juuso-oskari d77f0bea63 CK-UA: collapse MHA/GQA variants -- one binary per (head_dim, kBlockM)
After moving kBlockQ to runtime in the previous commit, the static
NumQPerKV in `variant_config<V>` and the runtime-vs-static assert in
the kernel became the only things still tying a compiled binary to a
specific num_queries_per_kv. Drop both and the existing instances now
serve every num_qpkv that divides kBlockM evenly.

Concretely:
  * `variant_config<V>` -- remove the NumQPerKV field from every
    specialization.
  * `unified_attention_kernel_traits` -- remove the `num_queries_per_kv`
    / `kBlockQ = kBlockM / num_qpkv` derivation. The BlockTile's 2nd
    entry (the static `kBlockQ` exposed via UnifiedAttentionShape) is
    anchored at kBlockM so it describes the "num_qpkv == 1" fallback;
    the actual kBlockQ is always the runtime value.
  * `unified_attention_kernel_launch` -- recompute kBlockQ at host time
    from `args.num_queries_per_kv` for the total_num_q_blocks math.
  * `unified_attention_kernel.hpp` -- drop the
    `assert(kBlockQ_dyn == kBlockQ)` (it enforced the very coupling we
    just removed).
  * `unified_attention.cpp::select_config` -- collapse the two
    per-num_qpkv code paths into a single (head_dim, avg_rows,
    max_rows) ladder, where avg_rows = avg_q * num_qpkv.

Variant renames (8 variants):
  prefill_d128_mha       -> prefill_d128
  decode_d128_mha_m128   -> decode_d128_m128
  decode_d128_mha_m32    -> decode_d128_m32
  decode_d128_mha_m16    -> decode_d128_m16
  prefill_d64_gqa8       -> prefill_d64
  decode_d64_gqa8_m128   -> decode_d64_m128
  decode_d64_gqa8_m64    -> decode_d64_m64
  decode_d64_gqa8_m16    -> decode_d64_m16

The 16 d=64 instance files lose their `_gqa8` infix to match the
d=128 naming (file count unchanged: 16 dtypes x mask combos per
head_dim).

Validation:
  * Correctness suite: 241/245 (same 4 pre-existing int32-overflow
    failures in the prefill rebased-pointer path).
  * d=128 GQA-8 (a NEW combo we never had a binary for) -- runs
    correctly on the existing decode_d128_m* binaries with num_qpkv=8
    at runtime. max abs diff <= 1e-2 vs the torch reference at ql in
    {1, 4, 16}.
  * d=64 MHA (also a new combo) -- runs correctly on the existing
    decode_d64_m* binaries with num_qpkv=1. Same tolerance.
  * Perf sweep (b=4..256, sk=120000, MI300):
      d=64  GQA-8: speedups 1.28x..1.84x vs Triton (within 0.6%
                   of baseline).
      d=128 MHA:   speedups 0.98x..1.14x vs Triton (within 0.3%
                   of baseline).

Unlocked: adding new (head_dim, num_qpkv) combos no longer requires
new kernel binaries -- just a host-side heuristic update mapping the
combo to the appropriate (kBlockM, BlockWarps) ladder.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-05-12 12:15:55 +00:00
2018-10-08 22:49:58 -05:00
2025-01-07 08:29:40 -08:00
2026-01-14 07:31:45 -08:00
2025-07-24 12:38:24 -07:00

Composable Kernel

Note

The published documentation is available at Composable Kernel in an organized, easy-to-read format, with search and a table of contents. The documentation source files reside in the docs folder of this repository. As with all ROCm projects, the documentation is open source. For more information on contributing to the documentation, see Contribute to ROCm documentation.

The Composable Kernel (CK) library provides a programming model for writing performance-critical kernels for machine learning workloads across multiple architectures (GPUs, CPUs, etc.). The CK library uses general purpose kernel languages, such as HIP C++.

CK uses two concepts to achieve performance portability and code maintainability:

  • A tile-based programming model
  • Algorithm complexity reduction for complex machine learning (ML) operators. This uses an innovative technique called Tensor Coordinate Transformation.

ALT

The current CK library is structured into four layers:

  • Templated Tile Operators
  • Templated Kernel and Invoker
  • Instantiated Kernel and Invoker
  • Client API

ALT

General information

CK is released under the MIT license.

Building CK

We recommend building CK inside Docker containers, which include all necessary packages. Pre-built Docker images are available on DockerHub.

  1. To build a new Docker image, use the Dockerfile provided with the source code:

    DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile .
    
  2. Launch the Docker container:

    docker run                                     \
    -it                                            \
    --privileged                                   \
    --group-add sudo                               \
    -w /root/workspace                             \
    -v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace  \
    ck:latest                                      \
    /bin/bash
    
  3. Clone CK source code from the GitHub repository and start the build:

    git clone https://github.com/ROCm/composable_kernel.git && \
    cd composable_kernel && \
    mkdir build && \
    cd build
    

    You must set the GPU_TARGETS macro to specify the GPU target architecture(s) you want to run CK on. You can specify single or multiple architectures. If you specify multiple architectures, use a semicolon between each; for example, gfx908;gfx90a;gfx942.

    cmake                                                                                             \
    -D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
    -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
    -D CMAKE_BUILD_TYPE=Release                                                                       \
    -D GPU_TARGETS="gfx908;gfx90a"                                                                    \
    ..
    

    If you don't set GPU_TARGETS on the cmake command line, CK is built for all GPU targets supported by the current compiler (this may take a long time). Tests and examples will only get built if the GPU_TARGETS is set by the user on the cmake command line.

    NOTE: If you try setting GPU_TARGETS to a list of architectures, the build will only work if the architectures are similar, e.g., gfx908;gfx90a, or gfx1100;gfx1101;gfx11012. Otherwise, if you want to build the library for a list of different architectures, you should use the GPU_ARCHS build argument, for example GPU_ARCHS=gfx908;gfx1030;gfx1100;gfx942.

    Convenience script for development builds:

    Alternatively, you can use the provided convenience script script/cmake-ck-dev.sh which automatically configures CK for development with sensible defaults. In the build directory:

    ../script/cmake-ck-dev.sh
    

    This script:

    • Cleans CMake cache files before configuring
    • Sets BUILD_DEV=ON for development mode
    • Defaults to GPU targets: gfx908;gfx90a;gfx942
    • Enables verbose makefile output
    • Sets additional compiler flags for better error messages

    By default, it considers the parent directory to be the project source directory.

    You can specify the source directory as the first argument. You can specify custom GPU targets (semicolon-separated) as the second argument:

    ../script/cmake-ck-dev.sh .. gfx1100
    

    Or pass additional cmake arguments:

    ../script/cmake-ck-dev.sh .. gfx90a -DCMAKE_BUILD_TYPE=Release
    
  4. Build the entire CK library:

    make -j"$(nproc)"
    
  5. Install CK:

    make -j install
    

    See Note on -j

Building for Windows

Install TheRock and run CMake configure as

    cmake                                                                                      \
    -D CMAKE_PREFIX_PATH="C:/dist/TheRock"                                                     \
    -D CMAKE_CXX_COMPILER="C:/dist/TheRock/bin/hipcc.exe"                                      \
    -D CMAKE_BUILD_TYPE=Release                                                                \
    -D GPU_TARGETS="gfx1151"                                                                   \
    -G Ninja                                                                                   \
    ..

Use Ninja to build either the whole library or individual targets.

Optional post-install steps

  • Build examples and tests:

    make -j examples tests
    
  • Build and run all examples and tests:

    make -j check
    

    You can find instructions for running each individual example in example.

  • Build and run smoke/regression examples and tests:

    make -j smoke # tests and examples that run for < 30 seconds each
    
    make -j regression # tests and examples that run for >= 30 seconds each
    
  • Build ckProfiler:

    make -j ckProfiler
    

    You can find instructions for running ckProfiler in profiler.

  • Build our documentation locally:

    cd docs
    pip3 install -r sphinx/requirements.txt
    python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
    

Notes

The -j option for building with multiple threads in parallel, which speeds up the build significantly. However, -j launches unlimited number of threads, which can cause the build to run out of memory and crash. On average, you should expect each thread to use ~2Gb of RAM. Depending on the number of CPU cores and the amount of RAM on your system, you may want to limit the number of threads. For example, if you have a 128-core CPU and 128 Gb of RAM it's advisable to use -j32.

Additional cmake flags can be used to significantly speed-up the build:

  • DTYPES (default is not set) can be set to any subset of "fp64;fp32;tf32;fp16;fp8;bf16;int8" to build instances of select data types only. The main default data types are fp32 and fp16; you can safely skip other data types.

  • DISABLE_DL_KERNELS (default is OFF) must be set to ON in order not to build instances, such as gemm_dl or batched_gemm_multi_d_dl. These instances are useful on architectures like the NAVI2x, as most other platforms have faster instances, such as xdl or wmma, available.

  • DISABLE_DPP_KERNELS (default is OFF) must be set to ON in order not to build instances, such as gemm_dpp. These instances offer a slightly better performance of fp16 gemms on NAVI2x. But on other architectures faster alternatives are available.

  • CK_USE_FP8_ON_UNSUPPORTED_ARCH (default is OFF) must be set to ON in order to build instances, such as gemm_universal, gemm_universal_streamk and gemm_multiply_multiply for fp8 data type for GPU targets which do not have native support for fp8 data type, such as gfx908 or gfx90a. These instances are useful on architectures like the MI100/MI200 for the functional support only.

Using sccache for building

The default CK Docker images come with a pre-installed version of sccache, which supports clang being used as hip-compiler (" -x hip"). Using sccache can help reduce the time to re-build code from hours to 1-2 minutes. In order to invoke sccache, you need to run:

 sccache --start-server

then add the following flags to the cmake command line:

 -DCMAKE_HIP_COMPILER_LAUNCHER=sccache -DCMAKE_CXX_COMPILER_LAUNCHER=sccache -DCMAKE_C_COMPILER_LAUNCHER=sccache

You may need to clean up the build folder and repeat the cmake and make steps in order to take advantage of the sccache during subsequent builds.

Using CK as pre-built kernel library

You can find instructions for using CK as a pre-built kernel library in client_example.

Contributing to CK

When you contribute to CK, make sure you run clang-format on all changed files. We highly recommend using git hooks that are managed by the pre-commit framework. To install hooks, run:

sudo script/install_precommit.sh

With this approach, pre-commit adds the appropriate hooks to your local repository and automatically runs clang-format (and possibly additional checks) before any commit is created.

If you need to uninstall hooks from the repository, you can do so by running the following command:

script/uninstall_precommit.sh

If you need to temporarily disable pre-commit hooks, you can add the --no-verify option to the git commit command.

Description
[DEPRECATED] Moved to ROCm/rocm-libraries repo. NOTE: develop branch is maintained as a read-only mirror
Readme MIT Cite this repository 241 MiB
Languages
C++ 91.3%
Python 6.1%
CMake 1.6%
Shell 0.5%
Pawn 0.2%
Other 0.1%