revert: MoE加权+attention输出布局两刀(评测净负35.85>34.64,大中间张量/跨步写代价>省的clone)。保留消同步刀单独测

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

代码/code/infer.py

@@ -115,16 +115,14 @@ def _triton_varlen_attn(q, k, v, meta):
     cu, blk_seq, blk_inseq, total_blocks = meta
     BLOCK_M = CONFIG.get("triton_block_m", 64)
     # contiguous 后连续访存更快（实测去 contiguous 用 stride 读反而慢：非连续跨步读 > 一次性 clone）。
+    # contiguous 输出（实测：为消调用方 clone 改跨步写，评测反而更慢 35.85>34.64，已退回）
+    out = torch.empty((1, H, S, Dh), device=q.device, dtype=torch.float16)
     qc = q.contiguous(); kc = k.contiguous(); vc = v.contiguous()
     sh, ss, sd = S * Dh, Dh, 1
-    # out 存储为 [1,S,H,Dh]，以 [1,H,S,Dh] 视图返回 → 调用方 permute(0,2,1,3) 得连续、reshape 免 clone
-    out_storage = torch.empty((1, S, H, Dh), device=q.device, dtype=torch.float16)
-    out = out_storage.permute(0, 2, 1, 3)              # [1,H,S,Dh] 视图
-    soh, sos, sod = out.stride(1), out.stride(2), out.stride(3)   # = Dh, H*Dh, 1
     grid = (total_blocks, H)
     _varlen_flash_fwd[grid](
         qc, kc, vc, out, cu, blk_seq, blk_inseq,
-        sh, ss, sd, soh, sos, sod, 1.0 / math.sqrt(Dh), cu.numel() - 1,
+        sh, ss, sd, sh, ss, sd, 1.0 / math.sqrt(Dh), cu.numel() - 1,
         BLOCK_M=BLOCK_M, BLOCK_N=64, D=Dh,
     )
     return out
@@ -154,7 +152,8 @@ CONFIG = {
     # 须靠提交验证。AUC中性、MoE仅占2%算力故风险极低。
     "vectorize_moe": True,    # True=稠密向量化MoE(无同步点)；False=原逐expert循环(.nonzero同步)
     "moe_baddbmm": True,      # MoE FFN 用 baddbmm(cutlass GEMM+bias epilogue融合),省 bias add kernel
-    "moe_fused_weight": True, # top-k加权用scatter+mul+sum(在[E,N,D]上),省permute大clone+gather;数学等价
+    # 评测净负:scatter+mul+sum 物化[E,N,D]大中间张量(访存)>省的clone。退回 gather 路径。
+    "moe_fused_weight": False, # True=top-k加权用scatter+mul+sum(评测慢,勿开)
     # 真稀疏MoE实测评测净负:lat 34.64->37.64s(本地快15%但argsort/scatter开销评测放大,如varlen)
     # +容量丢弃降AUC(0.7525->0.7507)。已退回 dense。
     "moe_sparse": False,      # True=真稀疏MoE(评测净负,勿开)