index
:
ffwd
master
Fastforward embedding inference engine
about
summary
refs
log
tree
commit
diff
log msg
author
committer
range
Commit message (
Expand
)
Author
Age
Files
Lines
*
cuda: add fast GEMM mode (--cuda-fast-gemm mode)
Onur Gündüz
2026-06-14
3
-17
/
+102
*
cuda: clarify comment gemm_compute
Onur Gündüz
2026-06-14
1
-2
/
+3
*
cuda: add opt-in BF16 tensor-core GEMM mode
Onur Gündüz
2026-06-14
1
-21
/
+34
*
cuda: replace long-seq attention with cuBLAS GEMM attention
Onur Gündüz
2026-06-14
1
-158
/
+124
*
cuda: replace tiled attention with warp-per-query reduction
Onur Gündüz
2026-06-14
1
-98
/
+105
*
cuda: extend tiled attention to ragged micro-batches
Onur Gündüz
2026-06-14
1
-3
/
+105
*
cuda: add tiled single attention
Onur Gündüz
2026-06-14
1
-4
/
+64
*
cuda: optimize rmsnorm reduction
Onur Gündüz
2026-06-14
1
-34
/
+29
*
cuda: add server contextual path
Onur Gündüz
2026-06-14
5
-21
/
+239
*
cuda: optimize projection path
Onur Gündüz
2026-06-14
1
-87
/
+118
*
cuda: fuse qkv projection
Onur Gündüz
2026-06-14
1
-44
/
+78
*
cuda: accumulate residual projections
Onur Gündüz
2026-06-14
1
-21
/
+18
*
cuda: optimize attention baseline
Onur Gündüz
2026-06-14
1
-44
/
+47
*
cuda: add initial CUDA embedding backend
Onur Gündüz
2026-06-14
6
-68
/
+911
*
kernels: fix OpenBLAS thread contention and numerical instability
Onur Gündüz
2026-06-14
2
-3
/
+33
*
mlx: split ragged contextual batches
Onur Gündüz
2026-06-14
1
-3
/
+137
*
mlx: tune ragged grouping for 4B
Onur Gündüz
2026-06-14
1
-5
/
+34
*
mlx: prepare contiguous linear weights
Onur Gündüz
2026-06-14
1
-4
/
+21
*
mlx: optimize ragged embedding batches
Onur Gündüz
2026-06-14
1
-2
/
+114
*
build: add address/undefined sanitizers to debug builds
Onur Gündüz
2026-06-14
1
-2
/
+4
*
chore: address static analyzer warnings
Onur Gündüz
2026-06-14
5
-27
/
+92
*
kernels: parallelize large rmsnorm rows
Onur Gündüz
2026-06-14
1
-5
/
+48
*
mlx: use prefix pooling for f32 contextual spans
Onur Gündüz
2026-06-14
1
-32
/
+138
*
mlx: cache transposed linear weights
Onur Gündüz
2026-06-14
1
-4
/
+46
*
server: expose worker stage timing
Onur Gündüz
2026-06-14
1
-13
/
+95
*
tokenizer: add reusable encode workspace
Onur Gündüz
2026-06-14
4
-2
/
+478
*
mlx: keep late maxsim device resident
Onur Gündüz
2026-06-14
2
-25
/
+405
*
late: add packed maxsim scoring
Onur Gündüz
2026-06-14
4
-7
/
+45
*
late: add mlx token vector support
Onur Gündüz
2026-06-14
2
-8
/
+235
*
late: add CPU token vectors and maxsim
Onur Gündüz
2026-06-14
2
-5
/
+223
*
all: add validation for late-interaction model compat
Onur Gündüz
2026-06-14
2
-1
/
+48
*
kernels: vectorize bf16 widening on arm
Onur Gündüz
2026-06-14
1
-0
/
+14
*
kernels: use sgemv for single-row f32 linear
Onur Gündüz
2026-06-14
2
-4
/
+11
*
kernels: add neon rope apply path
Onur Gündüz
2026-06-14
1
-0
/
+21
*
build: add tmp to .gitignore
Onur Gündüz
2026-06-14
1
-0
/
+1
*
dist: add fp16 activation transport
Onur Gündüz
2026-06-14
3
-23
/
+191
*
server: support http keep-alive
Onur Gündüz
2026-06-14
1
-11
/
+98
*
mlx: add opt-in 8-bit weight quantization
Onur Gündüz
2026-06-14
7
-33
/
+332
*
cpu: reuse tiled attention scratch buffers
Onur Gündüz
2026-06-14
3
-15
/
+77
*
mlx: harden model and buffer validation
Onur Gündüz
2026-06-14
1
-46
/
+156
*
server: bound mlx batches by padded token rows
Onur Gündüz
2026-06-14
1
-2
/
+25
*
mlx: simplify equal-length embedding pooling
Onur Gündüz
2026-06-14
1
-24
/
+35
*
mlx: pool contextual bf16 spans before widening
Onur Gündüz
2026-06-14
1
-11
/
+9
*
server: reject unsafe mlx model overcommit
Onur Gündüz
2026-06-14
5
-0
/
+129
*
server: simplify context separator logic
Onur Gündüz
2026-06-14
2
-13
/
+5
*
server: batch contextual document inference
Onur Gündüz
2026-06-14
5
-226
/
+567
*
mlx: pool contextual spans on device
Onur Gündüz
2026-06-14
1
-17
/
+95
*
cpu: tile packed attention through blas
Onur Gündüz
2026-06-14
1
-54
/
+110
*
server: batch standard requests by token budget
Onur Gündüz
2026-06-14
3
-11
/
+77
*
dist: add direct two-node fp32 inference
Onur Gündüz
2026-06-14
4
-19
/
+874
[next]