storytold/cloud-worker
2
0
Fork 0
mirror of https://github.com/storytold/cloud-worker.git synced 2026-10-09 00:09:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

MITIGATIONS.md: 24GB multi-ref RESCUED — full envelope verdicts per machine class

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Hanashi
2026-08-06 19:15:36 -04:00
parent e33192a283
commit 056d5785e7
1 changed files with 10 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/MITIGATIONS.md
+10 -6
View File
@@ -55,12 +55,16 @@ RAM ≥ 4× VRAM.
ref-resolution ladder (0.1 MP → native) are being measured on all three
GPU classes right now (`--suite refmatrix`).
## Open items
## Resolved verdicts (2026-08-06 late)
- 5090: 9-ref and 8-ref-`max` verdicts (running).
- 4090 with full stack (#15316 + cache-none + reserve 5): multi-ref verdict
(running).
- RTX PRO 4000: dropped from fleet consideration on economics; would need
re-test only if its price class matters.
- **24 GB multi-ref: RESCUED.** Full stack = #15316 patch + `--reserve-vram 5`
+ `--cache-none` + `--disable-pinned-memory` + valid ref images. Entire
practical envelope works (1–9 refs, all aspects, full-res canvas); only
9-ref `max` sizing remains out of reach (and is uneconomical everywhere).
- **5090 runs the complete envelope** including bf16 t2v via layerwise
streaming at 110 s warm — faster than its own pruned int8 (143 s).
- A recurring *non*-OOM failure class was corrupt reference downloads
(HTTP 429 partial files) masquerading as capability failures — bench images
are now staged on each region's network volume by the installer.
- Upstream: file the multi-ref OOM issue referencing #15316 (user decision —
outward-facing).