sycl: run on an integrated Arc (Arc 140T, xe): free memory from /proc/meminfo, no prompt-slot borrowing, a native launcher

[andersson82]

Title

sycl: run on an integrated Arc (Arc 140T, xe): free memory from /proc/meminfo, no prompt-slot borrowing, a native launcher

Issue: none; the "integrated Arc GPUs ... not tested" line of docs/INTEL_ARC.md (and #515 for the B390 on Windows, not covered here)

Summary

The SYCL port did not start on an integrated Arc: the Arc 140T has no free-memory query, so the expert cache was sized
from 0. With the three changes below it runs the Coder IQ1_M on the 140T with 62 GiB of RAM, through the normal server,
at 12.0-17.4 tok/s decode, and reads a 72,648-token prompt at 156.5 tok/s.

What changed

• sycl/include/dpct/device.hpp, get_memory_info: on a device with host_unified_memory and no ext_intel_free_memory,
  free memory is MemAvailable from /proc/meminfo (Linux only); STRATA_SYCL_FREE_MIB sets it by hand (the xe driver
  keeps freed GPU memory in a pool that MemAvailable does not count until a reboot). Other devices: the code path is the
  one before.
• sycl/src/program/generate.cpp: no prompt-slot borrowing on a host_unified_memory device. With borrowing, the answer
  after a borrowing prompt was token 0 (every expert activation NaN from the second prompt chunk; measured on the 0.1.40
  port, any context). Discrete cards keep the old rule.
• sycl/serve/strata-sycl-native.sh (new): the config's exe for the engine on the PC's own oneAPI, no Docker. Sets the
  same port defaults as strata-sycl.sh (each overridable from the config's "env") plus STRATA_NO_MULTI_GR=1
  (15.1 vs 18.3 tok/s on the 140T with the multi-token GR kernels, 0.1.40, same output) and an OOM score of 1000 for the
  engine.
• docs/INTEL.md: a new section "Integrated Arc: the Arc 140T" (build, what is different from a card, a config to run it by
  hand, the numbers). docs/INTEL_ARC.md: the 140T is no longer in "not tested".

Left alone: discrete cards (every change is behind host_unified_memory, which a card with its own VRAM should not
report; I have no discrete card to check), CUDA and HIP (no file of theirs is touched), setup (its own PR, which builds
on this one).

Measurements

Arc 140T (8086:7d51, Xe-LPG+), Core Ultra 9 285H, 62 GiB RAM, Fedora 44 (kernel 7.2.9, xe), oneAPI 2026.0, JIT build.
Coder IQ1_M, --spec 4 --mtp, --max-context 98304, --kv int8, --stream-experts, through
sycl/serve/server_intel.py, greedy, thinking off. Engine 0.1.43 with these changes and #2110 (without #2110 every request
stalls on the 140T).

decode, 100-token answers, 10 requests	12.0-17.4 tok/s, all 12,288 experts cached (first start after a boot)
the same, later start (the xe pool)	11.5-17.2 tok/s, 10,496 cached, 1,792 mirrored
a 72,648-token prompt	read at 156.5 tok/s, the fact on its first line recalled
first start / later starts	79 s / 7 s with SYCL_CACHE_PERSISTENT=1 (0.1.40)

Extra Notes

How I checked:

• Built this branch alone on origin/main (7695859) with sycl/tools/build.sh on the PC's oneAPI: 0 errors.
• bash -n on the launcher.

Not tested:

• Any discrete Intel card (I have none); the code paths there should be unchanged.
• Other integrated GPUs (Lunar Lake, Panther Lake), i915 on an iGPU, Windows.
• The general model at 0.1.43 (IQ2_XS ran on the 0.1.40 port: 13.3-16.2 tok/s decode).
• Where a decode round's time goes (not profiled).

🤖 Generated with Claude Code