aboutsummaryrefslogtreecommitdiff
path: root/embed_cuda.cu (unfollow)
Commit message (Expand)AuthorFilesLines
2026-06-16chore: move sources to src/ with proper names, drop qwen_ prefixOnur Gündüz1-2052/+0
2026-06-16cuda: memory-only bf16 for BERT (f32 compute on bf16 weights)Onur Gündüz1-6/+46
2026-06-16all: support BERT tanh-GeLU (gelu_new / gelu_pytorch_tanh)Onur Gündüz1-4/+18
2026-06-16cuda: add the BERT encoder family forward and loaderOnur Gündüz1-9/+238
2026-06-15all: add CLS first-token poolingOnur Gündüz1-0/+13
2026-06-15all: support GTE-Qwen2 embeddingsOnur Gündüz1-17/+55
2026-06-15late: add CUDA late-interaction rerank path (GPU encode, host MaxSim)Onur Gündüz1-4/+219
2026-06-15chore: run clang-formatOnur Gündüz1-442/+428
2026-06-14all: add support for Qwen3-Embedding + causal GQA attentionOnur Gündüz1-10/+91
2026-06-14chore: rename projectOnur Gündüz1-55/+55
2026-06-14cuda: restore BF16 attention internals, length-route batched attentionOnur Gündüz1-42/+130
2026-06-14cuda: drop BF16 attention internals, keep fused rope + batched F32 attentionOnur Gündüz1-125/+42
2026-06-14cuda: pointer-array batched attention for equal-length micro-batchesOnur Gündüz1-8/+167
2026-06-14cuda: fuse QK norm+rope launches, BF16 attention output pathOnur Gündüz1-45/+99
2026-06-14cuda: follow model file dtype for weight storage by defaultOnur Gündüz1-1/+15
2026-06-14chore: rename sources; binaries pplx-embed, pplx-embed-serverOnur Gündüz1-2/+2
2026-06-14gitignore: track new binary and library namesOnur Gündüz1-0/+0
2026-06-14cli: rename flags, add --stream, split serving into pplx-embed-serverOnur Gündüz1-2/+2
2026-06-14cuda: store BF16 activations directly in norm and SiLU kernelsOnur Gündüz1-51/+114
2026-06-14cuda: add optional BF16 CUDA weight storage (--cuda-weights bf16)Onur Gündüz1-68/+131
2026-06-14cuda: unify attention on one path; drop the max_seq>=64 dispatchOnur Gündüz1-73/+8
2026-06-14cuda: add fast GEMM mode (--cuda-fast-gemm mode)Onur Gündüz1-17/+75
2026-06-14cuda: clarify comment gemm_computeOnur Gündüz1-2/+3
2026-06-14cuda: add opt-in BF16 tensor-core GEMM modeOnur Gündüz1-21/+34
2026-06-14cuda: replace long-seq attention with cuBLAS GEMM attentionOnur Gündüz1-158/+124
2026-06-14cuda: replace tiled attention with warp-per-query reductionOnur Gündüz1-98/+105
2026-06-14cuda: extend tiled attention to ragged micro-batchesOnur Gündüz1-3/+105
2026-06-14cuda: add tiled single attentionOnur Gündüz1-4/+64
2026-06-14cuda: optimize rmsnorm reductionOnur Gündüz1-34/+29
2026-06-14cuda: add server contextual pathOnur Gündüz1-14/+159
2026-06-14cuda: optimize projection pathOnur Gündüz1-87/+118
2026-06-14cuda: fuse qkv projectionOnur Gündüz1-44/+78
2026-06-14cuda: accumulate residual projectionsOnur Gündüz1-21/+18
2026-06-14cuda: optimize attention baselineOnur Gündüz1-44/+47
2026-06-14cuda: add initial CUDA embedding backendOnur Gündüz1-0/+731