targets. Update CMakeLists.txt to have CTEST targets:
* smoke-builder
* regression-builder
* check-builder
Co-authored-by: John Shumway <jshumway@amd.com>
* use check-builder target
* get rid of test_ckb_all target
* call ninja check-builder separately
---------
Co-authored-by: John Shumway <jshumway@amd.com>
* allow using alternative compiler in all CI stages
* get rid of some redundancies in jenkinsfile
* clean up jenkinsfile a bit more
* further clean up jenkinsfile
* do not force user jenkins in ci dockers
* Renaming old code
* Adding GEMM code with new Architecture
* Partial Progress : Errors
* Partial Progress : Working code
* Changes to element wise function
* Removing Debugging statements
* Working GEMM Multi D code
* Removing Stale Code
* Address Copilot review comments
* Address Copilot review comments
* Changes to validation file
* Changes to common code snippets
* Creating common folder
* Removing duplicate files
* Pointing to right common file
* Pointing to right common file
* Pointing to right common file
* Changing to VERBOSE
* Changing CMAKE messages to verbose
* Updating Cmake with right layout datatype configs
* Working code for GEMM Multi D
* Adding GPU not found pattern
Also, failurePatterns does not need to be global. Moved variable to live in the failure notifications function scope.
* Testing new failure type
* Testing failure
* Removing the forced failure test
* Adding an additional failure pattern
* 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.
* refactor package generation, add dedicated switch
* allow building packages not only on gfx9
* enable last stage to post packages
* stash packages from different arch into separate stashes
* build packages daily automatically
* Partial Progress : CK Tile Engine GEMM
* Partial Progress : CK Tile Engine GEMM
* Partial Progress : Working GEMM Code
* Partial Progress : Working GEMM Code
* Changinf jenkins to remove preshuffle
* Partial Progress : CK TILE ENGINE GEMM Debugging
* Partial Progress : Removing changes that are not GEMM
* Partial Progress : Validation of full block size in GEMM
* Changes in Jenkins to run only fp16 and bf16
* Addressing Review Comments
* Partial Progress : Addressing CI issues
* Partial Progress - Runing GEMM for fp16,bf16 and rcr
* Clang
* Adding fp8 and bf8
* Adding fp8 and bf8
* Adding additional architrcture
* Limited datatypes and layouts
* Adding k_block_per_cu in test config
* Changes to faling CI errors
* Changes to faling CI errors
* Validation for GEMM
* Adding Layout support
* Adding Validations
* Adding layout in jenkins
* Update on Jenkins
* Distribution validation for GEMM
* Resolving merge conflicts
* Solving merge conflicts
