#6318 [GUI] Win 0.2.16 with V100+5070: panel downloads cuda13 engine "for V100" (no sm_70), launches on 5070; manual cuda12 works on V100 but UI still reports the 5070
open · @xglem2-eng · 2 コメント · GitHub で見る
本文
**Area**: GUI (model/GPU roles panel)
**Platform**: Desktop app (Windows)
**Strata version**: 0.2.16 (engine 0.1.62, feed sequence 128)
**Graphics cards**: Tesla V100-SXM2-16GB (TCC) + RTX 5070 (WDDM), driver 581.80
**Reproduction status**: every time on this hardware (clean install); workaround found and verified
### Environment
- Win11 Insider build 26300 (26H2 line)
- Tesla V100-SXM2-16GB on an adapter board with blower (PCIe bus 09:00.0, TCC mode, 2x8-pin power) + RTX 5070 (WDDM, drives the display)
- Driver 581.80 from the GeForce stack (install order: full Data Center 582.99 first, then full GeForce 581.80 on top; nvidia-smi then sees both)
- Ryzen 7 5700X, 80 GB DDR4. Model: Qwen3.8-Flash-Next IQ3_S (about 78 GiB mapped, packs+native, MTP), context 256K, kv int8
- Strata 0.2.16 clean install (after full uninstall), engine 0.1.62, feed sequence 128
### What happened (owner's live protocol, 2026-11-19)
1. Clean install: on first start the panel **detected the V100 as primary** and said it was downloading the engine **for the V100**.
2. What actually downloaded was the **cuda13 package**: `components\engine\0.1.62-windows-x86_64-cuda-cuda13\`, manifest `variant=cuda13`, archs = sm_75/sm_86/sm_89/sm_120 (+compute_120) - **no sm_70** (dlls are `*_13.dll`; checked right after install at 18:07).
3. Model launch went to the **5070** by default, despite the V100 being the chosen primary.
4. Manually unpacked the signed **cuda12 variant of 0.1.62 from the same feed** (sha256 verified): `components\engine\0.1.62-windows-x86_64-cuda-cuda12\` - launch **still tries the 5070**.
5. Manually selecting V100 = Primary: the panel shows the yellow banner "Tesla V100-SXM2-16GB cannot run Strata", **yet it launches and everything works well**.
6. With the model actually running on the V100, the model tab shows the VRAM budget as **"0 GB of 11 GB"** - 11-12 GB is the 5070's budget, not the V100's 16. The Monitor tab meanwhile draws the V100 correctly: 15.7 of 16.4 GB, 270-320 W, 1477-1507 MHz, ECC retired 0. So the runner accounting is right; the panel decision is stale / bound to the wrong card.
7. Control loop in the evening: cuda13 restored -> "everything as before" (expected, no sm_70); cuda12 put back via rename -> V100 runs again.
### Instrumental evidence (the assistant model logging its own engine)
- Default launch, engine log: `GPU 0: NVIDIA GeForce RTX 5070, compute capability 12.0` - single enumeration line, the V100 absent from the engine device list.
- After manual V100 primary: `GPU 0: Tesla V100-SXM2-16GB, compute capability 7.0` (engine 0.1.62); live process from `...\cuda-cuda12\strata.exe`; nvidia-smi: 15982/16384 MiB on V100, 5070 idle.
- Benchmarks on V100: **decode 46.5 t/s, fresh-prompt read 1332 t/s** (same config on 5070: ~29 t/s decode, 334-432 t/s read).
- **Pointer anomaly**: `components\.updates\state.json` keeps `engine.active = "0.1.62-windows-x86_64-cuda-cuda13"` while the cuda13 folder does NOT exist on disk (deleted by hand) and the live engine is cuda12. Checked three times in one evening (18:56, 20:33, 21:15) - `active` never switched to the actually running package.
Hypothesis: the panel decides "can this GPU run Strata", the VRAM budget and the default launch choice from the manifest of `engine.active` (cuda13: no sm_70, so the V100 "cannot"; recommended card = 5070 with its 11 GB), not from the actually running engine or the chosen roles. The cosmetics (yellow banner, "11 GB") survive even a successful run on the V100.
- Side note: `nvidia-smi -i 0 -pl 220` applied successfully (verified 220.00 W) but **reverted to 300 W after reboot** - not persistent on this GeForce-driver + TCC combination.
### Expectation vs fact
- Expectation: if the panel decides which engine the chosen GPU needs, it should fetch the package containing sm_70 (the cuda12 variant from the same feed) and launch on the chosen GPU; UI accounting should follow the real card.
- Fact: a package without sm_70 is fetched; launch goes to the 5070; manual cuda12 runs on the V100 while the UI keeps treating another card/package as the active one.
### Workaround (verified live, hours of uptime)
Manually unpack the feed's cuda12 variant into `components\engine\` and choose V100 = Primary. Ignore panel warnings as cosmetic. Archive used: `strata-engine-0.1.62-windows-x86_64-cuda-cuda12.tar.gz` (739 MB, sha256 verified against the feed).
### Related
#5380 (linux GUI does not see V100) - likely a sibling bug at the GUI layer ("the GUI looks the wrong way, the runner looks the right way").
Happy to provide engine log, desktop log and state.json on request.関連リンク
インストール・モデル・リリースへのサイト内リンク。