#6250 [Engine] Arc Pro B70 (xe): server reaches "ready", first request dies with verify flag A wait
open · @lanefritzel · 4 コメント · GitHub で見る
Setup & installModels & quantsLinux
本文
## Problem
On an Arc Pro B70 (xe driver), the server loads the model and prints "ready", but the first request makes the engine exit with a native verification failure. Reproduced with two quants.
## Environment
- Strata Terminal 0.1.62, engine 0.1.62 linux-x86_64-sycl
- GPU: Intel Arc Pro B70 32 GB (Battlemage G31, device 0xe223), kernel driver xe
- OS: Ubuntu 26.04.1 LTS, kernel 7.0.0-42-generic (differs from the Ubuntu 24.04 stack in INTEL_ARC.md)
- Compute runtime: intel-opencl-icd 26.05.37020.3-1, libze-intel-gpu1 26.05.37020.3-1, libze1 1.28.2-2 (legacy 24.35 packages also installed); no oneAPI toolkit installed
- CPU: Intel Core Ultra 7 270K Plus; no AVX-512 (engine uses AVX2)
- RAM: 31 GB, 8 GB swap
- Models: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF, IQ3_XXS and IQ2_XS, on a USB spinning disk (ext4); n-gram table read with --ple-io direct
## Reproduce
SYCL_CACHE_PERSISTENT=0 strata launch --model-path <folder with the shards> --model <iq3_xxs|iq2_xs> --data-dir <dir> --context 32k --vision off --drafting off --set STRATA_VERIFY_NO_HOST=1 --port 8085 --no-tui
## Expected
After "ready: http://127.0.0.1:8085/v1", the launcher's first request is answered.
## Actual
The pack build, expert cache fill and host mirror complete, and the server prints "ready" (about 1580 s for IQ3_XXS on this disk). The first request then fails after ~62 s ("reading the prompt: 53 tokens") and the engine exits with code -15:
"Unrecoverable native verification failure: verify: the GPU stopped waiting for the plan for its GPU share (flag A) of layer 0 (group 0, window 0 at position 0, 6 tokens, device 0 layers [0,48)): the CPU took 0 ms, longer than the GPU's wait of 1000000 reads; the window was discarded (73 waits ran out)"
The same result for IQ3_XXS and IQ2_XS. Without STRATA_VERIFY_NO_HOST=1, the same point fails earlier with flag M ("... CPU experts' rows ... CPU took 3 ms ...").
## Evidence
- Engine log shows "verify window: 6 tokens at position 0, host at layer step 1; the GPU rang 24; flags: served 0, plan (A) 0, copies (B) 0" (iq2_xs engine log lines 47 and 53), so the plan flag stayed 0 while the GPU rang 24-25 times.
- Kernel journal shows no GPU hang, reset or OOM at that time.
- Attached: iq3_xxs-32k.engine.log, iq2_xs-32k.engine.log, launch-iq2_xs-32k.server.log (credentials checked; none present).
## Reproduction status
Reproduced myself, consistently, on two quants. I have not tried other hardware.
## Other notes
- The launcher's startup wait is 600 + 8 s per GB above 40 GB (886 s for this 76 GB model). With the model on a spinning disk the load takes ~1600 s, so I patched strata_cli/launch/runner.py to 1800 to get this far.
- "RAM is tight - the model's experts take 40.0 GB of this PC's 31 GB" is printed even though the Intel engine reports "no host arena" and mirrors about 17 GiB.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。