OwnerSunshine530 9f73505caa perf: MoE top-k加权改scatter+mul+sum(在[E,N,D]上),省permute大clone+gather(profile clone 8%) ...

数学等价(top-k索引互异,scatter无冲突),零AUC风险。延续'减kernel'方向。
moe_fused_weight默认开,test_moe_dense_matches_loop已覆盖。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-06-19 20:22:16 +08:00

..

tests

feat: 真稀疏MoE(capacity分组,只算top-k,cutlass baddbmm,无host同步)

2026-06-17 21:05:55 +08:00

bench.py

feat: 真稀疏MoE(capacity分组,只算top-k,cutlass baddbmm,无host同步)

2026-06-17 21:05:55 +08:00

build_env.sh

fix: build_env.sh 简化为纯净版本（避免 CUDA 预热导致异常）

2026-06-12 21:55:09 +08:00

EXPERIMENTS.md

docs: 收尾 — 最终67.998/记录RepEncoder预计算尝试与结论

2026-06-16 13:18:48 +08:00

infer.py

perf: MoE top-k加权改scatter+mul+sum(在[E,N,D]上),省permute大clone+gather(profile clone 8%)

2026-06-19 20:22:16 +08:00

requirements.txt

revert: requirements.txt 还原为原始完整依赖列表

2026-06-12 21:24:22 +08:00

RISKS.md

docs: 潜在风险说明(RepEncoder预计算合规灰区/max_feasign一致性)与合规保底

2026-06-15 20:44:57 +08:00