* Split FMHA decode and GroupGemm template instantiations into per-kernel compilation units (#140) Split the monolithic template instantiation of xe_fmha_fwd_decode_runner.hpp into 72 separate .cpp files (one per QG_SZ × HEAD_DIM × PAGE_SIZE combination), each compiled as its own library. This enables parallel compilation and significantly speeds up build times. Changes: - Create xe_fmha_fwd_decode_kernel.cpp.in template for per-combination compilation - Create xe_fmha_fwd_decode_dispatch.hpp with function declarations for all 72 kernels - Move decode::mha_fwd() from header to flash_attention.cpp with dispatch table - Update src/CMakeLists.txt to generate .cpp files via configure_file() - Remove mha_fwd() definition from xe_fmha_fwd_decode_runner.hpp header Co-authored-by: airMeng <39229107+airMeng@users.noreply.github.com> Co-authored-by: jiwei1.sun <jiwei1.sun@intel.com> * Fix noncontiguous input for rmsnorm (#117) * fix norm with noncontiguous input * remove comment out test * support in kernel * Add MXFP4 Per Token Group Quant kernel and tests (#106) * Add MXFP4 Per Token Group Quant kernel and tests Remove commented out fp8 blockwise group gemm registration * Add benchmarking for per token group quant mxfp4 * Add test to run_suite.py * Fix group size constraint for mxfp4; Add benchmark test to CI flow * Remove reference provider from the benchmark script - Add check for quantized and scale values separately - Include eps value in ref quant function call * Fix MXFP4 quantization to match OCP MX spec - Replace ceil(log2(max/6.0)) scale computation with floor(log2(max)) - E2M1_EMAX per OCP MX spec - Fix roundTiesToEven at midpoints in SYCL kernel (change <= to < at odd-mantissa boundaries) - Replace naive argmin-based quantize_to_e2m1 reference with microxcaling _quantize_elemwise_core algorithm - Normalize signed zeros (+0.0 vs -0.0) before packed byte comparison in tests and benchmark * Fix lint issues * Remove unsupported group sizes * Apply formatting check * Add TODO for quantize_to_e2m1 * Update tests/test_per_token_group_quant_mxfp4.py Co-authored-by: Meng, Hengyu <airdldl@163.com> * trigger CI --------- Co-authored-by: Meng, Hengyu <airdldl@163.com> * add page 64 * Initial plan * add reduce.h * add XeFMHAFwdSplitKVKernel * const tensor for Q * add split kernel * save * cache_seqlens * head_dim =128 * 2026 * test for mingxu * Initial plan * Rebase onto main: integrate split-KV changes into flash_attention.cpp and fix SplitDeodeConfig typo * Rebase onto updated split_kv_decode: fix FMHAConfig undefined, add conditional dispatch between DecodeConfig and SplitDecodeConfig, comment out page_size 32 Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com> Agent-Logs-Url: https://github.com/sgl-project/sgl-kernel-xpu/sessions/82c107d1-3f61-4ce0-9444-d7f19f27a292 * bugfix * Refactor dispatch to function pointer tables following GroupGemmXe20 pattern Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com> Agent-Logs-Url: https://github.com/sgl-project/sgl-kernel-xpu/sessions/5c058f18-d408-4752-aba6-2625a6e5e638 * Refactor decode dispatch to struct operator() following GroupGemmXe20 pattern Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com> Agent-Logs-Url: https://github.com/sgl-project/sgl-kernel-xpu/sessions/871c6171-1314-44d4-b4f4-007ddffc6ed4 * Replace function pointer table with direct struct operator() dispatch (GroupGemmXe20 pattern) Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com> Agent-Logs-Url: https://github.com/sgl-project/sgl-kernel-xpu/sessions/85877884-e91a-400f-adcc-5b9a55692df9 * Add use_sink and use_causal_mask to Arguments; remove bool use_sink from operator() signature Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com> Agent-Logs-Url: https://github.com/sgl-project/sgl-kernel-xpu/sessions/f88d0fec-8f93-4f48-99b8-714d68fd14f4 * Replace non-ASCII em-dash in flash_attention.cpp comment with ASCII hyphen Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com> Agent-Logs-Url: https://github.com/sgl-project/sgl-kernel-xpu/sessions/8cc49274-a20f-42e3-aad5-39043ba2eefa --------- Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com> Co-authored-by: airMeng <39229107+airMeng@users.noreply.github.com> Co-authored-by: jiwei1.sun <jiwei1.sun@intel.com> Co-authored-by: Yang Wang <yang3.wang@intel.com> Co-authored-by: Suryaprakash Shanmugam <suryaprakash.shanmugam@intel.com> Co-authored-by: Meng, Hengyu <airdldl@163.com> Co-authored-by: sunjiweiswift <16934286+sunjiweiswift@users.noreply.github.com>

Commits

Commits on Mar 18, 2026

Commits on Mar 23, 2026

Commits on Mar 24, 2026

Commits on Mar 31, 2026

Commits on Apr 2, 2026

Commits on Apr 3, 2026

Commits on Apr 7, 2026

Commits on Apr 8, 2026