Markus Fritsche f574ed927c ggml-cpu/repack: ROCKET_NOREPACK_TENSORS env to force tensors off CPU_REPACK ...

Diagnostic knob for per-tensor rocket-vs-CPU_REPACK A/B. Listed name substrings
(space/comma sep, e.g. "ffn_down") skip repack in ggml_repack_get_optimal_repack_type
-> stay host-buffer -> the sched offload gate routes them to the rocket NPU backend.

Measured verdict (gemma-4-E2B-Q8, boltzmann): routing MORE weights to rocket LOSES
(baseline 46.5 -> ffn_down 35 -> all-MLP 22 tok/s); the default hybrid split is already
near-optimal. Kept as a measurement tool, NOT for production use.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EWpfhDgYNA21tETDP9ueBE

2026-07-18 19:16:40 +02:00

..

cmake

ggml : Parallelize quant LUT init (#23595)

2026-05-25 10:15:46 +03:00

include

TP: quantized KV cache support (#23792)

2026-06-01 12:30:10 +02:00

src

ggml-cpu/repack: ROCKET_NOREPACK_TENSORS env to force tensors off CPU_REPACK

2026-07-18 19:16:40 +02:00

.gitignore

…

CMakeLists.txt

ggml: add rocket NPU backend (RK3588 mainline accel driver)

2026-07-14 10:52:06 +02:00