Files
aituner/runs/frontier-fidelity-envelope-v1/smoke-report.md

16 lines
1.4 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Frontier measured-collective smoke
日期2026-07-17。设备local CPUFrontier simulation only。Frontier commit`d9cfeb6d8791fbf2f295dd9744c56a666171776e`,沿用既有 dirty compatibility patch setA2 额外 patch SHA256 为 `35cc6be846589faf8cb5fa3ce5fdfe0aee8f086ba7dbb5dbcdc677148f19a3c8`
固定 cellQwen3-30B-A3B BF16 profile-v2`TP2/MNS8/MBT8192``ISL=2048/OSL=1`64 requests8 req/sprefix offTTFT SLO 1256 ms。
| variant | CC path for 2048-token payload | TTFT p50/p95/max (ms) | pass rate |
|---|---|---:|---:|
| A0 analytical | analytical | 122.2406276 | 1.0 |
| A1 native Vidur | `4,194,304 > 100,000` elementslookup miss 后 analytical fallback | 122.2406276 | 1.0 |
| A2 measured + direct miss | measured random-forest estimatorexact payload memoized | 120.8898909 | 1.0 |
A1 与 A0 的所有 TTFT 数值完全一致,验证 measured CSV 虽成功加载和训练,但没有参与该 payload 的 prediction。A2 model 对该 payload 的预测为 `0.09877793 ms`;同一个 TP2/2048 row 的 measured critical-path median 是 `0.10891200 ms`。A2 相对 A1 的 E2E delta 为 `-1.3507367 ms``-1.105%`),与每层多次 collective 的累计量级一致因此通过“execution path 必须变化”的 smoke gate。
这个 smoke 只证明 profile consumption不证明 ranking 已恢复。下一步必须运行完整 A2 TP×MNS surface并与已冻结 real capacity比较。