aboutsummaryrefslogtreecommitdiff
path: root/libffwd (follow)
Commit message (Expand)AuthorAgeFilesLines
* cuda: default cuDNN attention to Qwen only, BERT opt-inOnur Gündüz2026-07-021-12/+19
* cuda: add cuDNN fused attention path with runtime fallbackOnur Gündüz2026-07-024-6/+603
* chore: clean up comment formatting and wordingOnur Gündüz2026-06-2742-452/+439
* cuda: skip Qwen flash softmax rescale when running max is stableOnur Gündüz2026-06-261-8/+33
* cuda: load Qwen flash V with ldmatrix.trans instead of gatherOnur Gündüz2026-06-261-35/+83
* build: vendor yyjson 0.12.0 in deps/yyjsonOnur Gündüz2026-06-251-0/+6
* cuda: skip O(L^2) score buffer for qwen flash path (enables 32k)Onur Gündüz2026-06-251-1/+7
* cuda: stage compact GQA K/V in qwen attention, halve the expandOnur Gündüz2026-06-251-14/+47
* chore: run clang-formatOnur Gündüz2026-06-2514-392/+345
* build: update build info and accelerator labelsOnur Gündüz2026-06-253-28/+24
* all: migrate to yyjsonOnur Gündüz2026-06-253-6/+3
* cuda: skip per-warp future key tiles in causal qwen attentionOnur Gündüz2026-06-251-65/+95
* cuda: double-buffer qwen flash K/V staging with cp.asyncOnur Gündüz2026-06-241-82/+228
* cuda: prefetch bert flash QK ahead of softmax to overlap the tensor coresOnur Gündüz2026-06-241-50/+99
* cuda: read bert flash K directly, drop the shared-memory transposeOnur Gündüz2026-06-241-40/+65
* cuda: broaden bert flash kv pipeline to all flash batchesOnur Gündüz2026-06-241-8/+10
* tokenizer: parse BPE vocab.json with yyjsonOnur Gündüz2026-06-241-127/+25
* api: parse model config with yyjson instead of a substring scannerOnur Gündüz2026-06-241-77/+94
* build: vendor yyjson 0.12.0 as the core JSON libraryOnur Gündüz2026-06-241-0/+3
* cuda: fix bert reduced-precision ffn gate-up output aliasing its inputOnur Gündüz2026-06-241-129/+156
* cuda: pipeline bert flash kv stagingOnur Gündüz2026-06-241-179/+649
* cuda: widen bert flash key tile to 64, hoist q fragmentsOnur Gündüz2026-06-241-100/+146
* cuda: trim bert reduced scratchOnur Gündüz2026-06-241-73/+213
* cuda: stage bert flash kv directly from qkv16Onur Gündüz2026-06-231-73/+143
* cuda: store bert qkv in 16-bit for flash batchesOnur Gündüz2026-06-231-40/+120
* cuda: fuse bert qkv bias via cublaslt epilogue (f32 output)Onur Gündüz2026-06-231-40/+59
* cuda: cache the cublaslt gate-up matmul plan per shapeOnur Gündüz2026-06-231-77/+114
* cuda: fuse bert gate-up bias+gelu via cublaslt epilogueOnur Gündüz2026-06-231-65/+225
* cuda: write bert residual delta 16-bit on the residual16 pathOnur Gündüz2026-06-231-10/+155
* chore: clang-format libffwd/cuda.cuOnur Gündüz2026-06-231-93/+92
* cuda: add bert ffn16 intermediate pathOnur Gündüz2026-06-233-168/+905
* cuda: auto-enable bert residual16 for long batchesOnur Gündüz2026-06-231-8/+12
* cuda: gate bert bf16 residual streamOnur Gündüz2026-06-231-33/+172
* cuda: store weight dtype per contextOnur Gündüz2026-06-222-48/+58
* cuda: prune attention experimentsOnur Gündüz2026-06-221-1330/+17
* cuda: add opt-in BERT flash attention for head_dim 32Onur Gündüz2026-06-221-49/+66
* cuda: make bf16 BERT flash attention the default on the head_dim-64 pathOnur Gündüz2026-06-221-6/+9
* cuda: add gated bf16 flash attention for the BERT shape (head_dim 64)Onur Gündüz2026-06-221-6/+237
* cuda: auto-enable bf16-activation BERT path above a token-count gateOnur Gündüz2026-06-221-9/+18
* cuda: add gated bf16-activation BERT pathOnur Gündüz2026-06-221-48/+142
* cuda: prepack streaming K as bf16Onur Gündüz2026-06-221-44/+65
* cuda: promote tc3 to the default Qwen bf16 attention pathOnur Gündüz2026-06-221-16/+32
* cuda: add qwen tc3 ldmatrix attentionOnur Gündüz2026-06-221-8/+460
* cuda: add K/V-sharing bf16 tensor-core attention (tc2)Onur Gündüz2026-06-221-30/+227
* cuda: add qwen bf16 tensor-core flash attentionOnur Gündüz2026-06-221-4/+459
* cuda: add register-resident streaming attention behind the flagOnur Gündüz2026-06-221-4/+172
* cuda: BERT bf16-compute GEMMs skip the f32 weight widenOnur Gündüz2026-06-211-6/+14
* cuda: add gated streaming attention prototypeOnur Gündüz2026-06-211-0/+383
* cuda: fuse BERT elementwise kernelsOnur Gündüz2026-06-211-73/+96
* cuda: fix bf16 BERT batched-attention buffer-layout mismatchOnur Gündüz2026-06-211-4/+16