ViperEkura 15862d4b56 perf: fuse fp8 linear fwd and bwd into single kernel calls ...

- fp8_linear_forward: cast + cublasLt GEMM + transpose + bias in one call
- fp8_linear_backward: scale-free, dtype derived from input tensor
- drops per-op Python dispatch (was ~6-8 launches per linear) and amax syncs
- 1024x1024 linear: 6.8x slow -> 0.67x (36.7us vs 24.8us bf16)
- small-model e2e still 1.71x slow; 15bt estimate ~0.78x (linear-heavy)

2026-08-14 01:24:37 +08:00

..

lib

refactor: move compiled kernel .so files into extension/lib

2026-07-31 15:36:32 +08:00

__init__.py

refactor : use factory for attention backends

2026-08-05 15:37:22 +08:00

attention_backend.py

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

attention_ops.py

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

fp8_dispatch.py

perf: fuse fp8 linear fwd and bwd into single kernel calls

2026-08-14 01:24:37 +08:00

fp8_ops.py

perf: fuse fp8 linear fwd and bwd into single kernel calls

2026-08-14 01:24:37 +08:00

loader.py

feat: add fp8 training via cublasLt dispatch

2026-08-14 00:39:49 +08:00

rotary_backend.py

refactor: unify rotary embedding interface and update docs

2026-07-31 16:52:25 +08:00

rotary_ops.py

refactor: unify rotary embedding interface and update docs

2026-07-31 16:52:25 +08:00