feat: Triton BLOCK_M 可调(triton_block_m,默认64);bench --triton-bm 扫描

突破:triton评测39.92s/69.72(vs chunked 47.84/67.998)。继续调BLOCK_M榨。

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

代码/code/bench.py

@@ -318,6 +318,7 @@ def _parse_args():
     ap.add_argument("--attn", choices=["sdpa", "chunked", "triton", "flex", "varlen"], default=None,
                     help="注意力：sdpa=稠密, chunked=分块SDPA, triton=varlen flash kernel, flex/varlen=对照")
     ap.add_argument("--chunk-users", type=int, default=None, help="chunked 模式每块用户数")
+    ap.add_argument("--triton-bm", type=int, default=None, help="Triton query 块大小(32/64/128)")
     ap.add_argument("--moe", choices=["dense", "loop"], default=None,
                     help="MoE实现：dense=向量化(新), loop=逐expert循环(原)")
     ap.add_argument("--compile", action="store_true", help="开启 torch.compile")
@@ -361,6 +362,8 @@ if __name__ == "__main__":
         cfg["attn"] = a.attn
     if a.chunk_users is not None:
         cfg["chunk_users"] = a.chunk_users
+    if a.triton_bm is not None:
+        cfg["triton_block_m"] = a.triton_bm
     if a.moe is not None:
         cfg["vectorize_moe"] = (a.moe == "dense")
     if a.emb_fp16: