#1861 tensor parallel intel arc b580

Beschreibung

Card: 2× Intel Arc B580 (BMG-G21), 8086:e20b, xe driver, Level Zero

Versions:
• Engine: 0.1.41-sycl, AOT build bmg-g21 (build-sycl-aot/strata)
• oneAPI: 2026.1, oneMKL 2026.1, Level Zero V2, NEO 26.09.x
• Host: 64 GB RAM, swap 7 GB, PCIe ~10-12 GB/s host→device

Model: Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS (IQ3_XXS, 48 layers, 24,576 experts), 32,768 context, INT8 KV, --spec 4 --spec-min-p 0.5 --mtp, --pool-workers 13, --vram-reserve-mib 1024

Config:

{
  "backend": "sycl",
  "gpu": [0, 1],
  "layer_split": "18",
  "env": {
    "STRATA_VERIFY_NO_HOST": "0",
    "STRATA_VERIFY_DEVICE_PLAN": "1",
    "ONEAPI_DEVICE_SELECTOR": "level_zero:*"
  }
}

Symptom: both GPUs are detected and the split initializes, but the first inference request kills the engine (exit code 1):

strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 18)
strata generate: CUDA1 runs layers 18-47, expert cache 2310 slots (4.08 GiB)
strata generate: layer split: CUDA0 runs layers 0-17
strata serve: layer split: layers 0-17 (CUDA0), 18-47 (CUDA1), one hand-off per window
strata serve: 1564 MiB of VRAM free with everything loaded
strata verify: captured the 6-token window (upload no error, sync no error)
Cannot submit to a queue with a dependency from a graph that is associated with a different context.
Exception caught at file:/work/sycl/src/core/verify.cpp, line:1706

Tried:
• STRATA_VERIFY_DEVICE_PLAN=1 (default) → crash
• STRATA_VERIFY_DEVICE_PLAN=0 → same crash (so it's not the device plan)
• Single GPU ("gpu": 0, level_zero:0) → works fine, Hello!, prompt 41.7 tok/s, decode 33.4 tok/s, 24/30 drafts

Relevant files:
• sycl/src/core/verify.cpp:1706 — the failing submission
• sycl/src/program/generate.cpp — the layer-split hand-off
• sycl/setup_intel.py:154 — host_arena = driver == "i915" (we're xe, so --stream-experts path; removed it → CPU expert pool)
• strata-swift-iq3_xxs.json — the failing config