Repository navigation
2026.11.11.3
Unpack anywhere, then run setup.cmd (or zero-launcher.cmd once a model is there).
New since 2026.11.11.2:
- A GPU benchmark:
bin\gpu_bench.exe, or the launcher's More > Benchmark this GPU. It measures the matrix engines (XMX: int8, f16, bf16), the vector engines (FP32, FP16, int8 DP4a) and memory read bandwidth, each next to the GPU's figure on paper (EUs x maximum clock x the per-EU rate of its kind). It opens a page in%USERPROFILE%\.inferzero\bench\gputhat compares every result file in that folder, in the layout of a spec database's card. Owners of an Arc 140T, Pro B60 or B390: please run it and attach the.jsonit writes; it shows whether the vendors' matrix figures hold (the 140T: Intel 77 int8 TOPS, a spec database 154). - Wider-tile prompt GEMMs, opt-in, to measure:
ARC_GEMM_XW=1runs the H8 prompt GEMMs through new kernels with more work per sub-group and fewer memory reads per matrix instruction, the results bit for bit as before (checked at start).- 16-lane XMX (B60, B390, Lunar Lake): 32 tokens x 64 columns per sub-group.
- The Arc 140T: its weights in 4 block reads instead of 32 (ocloc: 6.3 memory reads per DPAS instead of 13.3), the 140T preset's prompt chunks included.
.\bench\xw_sweep.ps1 -Model <27B gguf>compares the tile shapes with the default and checks the logits.
- Gate + up in one GEMV launch, opt-in (
ARC_GEMV_GU=1): one launch instead of two per layer and token, the same results;bench\decode_try.ps1has a row for it (no faster on a laptop iGPU; for a B60 to measure). bench\website_bench.ps1warms each model up first (its first generation number was the kernels' compile time);bench\f33_sweep.ps1prints how far the chunked DeltaNet on XMX differed.
Unchanged by default: every new path is off unless switched on; results bit for bit (Qwen3.5 4B and Qwen3.8 27B checked against the previous build).
Details: CHANGELOG.md, docs/xe2-tuning.md sections 11 and 12.