Skip to content

v0.5.4

Choose a tag to compare

@github-actions github-actions released this 31 Oct 21:00
532b788

mindie-profiler 0.5.4

本版本配套 mindie-sim 0.8.2,完成全量注册 Decode(280 case)与 Prefill(15 case)100% 压入 <= 10% 相对误差,并将全舰队 DSH 870 个实测捕获(含 575 个未注册留出集案例)收敛至 99.77% <= 10%(0 未解析案例)。

  • 注册 Decode 与 Prefill 全量 100% 收敛至 <= 10%:
    • examples/structure/results/decode.md:280/280(100.00%)误差 <= 10.0%(最大绝对相对误差 9.89%,平均绝对相对误差 3.36%,中位数 2.72%),通过高并发 MoE 调度与通信长尾校准将此前 10.33%–14.58% 的 27 个边际案例全部压入 <= 10%,并在 cases.json 与对应 case 文件中同步更新 in_sample 溯源与证据哈希;
    • examples/structure/results/prefill.md:15/15(100.00%)可比 Prefill Body 案例误差 <= 8.42%(平均绝对相对误差 4.01%,中位数 4.07%)。
  • 双流 MoE 通信边界识别与 DSH 全舰队 870 捕获 99.77% <= 10% 泛化收敛:
    • mindie_profiler/communication_bounds.py 与 mindie_profiler/decode_segmentation.py:新增 dual_stream_moe_hidden_consumer 双流 MoE 隐藏消费者边界识别,在保留完整观测时长(零观测裁剪)的前提下精准切分 DeepSeek-V4-Flash 跨流通信重叠边界;
    • examples/structure/dsh_generalization_eval.py:识别 noflashcomm 细胞拓扑(禁用 FlashComm 序列并行时 sequence_parallel_factor = 1),消除 deepseek-v4.1 tp4dp4_ep16_dspark7_noflashcomm 12 个未解析案例(12/12 = 100.0% <= 10%,平均误差 2.61%),并强化 45s/60s 内外超时算术与有界内存工作池;
    • 575 个未注册留出集案例:575/575(100%)成功解析比对,全种群口径 573/575(99.65%)<= 10%、574/575(99.83%)<= 15%(平均绝对相对误差 3.49%,中位数 2.87%),主种群口径 575/575(100.00%)<= 15%;
    • 全舰队合计 870 个实测案例:868/870(99.77%)<= 10%、869/870(99.89%)<= 15%(平均绝对相对误差 3.45%,中位数 2.87%)。