Skip to content

2026.11.20.03

Choose a tag to compare

@rizagp rizagp released this 20 Nov 10:12

Windows: unpack inferzero-2026.11.20.03-a5acc31-win64.zip anywhere, then run setup.cmd (or zero-launcher.cmd once a model is there).

Linux: inferzero-2026.11.20.03-a5acc31-linux-x64.tar.gz (Ubuntu 24.04 and newer): unpack it, then run ./zero-launcher.sh --check and ./zero-launcher.sh. It needs Intel's GPU runtime and nothing else. The steps: docs/install.md.

New since 2026.11.20.02:

  • B-series cards (B580, Pro B50 / B60 / B65 / B70):
    • The FFN's gate and up GEMMs in one launch. A Pro B60's 27B prompt is +3.3 % at 4K (1,806 -> 1,866 tok/s) and +3.8 % at 16K, with the same answers bit for bit (skipzx-lab). ARC_GEMM_XW_GU=0 turns it off.
    • The CPU computes the MoE experts missing from the GPU's cache. With the model partly in RAM: Flash Next on a B60 in a PCIe 3.0 x8 slot generates +17 % (22.5 -> 26.3 tok/s, skipzx-lab). ARC_XCPU=0 turns it off, and is the setting to use when you need a temp-0 answer to repeat exactly run to run.
    • Flash Next's 4,096-token prompt chunks only from 20 GiB of GPU memory. A 16 GB card keeps 2,048: on an A770 that is 14-19 % faster generation after a prompt.
  • Precise mode (--precise) with the CPU experts: the CPU now runs one int8 pass by default (KL against f32 0.0013-0.0025, top token >= 98.4 %, ~40 % less CPU time an expert). ARC_XCPU_Q2=1 gives both passes, as the GPU does.
  • The old-driver warning in plain sight: the web UI shows a bar, the Windows launcher its status line, zero-server status / doctor a line. It now starts below 32.0.101.8976, the oldest driver measured right (32.0.101.8991 and newer no longer warn).
  • Fixes (a self audit):
    • Lunar Lake / Panther Lake: wrong answers after a short follow-up with a resident model.
    • ARC_XCPU did nothing on a launcher package; now it works there.
    • Four races and a leak in the expert cache's background fills.
    • Re-packing a model moved or copied to another drive is no longer refused.
    • Two processes building the same expert file no longer remove each other's .tmp.
    • Ctrl+C / SIGTERM during builds.
    • The warm-up reading past a short context.
  • Opt-in, to be measured by owners:
    • ARC_ATTN_X8=1 (the prompt attention on 8-lane XMX: slower on an A770 so far).
    • ARC_GEMM_XW_NC_E=2 (Alchemist / 140T: the n_embd-wide prompt GEMMs at 32 x 32).
    • bench/xw_pergemm.ps1 / .sh (which tile each of a model's GEMMs wants).

The full list: CHANGELOG.md.