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: stage compact GQA K/V in qwen attention, halve the expand
Onur Gündüz
2026-06-25
1
-14
/
+47
*
chore: fix permissions for test scripts and devtools
Onur Gündüz
2026-06-25
9
-0
/
+0
*
chore: run clang-format
Onur Gündüz
2026-06-25
20
-432
/
+380
*
build: update build info and accelerator labels
Onur Gündüz
2026-06-25
5
-34
/
+30
*
server: store the Server-Timing header inline in the job
Onur Gündüz
2026-06-25
2
-10
/
+11
*
server: encode base64 embeddings directly into the response buffer
Onur Gündüz
2026-06-25
5
-67
/
+97
*
server: drop the throwaway buffer in the base64 float path
Onur Gündüz
2026-06-25
1
-8
/
+4
*
server: remove the dead auth field
Onur Gündüz
2026-06-25
2
-2
/
+0
*
server: move the request body instead of copying it
Onur Gündüz
2026-06-25
1
-2
/
+3
*
chore: update README
Onur Gündüz
2026-06-25
1
-13
/
+20
*
all: migrate to yyjson
Onur Gündüz
2026-06-25
19
-323
/
+354
*
cuda: skip per-warp future key tiles in causal qwen attention
Onur Gündüz
2026-06-25
1
-65
/
+95
*
cuda: double-buffer qwen flash K/V staging with cp.async
Onur Gündüz
2026-06-24
1
-82
/
+228
*
cuda: prefetch bert flash QK ahead of softmax to overlap the tensor cores
Onur Gündüz
2026-06-24
1
-50
/
+99
*
cuda: read bert flash K directly, drop the shared-memory transpose
Onur Gündüz
2026-06-24
1
-40
/
+65
*
cuda: broaden bert flash kv pipeline to all flash batches
Onur Gündüz
2026-06-24
1
-8
/
+10
*
tokenizer: parse BPE vocab.json with yyjson
Onur Gündüz
2026-06-24
1
-127
/
+25
*
api: parse model config with yyjson instead of a substring scanner
Onur Gündüz
2026-06-24
1
-77
/
+94
*
build: vendor yyjson 0.12.0 as the core JSON library
Onur Gündüz
2026-06-24
2
-0
/
+6
*
tests: cover CUDA reduced-precision stack and add determinism gate
Onur Gündüz
2026-06-24
1
-25
/
+45
*
cuda: fix bert reduced-precision ffn gate-up output aliasing its input
Onur Gündüz
2026-06-24
1
-129
/
+156
*
cuda: pipeline bert flash kv staging
Onur Gündüz
2026-06-24
1
-179
/
+649
*
cuda: widen bert flash key tile to 64, hoist q fragments
Onur Gündüz
2026-06-24
1
-100
/
+146
*
cuda: trim bert reduced scratch
Onur Gündüz
2026-06-24
1
-73
/
+213
*
server: bound merged batches by token budget
Onur Gündüz
2026-06-24
2
-6
/
+46
*
server: fix tokenizer data race under concurrency
Onur Gündüz
2026-06-24
3
-22
/
+8
*
server: make tokenize/render overlap adaptive, drop env switches
Onur Gündüz
2026-06-24
4
-146
/
+50
*
chore: document server diagnostic env switches
Onur Gündüz
2026-06-24
1
-0
/
+12
*
server: tokenize requests off the worker thread
Onur Gündüz
2026-06-24
3
-23
/
+209
*
server: render inline when idle, defer only under backlog
Onur Gündüz
2026-06-24
3
-6
/
+35
*
server: pack all compatible requests into token-budget batches
Onur Gündüz
2026-06-24
1
-20
/
+46
*
server: render embedding responses off worker
Onur Gündüz
2026-06-23
6
-16
/
+238
*
cuda: stage bert flash kv directly from qkv16
Onur Gündüz
2026-06-23
1
-73
/
+143
*
cuda: store bert qkv in 16-bit for flash batches
Onur Gündüz
2026-06-23
1
-40
/
+120
*
cuda: fuse bert qkv bias via cublaslt epilogue (f32 output)
Onur Gündüz
2026-06-23
1
-40
/
+59
*
cuda: cache the cublaslt gate-up matmul plan per shape
Onur Gündüz
2026-06-23
1
-77
/
+114
*
cuda: fuse bert gate-up bias+gelu via cublaslt epilogue
Onur Gündüz
2026-06-23
2
-66
/
+226
*
cuda: write bert residual delta 16-bit on the residual16 path
Onur Gündüz
2026-06-23
1
-10
/
+155
*
chore: clang-format libffwd/cuda.cu
Onur Gündüz
2026-06-23
1
-93
/
+92
*
cuda: add bert ffn16 intermediate path
Onur Gündüz
2026-06-23
5
-170
/
+907
*
cuda: auto-enable bert residual16 for long batches
Onur Gündüz
2026-06-23
1
-8
/
+12
*
cuda: gate bert bf16 residual stream
Onur Gündüz
2026-06-23
1
-33
/
+172
*
cuda: store weight dtype per context
Onur Gündüz
2026-06-22
2
-48
/
+58
*
cuda: prune attention experiments
Onur Gündüz
2026-06-22
4
-1357
/
+102
*
cuda: add opt-in BERT flash attention for head_dim 32
Onur Gündüz
2026-06-22
1
-49
/
+66
*
cuda: make bf16 BERT flash attention the default on the head_dim-64 path
Onur Gündüz
2026-06-22
1
-6
/
+9
*
cuda: add gated bf16 flash attention for the BERT shape (head_dim 64)
Onur Gündüz
2026-06-22
1
-6
/
+237
*
cuda: auto-enable bf16-activation BERT path above a token-count gate
Onur Gündüz
2026-06-22
1
-9
/
+18
*
cuda: add gated bf16-activation BERT path
Onur Gündüz
2026-06-22
1
-48
/
+142
*
cuda: prepack streaming K as bf16
Onur Gündüz
2026-06-22
1
-44
/
+65
[next]