aboutsummaryrefslogtreecommitdiff
Commit message (Expand)AuthorAgeFilesLines
* cuda: add fast GEMM mode (--cuda-fast-gemm mode)Onur Gündüz2026-06-143-17/+102
* cuda: clarify comment gemm_computeOnur Gündüz2026-06-141-2/+3
* cuda: add opt-in BF16 tensor-core GEMM modeOnur Gündüz2026-06-141-21/+34
* cuda: replace long-seq attention with cuBLAS GEMM attentionOnur Gündüz2026-06-141-158/+124
* cuda: replace tiled attention with warp-per-query reductionOnur Gündüz2026-06-141-98/+105
* cuda: extend tiled attention to ragged micro-batchesOnur Gündüz2026-06-141-3/+105
* cuda: add tiled single attentionOnur Gündüz2026-06-141-4/+64
* cuda: optimize rmsnorm reductionOnur Gündüz2026-06-141-34/+29
* cuda: add server contextual pathOnur Gündüz2026-06-145-21/+239
* cuda: optimize projection pathOnur Gündüz2026-06-141-87/+118
* cuda: fuse qkv projectionOnur Gündüz2026-06-141-44/+78
* cuda: accumulate residual projectionsOnur Gündüz2026-06-141-21/+18
* cuda: optimize attention baselineOnur Gündüz2026-06-141-44/+47
* cuda: add initial CUDA embedding backendOnur Gündüz2026-06-146-68/+911
* kernels: fix OpenBLAS thread contention and numerical instabilityOnur Gündüz2026-06-142-3/+33
* mlx: split ragged contextual batchesOnur Gündüz2026-06-141-3/+137
* mlx: tune ragged grouping for 4BOnur Gündüz2026-06-141-5/+34
* mlx: prepare contiguous linear weightsOnur Gündüz2026-06-141-4/+21
* mlx: optimize ragged embedding batchesOnur Gündüz2026-06-141-2/+114
* build: add address/undefined sanitizers to debug buildsOnur Gündüz2026-06-141-2/+4
* chore: address static analyzer warningsOnur Gündüz2026-06-145-27/+92
* kernels: parallelize large rmsnorm rowsOnur Gündüz2026-06-141-5/+48
* mlx: use prefix pooling for f32 contextual spansOnur Gündüz2026-06-141-32/+138
* mlx: cache transposed linear weightsOnur Gündüz2026-06-141-4/+46
* server: expose worker stage timingOnur Gündüz2026-06-141-13/+95
* tokenizer: add reusable encode workspaceOnur Gündüz2026-06-144-2/+478
* mlx: keep late maxsim device residentOnur Gündüz2026-06-142-25/+405
* late: add packed maxsim scoringOnur Gündüz2026-06-144-7/+45
* late: add mlx token vector supportOnur Gündüz2026-06-142-8/+235
* late: add CPU token vectors and maxsimOnur Gündüz2026-06-142-5/+223
* all: add validation for late-interaction model compatOnur Gündüz2026-06-142-1/+48
* kernels: vectorize bf16 widening on armOnur Gündüz2026-06-141-0/+14
* kernels: use sgemv for single-row f32 linearOnur Gündüz2026-06-142-4/+11
* kernels: add neon rope apply pathOnur Gündüz2026-06-141-0/+21
* build: add tmp to .gitignoreOnur Gündüz2026-06-141-0/+1
* dist: add fp16 activation transportOnur Gündüz2026-06-143-23/+191
* server: support http keep-aliveOnur Gündüz2026-06-141-11/+98
* mlx: add opt-in 8-bit weight quantizationOnur Gündüz2026-06-147-33/+332
* cpu: reuse tiled attention scratch buffersOnur Gündüz2026-06-143-15/+77
* mlx: harden model and buffer validationOnur Gündüz2026-06-141-46/+156
* server: bound mlx batches by padded token rowsOnur Gündüz2026-06-141-2/+25
* mlx: simplify equal-length embedding poolingOnur Gündüz2026-06-141-24/+35
* mlx: pool contextual bf16 spans before wideningOnur Gündüz2026-06-141-11/+9
* server: reject unsafe mlx model overcommitOnur Gündüz2026-06-145-0/+129
* server: simplify context separator logicOnur Gündüz2026-06-142-13/+5
* server: batch contextual document inferenceOnur Gündüz2026-06-145-226/+567
* mlx: pool contextual spans on deviceOnur Gündüz2026-06-141-17/+95
* cpu: tile packed attention through blasOnur Gündüz2026-06-141-54/+110
* server: batch standard requests by token budgetOnur Gündüz2026-06-143-11/+77
* dist: add direct two-node fp32 inferenceOnur Gündüz2026-06-144-19/+874