| Priority workload | SuperCollider | Record/Replay | Sampled | Inline Shadow |
|---|---|---|---|---|
| P0 Qwen3-0.6B prefill | 🟩 20/20; clean + exact-fault trap; overhead: 1.0x | 🟩 20/20; clean + qualified miss; overhead: 1.0x | 🟩 20/20; clean + measured statistical detection; repeated-dispatch metadata hardening tracked; overhead: 1.0x | 🟩 20/20; clean + attributed diagnostic; overhead: 1.0x |
| P1 Sharktank TP1 prefill | 🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; instability marker retained; overhead: 1.1x | 🟩 Clean 352/352 + 92/92 barriers; exact pair removal corrupts oracle, qualified replay miss; overhead: 1.7x | 🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; four unpatched supported barriers are decode-only; overhead: 1.0x | 🟩 Clean 352/352 + 92/92; exact pair removal emits an attributed diagnostic; overhead: 4.4x |
| P1 Sharktank TP1 decode/combined | 🟩 Clean 352/352; exact pair removal is a schedule-masked qualified miss; clean/fault instability marker is explicitly nondifferential; overhead: 1.1x | 🟩 Clean 352/352 + 92/92; exact pair removal corrupts the decode oracle, qualified replay miss; overhead: 1.5x | 🟩 Clean 352/352 + 18/18 applicable barriers; ordinary defaults complete repeated calls; exact pair removal is a schedule-masked qualified miss; overhead: 1.1x | 🟩 Clean and dynamically complete at 352/352 + 92/92; exact pair removal is a schedule-masked qualified diagnostic miss; overhead: 2.7x |
| P2 Sharktank TP2 family | 🟩 Clean 992/992; exact pair removal is schedule-masked and the instability marker is nondifferential; overhead: 1.2x | 🟩 Clean 992/992 + 152/152 barriers; exact pair removal corrupts the prefill oracle, qualified replay miss; overhead: 1.6x | 🟩 Clean 992/992 + 12/12 applicable barriers; exact pair removal is schedule/sampling-masked; overhead: 1.1x | 🟩 Clean and dynamically complete at 992/992 + 152/152; exact pair removal emits an attributed diagnostic; overhead: 3.0x |
| P3 CLIP BF16 | 🟨 Clean 85/85; instability marker | 🟨 Clean 85/85 + 72/72 barriers | 🟨 Clean 85/85 + 70/70 barriers; automatic standard sampling profile | 🟨 Clean 85/85 + 72/72; dynamically complete |
| P4 hip-moi D128 block attention | 🟨 Clean; 8/8 admitted generic-LDS terminals patched through direct or indirect caves | 🟨 Clean; 15/16 relevant sites (accesses 7/8 + barriers 8/8) | 🟨 Clean; accesses 8/8; decoded ownerless runtime atomics excluded | 🟨 Clean; barriers 8/8; exact dropped pair corrupts output, qualified diagnostic miss |
| P4 hip-moi D128 pressure attention | 🟨 Clean; 8/8 admitted generic-LDS terminals patched through local islands or bounded straight-line prefix relocation | 🟨 Clean; 12/16 relevant sites (accesses 4/8 + barriers 8/8); far function probes and dynamic-stack entry initialization compose through bounded indirect relays | 🟨 Clean; accesses 8/8 through local islands or bounded straight-line prefix relocation; decoded ownerless runtime atomics excluded | 🟨 Clean; barriers 8/8; shared barrier-only path qualified |
| P4 hip-moi WMMA attention | 🟨 Clean; 8/8 admitted generic-LDS terminals patched through direct or indirect caves | 🟨 Clean; 12/16 relevant sites (accesses 4/8 + barriers 8/8) | 🟨 Clean; accesses 8/8; decoded ownerless runtime atomics excluded | 🟨 Clean; barriers 8/8; shared barrier-only path qualified |
| P4 hip-moi Stream-K arrival | 🟨 Clean; 4/4 admitted generic-LDS terminals patched through direct or indirect caves | 🟨 Clean; 42/43 relevant sites (accesses 3/4 + barriers 8/8 + atomics 15/15 + fences 16/16) | 🟨 Clean; 14/19 relevant sites (accesses 4/4 + atomics 10/15); five far helper variants are typed placement misses | 🟨 Clean AcqRel/Relaxed; atomics 15/15 static; valid release evidence retained; one concurrent AcqRel publisher is typed undercoverage, never malformed metadata |
| P4 hip-moi tree atomic-OR | 🟨 Clean; 4/4 admitted generic-LDS terminals patched through direct or indirect caves | 🟨 Clean; 35/43 relevant sites (accesses 1/4 + barriers 8/8 + atomics 15/15 + fences 11/16) | 🟨 Clean; 10/19 relevant sites (accesses 4/4 + atomics 6/15); nine far helper variants are typed placement misses | 🟨 Clean AcqRel/Relaxed; atomics 15/15 static; valid release evidence retained; two typed dynamic undercoverage events and a qualified Relaxed diagnostic miss remain |
| P4 Jakub attention variants | 🟨 Clean 3/3 schedules; accesses 31/31 through owner-safe indirect islands | 🟨 Clean 3/3 schedules; 39/39 relevant sites (accesses 31/31 + barriers 8/8); far helper probes use owner-safe indirect islands | 🟨 Clean 3/3 schedules; accesses 31/31; decoded ownerless runtime atomics excluded | 🟨 Clean 3/3 schedules; 24/39 relevant access/barrier sites (16/31 + 8/8); dynamic undercoverage |
-
-
Save bjacob/4e7c5ebfef80f62c6867aa5486b033de to your computer and use it in GitHub Desktop.
a.md
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment