#2490 setup: Alchemist on xe also gets the Level Zero v2 adapter (UR_LOADER_USE_LEVEL_ZERO_V2=1; A770 decode +7%, prompt +3%)

本文

## Title
setup: Alchemist on xe also gets the Level Zero v2 adapter (UR_LOADER_USE_LEVEL_ZERO_V2=1; A770 decode +7%, prompt +3%)

## Summary
On an Arc A770 under xe, each of the stepped verify window's ~97 graph segments per round took 53-200 µs to submit with the default Level Zero adapter. With the v2 adapter it takes about 25 µs (timed in the window's serve).

Setup now writes `UR_LOADER_USE_LEVEL_ZERO_V2=1` for an Alchemist card on xe, next to `UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS`. **This builds on #2394**, which adds that block.

## What changed
- `sycl/setup_intel.py`: the env entry, for Alchemist on xe only.
- `tools/test_setup_sycl.py`: the existing Alchemist/xe test also asserts the variable, and that a B-series card and i915 do not get it.
- `docs/INTEL.md`: one paragraph.

## Extra Notes
Arc A770 16 GB, xe, compute runtime 26.05, oneAPI 2026.1. Same day, same binary, with and without the variable:

| | v2 | default |
|---|---|---|
| 8K decode (3 pairs) | 35.49 / 36.38 / 34.29 | 33.90 / 32.39 / 32.56 tok/s |
| 8K prefill | 791-795 | 768-775 tok/s |
| 32K | 795.9 prefill, 29.97 decode | 782.4 prefill, 28.21 decode |
| GPU wait per round | 39.4-39.7 | 43-48 ms |

**Stability.** With the default adapter, about 1 in 10 runs had a decode about 3× slower (serves stalled under memory pressure); none did with v2.

**Quality.** The text is equivalent. The ids vary from run to run, as they already do with the default adapter, because the adaptive tier's swap timing moves experts between CPU and GPU.

`python tools/test_setup_sycl.py`: 19 tests OK.

Not tested: A750 / A580, i915, Windows, other compute runtimes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)