index
:
ffwd
master
Fastforward embedding inference engine
about
summary
refs
log
tree
commit
diff
log msg
author
committer
range
path:
root
/
embed_cuda.cu
(
follow
)
Commit message (
Collapse
)
Author
Age
Files
Lines
*
chore: move sources to src/ with proper names, drop qwen_ prefix
Onur Gündüz
2026-06-16
1
-2052
/
+0
|
*
cuda: memory-only bf16 for BERT (f32 compute on bf16 weights)
Onur Gündüz
2026-06-16
1
-6
/
+46
|
*
all: support BERT tanh-GeLU (gelu_new / gelu_pytorch_tanh)
Onur Gündüz
2026-06-16
1
-4
/
+18
|
*
cuda: add the BERT encoder family forward and loader
Onur Gündüz
2026-06-16
1
-9
/
+238
|
*
all: add CLS first-token pooling
Onur Gündüz
2026-06-15
1
-0
/
+13
|
*
all: support GTE-Qwen2 embeddings
Onur Gündüz
2026-06-15
1
-17
/
+55
|
*
late: add CUDA late-interaction rerank path (GPU encode, host MaxSim)
Onur Gündüz
2026-06-15
1
-4
/
+219
|
*
chore: run clang-format
Onur Gündüz
2026-06-15
1
-442
/
+428
|
*
all: add support for Qwen3-Embedding + causal GQA attention
Onur Gündüz
2026-06-14
1
-10
/
+91
|
*
chore: rename project
Onur Gündüz
2026-06-14
1
-55
/
+55
|
*
cuda: restore BF16 attention internals, length-route batched attention
Onur Gündüz
2026-06-14
1
-42
/
+130
|
*
cuda: drop BF16 attention internals, keep fused rope + batched F32 attention
Onur Gündüz
2026-06-14
1
-125
/
+42
|
*
cuda: pointer-array batched attention for equal-length micro-batches
Onur Gündüz
2026-06-14
1
-8
/
+167
|
*
cuda: fuse QK norm+rope launches, BF16 attention output path
Onur Gündüz
2026-06-14
1
-45
/
+99
|
*
cuda: follow model file dtype for weight storage by default
Onur Gündüz
2026-06-14
1
-1
/
+15
|
*
chore: rename sources; binaries pplx-embed, pplx-embed-server
Onur Gündüz
2026-06-14
1
-2
/
+2
|
*
gitignore: track new binary and library names
Onur Gündüz
2026-06-14
1
-0
/
+1204