#3605 [Engine] During initial strata launch migrating from 0.1.45 to 0.2.1 it hangs while test-starting the engine

说明

Windows 11, 64GB, RTX 4070 Ti Super, source installation

Updated from 0.1.45 source installation.

Ran "M:\Strata\.venv\Scripts\python.exe" strata.py launch

It found the installed model (iq3_xxs) and after I answered the setup questions it downloaded engine 0.1.47 windows-x86_64-cuda-cuda13 and then "test-started" it. This successfully starts the engine but it never responds and eventually times out. At this point the script deletes the engine.

I have tried this with and without `set STRATA_COMPONENTS_DIR=M:\Strata-Components`. The 2nd time I allowed it to install into %appdata%\local\ but that did not help. The log is included below. It did leave a file in 
%appdata%\Local\Strata\components\.updates\partial in both cases but also what seemed like a full engine in the .s directory before it was deleted. The failure method was the same in both cases.

test-start-0.1.7.log
```
ggml_cuda_init: failed to initialize CUDA: no CUDA-capable device is detected
load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842

strata-vision: model files loaded in 1.0 s
strata-vision: on the CPU, 12 threads, no warm-up
[strata] 2026-10-21 15:40:58 engine started: C:\Users\njcr\AppData\Local\Strata\components\.s\fcc4\engine\0.1.47-windows-x86_64-cuda-cuda13\strata.exe --serve --pack M:\Strata-data\packs\iq3_xxs --native M:\Strata-data\models\IQ3_XXS\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf --ple-gguf M:\Strata-data\models\IQ3_XXS\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf --expert-profile expert-profile-learned.bin --expert-cache auto --prefill auto --spec 4 --mtp M:\Strata-data\mtp\rt --max-context 131072 --kv int8 --kv-resident 32768 --vision --pcie-frac 0.35 --spec-min-p 0.70 --adapt-every 1 --adapt-swaps 80 --adapt-decay 0.97 --vram-reserve-mib 1600 --expert-profile-save expert-profile-learned.bin
strata generate: memory guard: limit 63.4 GiB (source: meminfo), current 10.5 GiB, reclaimable cache credited 0.0 GiB; available 52.8 GiB (MemAvailable 52.8 GiB), floor 2560 MiB
strata generate: native pack: M:\Strata-data\packs\iq3_xxs experts (largest blob 2.33 MB), token embedding IQ3_S in mapped host memory (260 MiB, 0.3 s)
strata generate: 1466 MiB of weights loaded from M:\Strata-data\packs\iq3_xxs in 0.2 s (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 1781.11 MiB of weights, in 1.4 s
strata generate: PLE direct read probe: 205045 rows/s (fine)
strata generate: the SSD is kept awake while rows are read: one page of the table after 100 ms without a read, until 60 s after the last request (STRATA_SSD_KEEPALIVE=0 turns it off)
strata generate: profile expert-profile-learned.bin: 24576 ranked pairs, built for 24576 slots
strata generate: KV streaming: 32768 of 131072 cells per QSA layer in VRAM, the K/V in 1.55 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of M:\Strata-data\models\IQ3_XXS\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.17 s (4566 MiB/s)
strata generate: experimental native Q5_K head, 437043200 bytes, in 0.3 s
strata generate: GPU 0: NVIDIA GeForce RTX 4070 Ti SUPER, compute capability 8.9
strata: pinned-memory headroom: holding 800 MiB for the buffers allocated after the expert arena (released after it: 640 MiB; kept until needed: 160 MiB; STRATA_PIN_RESERVE_MIB changes it)
strata generate: expert arena read unbuffered (0 of 16 probe reads from the file cache; 9.8 GiB available, 43.8 GiB of files)
strata: pinned-memory headroom: 640 MiB handed back after the expert arena
strata generate: expert arena: cudaHostRegister PORTABLE ok; large pages refused for 42916118528 B (GetLargePageMinimum=2097152, VirtualAlloc error 1314); using 4 KB pages
strata generate: loaded 39.97 GiB at 5.96 GiB/s
strata generate: CPU pool tasks/phase: 36 (automatic), participating threads: 12
strata generate: CPU pool: 11 workers by default (12 physical cores, 24 logical processors; bounded by every physical core but the host's)
strata generate: 11 pool workers on logical processors 2,4,6,8,10,12,14,16,18,20,22, the host thread on 0 (draining too) (--host-core first)
strata generate: CPU workers on logical processors 2,4,...,22 (11 workers, one per physical core; host on 0)
strata aux cpus: not available on this system
strata generate: expert cache auto: 9.36 GiB free, 1600 MiB reserved (+184 MiB for the draft head) -> 3512 slots
strata generate: expert cache 4661 slots, 7.62 GiB of VRAM; policy is
strata generate: the GPU computes the experts in the cache; it rounds differently from the CPU,
                 so a reply can differ slightly from a run without the cache (same quality:
                 bench/results/2026-09-27-cache-parity).
                 PROFILE, ranked by routing frequency, no eviction.
strata generate: pre-filled 4661 of 4661 slots from the profile in 0.4 s (22912 MB/s); slot 0 verified
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 11 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.47)
strata generate: token graph hit path: 4661 resident experts, decided on the device
strata serve: Windows commit: 18.7 GiB left of a 95.4 GiB limit (RAM + page file)
prefill: the CPU share is ON by default for prompt chunks below 1024 tokens (the idle CPU takes some of the experts the GPU would stream; answers can differ slightly from 0.1.40.3, mean KL ~0.004). STRATA_PREFILL_CPU_SHARE=0 turns it off.
strata serve: prompt chunk auto: 8192 tokens, a 384-slot ring
strata serve: the prompt path borrows 2537 CUDA0 cache slots (4.12 GiB)
strata hc: CUDA0: the hyper-connection read runs as staged (the norm per token and stream, the down projection's activations staged ahead by cp.async); checked bit for bit against the plain read on this card (STRATA_HC_SPLIT=1 or 0 for the earlier ones)
strata verify: window up to 6 tokens, 70.1 MiB of device buffers
strata mtp: draft head over 106299 tokens (178.4 MiB)
strata serve: STRATA_ROUTE_TAIL_SKIP=7 is ON (default): a missed expert that every token of a verify window routes at rank 7 or lower is skipped (+10..20% decode on 12-16 GB cards that miss experts, answers differ slightly from 0.1.41); STRATA_ROUTE_TAIL_SKIP=0 turns it off
strata serve: adaptive overlap: off (STRATA_ADAPT_OVERLAP=1 turns it on)
strata serve: 1253 MiB of VRAM free with everything loaded
```
state.json
```{
 "engine": {
  "failed": {
   "0.1.47": {
    "at": "2026-10-21T14:50:59.760237+00:00",
    "manifest_sha256": "95eb1eef404859e3afae56811cccb3da27bbab6b589e276fe4e4e701dfbe9427",
    "reason": "it did not answer a one-token probe (startup_wait_expired) | strata serve: STRATA_ROUTE_TAIL_SKIP=7 is ON (default): a missed expert that every token of a verify window routes at rank 7 or lower is skipped (+10..20% decode on 12-16 GB cards that miss experts, answers differ slightly from 0.1.41); ST"
   }
  }
 },
 "feed": {
  "channel": "stable",
  "checked_at": "2026-10-21T14:39:45.345211+00:00",
  "key_id": "5d1bc23abbeffd0d",
  "sequence": 107,
  "source": "https://github.com/Niko1221/Strata/releases/latest/download/strata-feed.json"
 },
 "history": [
  {
   "action": "update-engine",
   "at": "2026-10-21T14:50:59.760237+00:00",
   "detail": "test start failed: it did not answer a one-token probe (startup_wait_expired) | strata serve: STRATA_ROUTE_TAIL_SKIP=7 is ON (default): a missed expert that every token of a verify window routes at rank 7 or lower is sk",
   "from": null,
   "ok": false,
   "to": "0.1.47"
  }
 ],
 "last_check": {
  "app": null,
  "at": "2026-10-21T11:55:02.868460+00:00",
  "control": null,
  "engine": "0.1.47",
  "engine_installed": null
 },
 "schema_version": 1
}


```