c8a8449eec
[CK] Add split-K support for ABQuantGrouped in block_scale_gemm (#4816) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## Changes ### Split-K support in `gemm_quant_kernel.hpp` - **`SplitKBatchOffset`**: Added `aq_group_offset` and `aq_k_split_offset` fields (mirroring the existing `bq_*` fields for B) to track each split-K batch's position within the AQ scale tensor. For `ABQuantGrouped`, both offsets are computed from `k_id * KRead` divided by `AQuantGroupSize::kK`. - **`MakeAQBlockWindow`**: Added an `aq_group_offset` parameter (defaulting to 0 for non-split-K paths) so the AQ tensor view's K-group dimension reflects only the remaining K-groups from the split-K offset, consistent with how `MakeBQBlockWindow` handles the BQ tensor. - **`RunGemm`**: Threads the `aq_k_split_offset` through to `MakeAQBlockWindow` when in split-K mode. ### Constraints in `IsSupportedArgument()` Four constraints gate split-K (`k_batch > 1`) for ABQuantGrouped: 1. **Mode check** — split-K is only allowed for `BQuantGrouped` (no preshuffle) or `ABQuantGrouped` (no `APreshuffleQuant`). Any other quant mode with `k_batch > 1` returns `false`. 2. **B quant group alignment** — `KRead` (per-batch K slice) must be divisible by `BQuantGroupSize::kK`. Each batch must operate on complete B quantization groups; a partial group would require splitting a scale value across batches. 3. **A quant group alignment** (new, ABQuantGrouped only) — `KRead` must also be divisible by `AQuantGroupSize::kK` for the same reason applied to the AQ scale tensor. 4. **Minimum 2 K-tile iterations per batch** (new) — The software-pipelined GEMM kernels (CompV3 family) prefetch one tile ahead, so they require `per_batch_num_loop = KRead / KPerBlock >= 2`. When `KRead == KPerBlock` (i.e. each batch is exactly one tile), the prefetch reads into the next batch's memory region and produces incorrect results. Configurations where `K == k_batch * KPerBlock` are therefore rejected. ### Example update (`run_gemm_quant_example.inc`) Updated the comment above the `IsSupportedArgument` call to document that split-K is now supported for both `BQuantGrouped` (no preshuffle) and `ABQuantGrouped` (no `APreshuffleQuant`). ## Unit Tests Two new test files covering decode and prefill tile shapes across a range of `k_batch` values (2–8), data types (FP8, BF8), and quantization group sizes (1×1×128 and 1×128×128 for B): - `test_gemm_quant_abquant_splitk_decode.cpp` — uses the decode tile shape (M=16, N=64, K_tile=256) - `test_gemm_quant_abquant_splitk_prefill.cpp` — uses the prefill tile shape (M=128, N=128, K_tile=128) Each test calls `run_test_with_validation` which runs the kernel and checks correctness against a CPU reference. Configurations excluded from tests are annotated with comments explaining which constraint they violate (typically the `per_batch_num_loop >= 2` requirement). ## Prerequisites This PR depends on #4429, which must be merged before this can be merged.
Composable Kernel Tile
concept
ck_tile provides a programming model with templated abstractions to enable users to implement performance-critical kernels for machine learning workloads. introduces following basic concepts to help users building your own operator
- tensor coordinate transformation, this is the core concept of layout/index transform abstraction in both compiler time and run time.
- tile-based programming model, including tile-level api and the concept of distributed tensor.
ck_tile is independently from the old ck, located under /include/ck_tile. You don't need to include anything from old CK, ck_tile has similiar (indeed almost the same) implementations for users to build operators. We will have a transition period to pull everything from old ck into ck_tile, stay tuned.
component
ck_tile is splitted into several componenets including core, host, ops/gemm, ops/fmha... each component you only need to include a single header (e.g #include "ck_tile/core.hpp", #include "ck_tile/ops/fmha.hpp") then you are able to use the function/structure inside (different from old ck)
[core]
ck_tile/core contains all the basic data structure and function to build the kernel, you can only include this header and build your own operators that utilizing all the basic building blocks introduced in ck.
core/container
- array, store runtime variables with fixed length (tensor index, register buffer, etc...)
- tuple, same as std::tuple, hold different type of data, and one of the solution to achieve multiple buffer.
- sequence, compile time integer sequence used to build various internal structures, or to describe tile size
- other convenient structure build on top of above 3
core/numeric
- gpu data type like
fp16_t,bf16_t,fp8_t... and the conversion between each other - constexpr integer similiar to std::integral_constant to be used as compile time integer.
- math functions and numeric utilities
core/algorithm
- coordinate transformation system, used to build tensor transform and compile time indexing. This is the core idea introduced in old
ckto describe how a tensor is build by several basic transform primitives likemerge/unmerge/embedetc... and how we indexing into a ND tensor that finally mapped to 1D memory offset.
core/tensor
- tensor descriptor, to describe how a ND tensor
- distributed tensor, describe the storage of this tensor, and the distribution of how a collection of threads collaborately work for this tensor.
- tile level API, including
load_tile,store_tile,shuffle_tile,slice_tile, etc...
[host]
ck_tile/host contains all the host side utilities to launch a kernel, create the device buffer, and some reference implementations. This can be used to create examples (like that under ck_tile example folder) and simple executable to invoke this kernel, so if you only need ck_tile to build your own device library then it's OK to not include this. Based on this, it is recommended to include the specific header you needed under this folder to avoid including unwanted headers (e.g, only include ck_tile/host/kernel_launch.hpp), unless you are writing a host executable.
[ops/gemm, ops/fmha, ops/reduce...]
our implementation of different device operators.
- warp, warp tile level operator
- block, block tile level operator
- pipeline, pipeline that can achieve a customized tile level mainloop (or epilogue). By switching different pipeline to the kernel template you can have different kind of pipeline optimizations.
- kernel, template interface for users to instantiate a particular kernel
[ops/epilogue]
epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues.
[ref]
reference implementation of cpu or gpu. This folder is supposed to include a specific header on demand.
examples
currently we put all ck_tile related example under /example/ck_tile folder. Please check each example's subfolder.