perf: add fused CUDA rotary embedding kernel

- Single-kernel rotary embedding (cos/sin lookup + rotation) replaces PyTorch complex-multiply path (3 kernel launches + f32 upcast per call) - RotaryEmbedding now stores cos_table/sin_table and returns (cos, sin) f32 tuple instead of a complex tensor - apply_rotary_emb in rotary_backend.py auto-dispatches: CUDA kernel if available, else torch complex-multiply fallback; backend-agnostic (both attention backends benefit) - Kernel: 256-thread blocks, grid-stride loop, vectorized __nv_bfloat162 load/store, f32 compute, bf16 out - Standalone kernel 6-9x faster than torch across decode/prefill shapes, max diff 0 (decode) to 3e-2 (large prefill, bf16) - Benchmark (L20, bf16, CUDA backend): B=1 9.48->7.25ms (+31%), B=4 10.73->7.67ms (+40%), B=8 10.77->7.81ms (+38%), B=16 10.79->7.83ms (+38%)
2026-07-31 15:27:31 +08:00
parent 50cfd0d555
commit 3e67b4f88d
9 changed files with 218 additions and 24 deletions
@@ -6,11 +6,11 @@ import torch.nn.functional as F
 from torch import Tensor

 from astrai.extension import attention
+from astrai.extension.rotary_backend import apply_rotary_emb
 from astrai.factory import BaseFactory
 from astrai.inference.core.cache import KVCache
 from astrai.model.components.linear import Linear
 from astrai.model.components.norm import RMSNorm
-from astrai.model.components.rope import apply_rotary_emb


 class AttnFactory(BaseFactory[nn.Module]):