aboutsummaryrefslogtreecommitdiff
path: root/libffwd (unfollow)
Commit message (Expand)AuthorFilesLines
2026-07-02cuda: default cuDNN attention to Qwen only, BERT opt-inOnur Gündüz1-12/+19
2026-07-02cuda: add cuDNN fused attention path with runtime fallbackOnur Gündüz4-6/+603
2026-06-27chore: clean up comment formatting and wordingOnur Gündüz42-452/+439
2026-06-26cuda: skip Qwen flash softmax rescale when running max is stableOnur Gündüz1-8/+33
2026-06-26cuda: load Qwen flash V with ldmatrix.trans instead of gatherOnur Gündüz1-35/+83
2026-06-25build: vendor yyjson 0.12.0 in deps/yyjsonOnur Gündüz1-0/+6
2026-06-25cuda: skip O(L^2) score buffer for qwen flash path (enables 32k)Onur Gündüz1-1/+7
2026-06-25cuda: stage compact GQA K/V in qwen attention, halve the expandOnur Gündüz1-14/+47
2026-06-25chore: run clang-formatOnur Gündüz14-392/+345
2026-06-25build: update build info and accelerator labelsOnur Gündüz3-28/+24
2026-06-25all: migrate to yyjsonOnur Gündüz3-6/+3
2026-06-25cuda: skip per-warp future key tiles in causal qwen attentionOnur Gündüz1-65/+95
2026-06-24cuda: double-buffer qwen flash K/V staging with cp.asyncOnur Gündüz1-82/+228
2026-06-24cuda: prefetch bert flash QK ahead of softmax to overlap the tensor coresOnur Gündüz1-50/+99
2026-06-24cuda: read bert flash K directly, drop the shared-memory transposeOnur Gündüz1-40/+65
2026-06-24cuda: broaden bert flash kv pipeline to all flash batchesOnur Gündüz1-8/+10
2026-06-24tokenizer: parse BPE vocab.json with yyjsonOnur Gündüz1-127/+25
2026-06-24api: parse model config with yyjson instead of a substring scannerOnur Gündüz1-77/+94
2026-06-24build: vendor yyjson 0.12.0 as the core JSON libraryOnur Gündüz1-0/+3
2026-06-24cuda: fix bert reduced-precision ffn gate-up output aliasing its inputOnur Gündüz1-129/+156
2026-06-24cuda: pipeline bert flash kv stagingOnur Gündüz1-179/+649
2026-06-24cuda: widen bert flash key tile to 64, hoist q fragmentsOnur Gündüz1-100/+146
2026-06-24cuda: trim bert reduced scratchOnur Gündüz1-73/+213
2026-06-23cuda: stage bert flash kv directly from qkv16Onur Gündüz1-73/+143
2026-06-23cuda: store bert qkv in 16-bit for flash batchesOnur Gündüz1-40/+120
2026-06-23cuda: fuse bert qkv bias via cublaslt epilogue (f32 output)Onur Gündüz1-40/+59
2026-06-23cuda: cache the cublaslt gate-up matmul plan per shapeOnur Gündüz1-77/+114
2026-06-23cuda: fuse bert gate-up bias+gelu via cublaslt epilogueOnur Gündüz1-65/+225
2026-06-23cuda: write bert residual delta 16-bit on the residual16 pathOnur Gündüz1-10/+155
2026-06-23chore: clang-format libffwd/cuda.cuOnur Gündüz1-93/+92
2026-06-23cuda: add bert ffn16 intermediate pathOnur Gündüz3-168/+905
2026-06-23cuda: auto-enable bert residual16 for long batchesOnur Gündüz1-8/+12
2026-06-23cuda: gate bert bf16 residual streamOnur Gündüz1-33/+172
2026-06-22cuda: store weight dtype per contextOnur Gündüz2-48/+58
2026-06-22cuda: prune attention experimentsOnur Gündüz1-1330/+17
2026-06-22cuda: add opt-in BERT flash attention for head_dim 32Onur Gündüz1-49/+66
2026-06-22cuda: make bf16 BERT flash attention the default on the head_dim-64 pathOnur Gündüz1-6/+9
2026-06-22cuda: add gated bf16 flash attention for the BERT shape (head_dim 64)Onur Gündüz1-6/+237
2026-06-22cuda: auto-enable bf16-activation BERT path above a token-count gateOnur Gündüz1-9/+18
2026-06-22cuda: add gated bf16-activation BERT pathOnur Gündüz1-48/+142
2026-06-22cuda: prepack streaming K as bf16Onur Gündüz1-44/+65
2026-06-22cuda: promote tc3 to the default Qwen bf16 attention pathOnur Gündüz1-16/+32
2026-06-22cuda: add qwen tc3 ldmatrix attentionOnur Gündüz1-8/+460
2026-06-22cuda: add K/V-sharing bf16 tensor-core attention (tc2)Onur Gündüz1-30/+227
2026-06-22cuda: add qwen bf16 tensor-core flash attentionOnur Gündüz1-4/+459
2026-06-22cuda: add register-resident streaming attention behind the flagOnur Gündüz1-4/+172
2026-06-21cuda: BERT bf16-compute GEMMs skip the f32 weight widenOnur Gündüz1-6/+14
2026-06-21cuda: add gated streaming attention prototypeOnur Gündüz1-0/+383
2026-06-21cuda: fuse BERT elementwise kernelsOnur Gündüz1-73/+96
2026-06-21cuda: fix bf16 BERT batched-attention buffer-layout mismatchOnur Gündüz1-4/+16