marfrit/rk-llama.cpp
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity
9,121 Commits 6 Branches 0 Tags

Default Branch

perf-rknpu2

137b4ceebe · rknpu2: add Q4_K dequant + W4A4_HADAMARD default pattern · Updated 2026-07-09 21:24:08 +00:00
 ZIP  TAR.GZ

Branches

performance-enhancement

396159c9d5 · ggml-rknpu2: graceful abort with W8A8 hint on NPU IOVA exhaustion · Updated 2026-07-20 02:21:49 +00:00    marfrit

0
470
 ZIP  TAR.GZ
rocket-npu-backend

67316774a0 · ggml-rocket: make ROCKET_MIN_BATCH runtime-tunable via GGML_ROCKET_MIN_BATCH · Updated 2026-07-18 13:21:55 +00:00    marfrit

0
465
 ZIP  TAR.GZ
rocket-ksplit-wip

6d5a075a6c · ggml-rocket: K-segmentation for K>8192 (int8-partial float-accumulate) · Updated 2026-07-14 20:52:59 +00:00    marfrit

0
465
 ZIP  TAR.GZ
mtp-b9549

0ebb855635 · docs: int4 closed — no weight-only int4 dtype on RK3588 (hardware limit) · Updated 2026-07-10 14:58:33 +00:00    marfrit

0
457
 ZIP  TAR.GZ
rknpu-int8-clamp

447672d647 · rknpu2: saturate int8 activation quantization · Updated 2026-07-10 07:01:57 +00:00    marfrit

0
453
 ZIP  TAR.GZ