…5 -> 90.9), MXFP4 weights (#282)
llama.cpp fork since d60cc4f:
- #64: a kernel for the MUL_MAT ggml hints as GGML_HINT_SRC0_IS_HADAMARD. Bonsai's 1024-point rotation
of the 17408-wide FFN input ran on the CPU in every layer at prompt sizes. Ternary-Bonsai-2-27B,
balanced mode: pp512 13.5 -> 90.9 tok/s, tg128 15.8 -> 19.0. test-hrx-hadamard: six shapes vs CPU
and an exact product, x4, bitwise-identical repeats; 8 sanitizer-clean check.cases.
- #65: MXFP4 in the shared dequantizer, the E8M0 scale built exactly as GGML_E8M0_TO_FP32_HALF.
test-hrx-mxfp4 bit-exact vs ggml on HRX0 (subnormal scales e = 0/1 flush to zero);
test-backend-ops -b HRX0 1051/1051. MUL_MAT_ID admits only multiples of 256 (its kernels declare
it): gpt-oss-20b's 2880-wide experts fall back to the CPU instead of failing the JIT.
Docs: docs/hrx.md Ternary Bonsai section and "Our patches". Registry regenerated (no mapping changes).
Co-authored-by: bong-water-water-bong <bong-water-water-bong@1bit.gg>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Adds the Docs7 analytics script to every page, next to GoatCounter:
<script defer src="https://context7.com/docs7-analytics.js" data-site="…"></script>DOCS7_SITEholds the Docs7 site id. The variable is set to f172a149-d085-45bc-8c58-ae0de677b0bf.🤖 Generated with Claude Code