marfrit 6d5a075a6c ggml-rocket: K-segmentation for K>8192 (int8-partial float-accumulate) ...

Enables MUL_MAT with K>ROCKET_K_MAX (e.g. gemma-4 ffn_down K=12288) on the
NPU via 6144-wide K-chunks, mirroring ggml-rknpu2 compute_k_segments.
Validated correct (coherent long-prompt greedy output; int8 2-way partials
lossless enough). pp512-NEUTRAL on gemma-4-E2B (46.61 vs 46.69) — kept on a
branch as a capability, not defaulted. May help models with heavy K>8192.

2026-07-14 22:52:59 +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-rocket: K-segmentation for K>8192 (int8-partial float-accumulate)

2026-07-14 22:52:59 +02:00

.gitignore

…

CMakeLists.txt

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

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