feat(ck-tile): add block-scale GEMM operators (aquant,
bquant, abquant) (#8519)
JIRA ID - AICK-1289
Motivation
Adds three new block-scale quantized GEMM operators to the CK Tile
Engine for FP8/BF8 inference workloads.
Technical Details
gemm_aquant: A-matrix quantized GEMM with per-row-group scale tensor [M,
K/group_size_k]
gemm_bquant: B-matrix quantized GEMM with per-column-group scale tensor
[K/group_size_k, N]
gemm_abquant: Both A and B quantized with independent group-scale
tensors
Each operator includes CMakeLists, Python instance builder with tier
sampling, C++ benchmark/profiler with host reference verification, and
config JSONs. Supporting changes to gemm_instance_builder.py,
gemm_validation_utils.py, sampling infra, and the operation support
matrix.
Test Plan
Build and run all three operators with fp8/bf8 on gfx942/gfx950
Verify correctness against CPU reference
Verify CI config builds pass
[CK Tile Engine] Daily tier sampling for tile engine GEMM (#7311)
Summary
- Replace uniform random instance sampling (random.shuffle) with
scrambled Sobol + Latin Hypercube + maximin space-filling
sampling, per the Tile Engine Benchmark Sampling RFC
- Add op-weighted budget allocation via new
TILE_ENGINE_SAMPLING_TIER=daily CMake knob that auto-distributes 8,000
instances across
ops proportional to registered weights in op_weights.json
- Emit chosen_instances.json manifests for reproducibility tracking
- Consolidate 5 copies of sampling logic into single _apply_sampling()
method on the base class
Jenkinsfile changes
Replace per-op -D *_MAX_INSTANCES=250 with single -D
TILE_ENGINE_SAMPLING_TIER=daily in gfx942/gfx950/gfx1201 stages. Budget
auto-distributes (8000 total per GPU target).
---------
Co-authored-by: Claude Sonnet 4 <noreply@anthropic.com>
