index
:
ffwd
master
Fastforward embedding inference engine
about
summary
refs
log
tree
commit
diff
log msg
author
committer
range
path:
root
/
src
/
cuda.cu
(
unfollow
)
Commit message (
Expand
)
Author
Files
Lines
2026-06-21
all: split Makefile and reorg library/tools
Onur Gündüz
1
-2056
/
+0
2026-06-16
all: add SentencePiece XLM-R support
Onur Gündüz
1
-3
/
+7
2026-06-16
chore: move sources to src/ with proper names, drop qwen_ prefix
Onur Gündüz
1
-3
/
+3
2026-06-16
cuda: memory-only bf16 for BERT (f32 compute on bf16 weights)
Onur Gündüz
1
-6
/
+46
2026-06-16
all: support BERT tanh-GeLU (gelu_new / gelu_pytorch_tanh)
Onur Gündüz
1
-4
/
+18
2026-06-16
cuda: add the BERT encoder family forward and loader
Onur Gündüz
1
-9
/
+238
2026-06-15
all: add CLS first-token pooling
Onur Gündüz
1
-0
/
+13
2026-06-15
all: support GTE-Qwen2 embeddings
Onur Gündüz
1
-17
/
+55
2026-06-15
late: add CUDA late-interaction rerank path (GPU encode, host MaxSim)
Onur Gündüz
1
-4
/
+219
2026-06-15
chore: run clang-format
Onur Gündüz
1
-442
/
+428
2026-06-14
all: add support for Qwen3-Embedding + causal GQA attention
Onur Gündüz
1
-10
/
+91
2026-06-14
chore: rename project
Onur Gündüz
1
-55
/
+55
2026-06-14
cuda: restore BF16 attention internals, length-route batched attention
Onur Gündüz
1
-42
/
+130
2026-06-14
cuda: drop BF16 attention internals, keep fused rope + batched F32 attention
Onur Gündüz
1
-125
/
+42
2026-06-14
cuda: pointer-array batched attention for equal-length micro-batches
Onur Gündüz
1
-8
/
+167
2026-06-14
cuda: fuse QK norm+rope launches, BF16 attention output path
Onur Gündüz
1
-45
/
+99
2026-06-14
cuda: follow model file dtype for weight storage by default
Onur Gündüz
1
-1
/
+15
2026-06-14
chore: rename sources; binaries pplx-embed, pplx-embed-server
Onur Gündüz
1
-2
/
+2
2026-06-14
gitignore: track new binary and library names
Onur Gündüz
1
-0
/
+0
2026-06-14
cli: rename flags, add --stream, split serving into pplx-embed-server
Onur Gündüz
1
-2
/
+2
2026-06-14
cuda: store BF16 activations directly in norm and SiLU kernels
Onur Gündüz
1
-51
/
+114
2026-06-14
cuda: add optional BF16 CUDA weight storage (--cuda-weights bf16)
Onur Gündüz
1
-68
/
+131
2026-06-14
cuda: unify attention on one path; drop the max_seq>=64 dispatch
Onur Gündüz
1
-73
/
+8
2026-06-14
cuda: add fast GEMM mode (--cuda-fast-gemm mode)
Onur Gündüz
1
-17
/
+75
2026-06-14
cuda: clarify comment gemm_compute
Onur Gündüz
1
-2
/
+3
2026-06-14
cuda: add opt-in BF16 tensor-core GEMM mode
Onur Gündüz
1
-21
/
+34
2026-06-14
cuda: replace long-seq attention with cuBLAS GEMM attention
Onur Gündüz
1
-158
/
+124
2026-06-14
cuda: replace tiled attention with warp-per-query reduction
Onur Gündüz
1
-98
/
+105
2026-06-14
cuda: extend tiled attention to ragged micro-batches
Onur Gündüz
1
-3
/
+105
2026-06-14
cuda: add tiled single attention
Onur Gündüz
1
-4
/
+64
2026-06-14
cuda: optimize rmsnorm reduction
Onur Gündüz
1
-34
/
+29
2026-06-14
cuda: add server contextual path
Onur Gündüz
1
-14
/
+159
2026-06-14
cuda: optimize projection path
Onur Gündüz
1
-87
/
+118
2026-06-14
cuda: fuse qkv projection
Onur Gündüz
1
-44
/
+78
2026-06-14
cuda: accumulate residual projections
Onur Gündüz
1
-21
/
+18
2026-06-14
cuda: optimize attention baseline
Onur Gündüz
1
-44
/
+47
2026-06-14
cuda: add initial CUDA embedding backend
Onur Gündüz
1
-0
/
+731